The Robotics Primer (Maja Mataric) - 2007

The
Robotics
Primer
The
Robotics
Primer
Maja J Matarić
Illustrations by Nathan Koenig

The MIT Press
Cambridge, Massachusetts
London, England
© 2007 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any
electronic or mechanical means (including photocopying, recording, or information
storage and retrieval) without permission in writing from the publisher.
Typeset in 10/13 Lucida Bright by the authors using LATEX 2ε .

Printed and bound in the United States of America.
Library of Congress Cataloging-in-Publication Information
Matarić, Maja J
The robotics primer / Maja J Matarić ; illustrations by Nathan Koenig.
p. cm. (Intelligent robotics and autonomous agents series)
Includes bibliographical references and index.
ISBN 978-0-262-63354-3 (pbk. : alk. paper) 1. Robotics I. Title
TJ211.M3673 2007
629.8’92—dc22 2007000981
To Helena and Nicholas, who teach me what matters, every day.
Brief Contents
1 What Is a Robot? 1
2 Where Do Robots Come From? 7
3 What’s in a Robot? 19
4 Arms, Legs, Wheels, Tracks, and What Really Drives Them 29
5 Move It! 47
6 Grasping at Straws 59
7 What’s Going On? 69
8 Switch on the Light 81
9 Sonars, Lasers, and Cameras 97
10 Stay in Control 121
11 The Building Blocks of Control 135
12 What’s in Your Head? 145
13 Think Hard, Act Later 151
14 Don’t Think, React! 161
15 Think and Act Separately, in Parallel 177
16 Think the Way You Act 187
17 Making Your Robot Behave 207
18 When the Unexpected Happens 215
19 Going Places 223
20 Go, Team! 233
21 Things Keep Getting Better 255
22 Where To Next? 269
Bibliography 289
viii Brief Contents
Glossary 293
Index 303
Contents
Preface xv
1 What Is a Robot? 1
2 Where Do Robots Come From? 7

2.1 Control Theory 7
2.2 Cybernetics 8
2.2.1 Grey Walter’s Tortoise 9
2.2.2 Braitenberg’s Vehicles 11
2.3 Artificial Intelligence 13
3 What’s in a Robot? 19
3.1 Embodiment 20
3.2 Sensing 21
3.3 Action 24
3.4 Brains and Brawn 25
3.5 Autonomy 26
4 Arms, Legs, Wheels, Tracks, and What Really Drives Them 29

4.1 Active vs. Passive Actuation 30
4.2 Types of Actuators 31
4.3 Motors 32
4.3.1 Direct-Current (DC) Motors 32
4.3.2 Gearing 34
4.3.3 Servo Motors 37
4.4 Degrees of Freedom 38
x Contents
5 Move It! 47
5.1 Stability 48
5.2 Moving and Gaits 51
5.3 Wheels and Steering 53
5.4 Staying on the Path vs. Getting There 55
6 Grasping at Straws 59
6.1 Endeffectors 59
6.2 Teleoperation 60
6.3 Why Is Manipulation Hard? 62
7 What’s Going On? 69

7.1 Levels of Processing 73
8 Switch on the Light 81

8.1 Passive vs. Active Sensors 81
8.2 Switches 82
8.3 Light Sensors 84
8.3.1 Polarized Light 86
8.3.2 Reflective Optosensors 86
8.3.3 Reflectance Sensors 88
8.3.4 Infra Red Light 89
8.3.5 Modulation and Demodulation of Light 90
8.3.6 Break Beam Sensors 90
8.3.7 Shaft Encoders 91
8.4 Resistive Position Sensors 94
8.4.1 Potentiometers 95
9 Sonars, Lasers, and Cameras 97

9.1 Ultrasonic or Sonar Sensing 97
9.1.1 Sonar Before and Beyond Robotics 100
9.1.2 Specular Reflection 101
9.2 Laser Sensing 104
9.3 Visual Sensing 107
9.3.1 Cameras 108
9.3.2 Edge Detection 110
9.3.3 Model-Based Vision 112
9.3.4 Motion Vision 113
9.3.5 Stereo Vision 114
9.3.6 Texture, Shading, Contours 116
Contents xi
9.3.7 Biological Vision 116

9.3.8 Vision for Robots 117
10 Stay in Control 121

10.1 Feedback or Closed Loop Control 121
10.2 The Many Faces of Error 122
10.3 An Example of a Feedback Control Robot 124
10.4 Types of Feedback Control 126
10.4.1 Proportional Control 126
10.4.2 Derivative Control 128
10.4.3 Integral Control 129
10.4.4 PD and PID Control 130
10.5 Feedforward or Open Loop Control 131
11 The Building Blocks of Control 135

11.1 Who Needs Control Architectures? 135
11.2 Languages for Programming Robots 137
11.3 And the Architectures are... 139
11.3.1 Time 141
11.3.2 Modularity 141
11.3.3 Representation 142
12 What’s in Your Head? 145

12.1 The Many Ways to Make a Map 146
12.2 What Can the Robot Represent? 148
12.3 Costs of Representing 148
13 Think Hard, Act Later 151

13.1 What Is Planning? 152
13.2 Costs of Planning 154
14 Don’t Think, React! 161

14.1 Action Selection 166
14.2 Subsumption Architecture 169
14.3 Herbert, or How to Sequence Behaviors Through the
World 172
15 Think and Act Separately, in Parallel 177

15.1 Dealing with Changes in the World/Map/Task 179
15.2 Planning and Replanning 180
xii Contents
15.3 Avoiding Replanning 181

15.4 On-Line and Off-Line Planning 182
16 Think the Way You Act 187

16.1 Distributed Representation 192
16.2 An Example: Distributed Mapping 193
16.2.1 Toto the Robot 194
16.2.2 Toto’s Navigation 194
16.2.3 Toto’s Landmark Detection 196
16.2.4 Toto’s Mapping Behaviors 197
16.2.5 Path Planning in Toto’s Behavior Map 200
16.2.6 Toto’s Map-Following 202
17 Making Your Robot Behave 207

17.1 Behavior Arbitration: Make a Choice 207
17.2 Behavior Fusion: Sum It Up 209
18 When the Unexpected Happens 215

18.1 An Example: Emergent Wall-Following 215
18.2 The Whole Is Greater Than the Sum of Its Parts 216
18.3 Components of Emergence 218
18.4 Expect the Unexpected 218
18.5 Predictability of Surprise 219
18.6 Good vs. Bad Emergent Behavior 220
18.7 Architectures and Emergence 221
19 Going Places 223

19.1 Localization 225
19.2 Search and Path Planning 228
19.3 SLAM 229
19.4 Coverage 230
20 Go, Team! 233

20.1 Benefits of Teamwork 234
20.2 Challenges of Teamwork 236
20.3 Types of Groups and Teams 238
20.4 Communication 241
20.4.1 Kin Recognition 247
20.5 Getting a Team to Play Together 248
20.5.1 I’m the Boss: Centralized Control 248
Contents xiii
20.5.2 Work It Out as a Team: Distributed Control 248

20.6 Architectures for Multi-Robot Control 250
20.6.1 Pecking Orders: Hierarchies 251
21 Things Keep Getting Better 255

21.1 Reinforcement Learning 256
21.2 Supervised Learning 260
21.3 Learning by Imitation/From Demonstration 262
21.4 Learning and Forgetting 266
22 Where To Next? 269
22.1 Space Robotics 273
22.2 Surgical Robotics 274
22.3 Self-Reconfigurable Robotics 276
22.4 Humanoid Robotics 277
22.5 Social Robotics and Human-Robot Interaction 278
22.6 Service, Assistive and Rehabilitation Robotics 280
22.7 Educational Robotics 283
22.8 Ethical Implications 285
Bibliography 289
Glossary 293
Index 303
Preface
Upon arriving at USC in 1997 as an assistant professor in Computer Science, I

designed the Introduction to Robotics course (http://www-scf.usc.edu/∼csci4
45). The highlights of the course were the intensive hands-on LEGO robotics
lab, team projects, and the end-of-semester contest held at the nearby Cal-
ifornia Science Center. I posted my lecture notes on the web, for students’
convenience, and found that, over the years, a growing number of faculty
and K-12 teachers around the world contacted me about using those in their
courses and getting additional materials from me. In 2001, shortly after re-
ceiving tenure, being promoted to rank of associate professor, and expecting
our second child, perhaps influence of the euphoria of it all, I had a decep-
tively simple thought: why not turn all those course notes into a book?
As I started converting my sometimes-cryptic lecture notes into book chap-
ters, I realized the true magnitude of the challenge I had set up for myself,
that of writing a textbook for an audience that intended to span K-12 and uni-
versity educators and students, as well as robotics enthusiasts who wanted
to go beyond the popular press and get under the surface of the topic. The
readers of the book would (hopefully) span the ages of early adolescence to
retirement, and education levels of elementary school to PhD. This was not
going to be easy.
My motivation came from the fact that robotics is a wonderful field of
study. I am not a typical engineer; instead of making robots out of old
watches and radios in my parent’s basement as a kid, I was interested in
art and fashion design. I discovered the field late in college, through extra-
curricular reading while majoring in computer science; back then there were
no robotics courses at the University of Kansas. Upon entering graduate
school at MIT, I serendipitously chose robotics as my direction of study,
based mostly on the charisma of my PhD advisor, Rodney Brooks, who was
the first to make a compelling case for it. The goal of this book is to make a
xvi Preface
compelling case for studying robotics to a wide variety of readers, because

the substance of robotics, while challenging, is incredibly interesting.
Creating intelligent machines can fuel our imagination, creativity, and de-
sire to make a difference. Robotics technology will play a key role in the
future of humanity, and we are currently very directly shaping that future in
the research labs and classrooms around the country and the world. Most
importantly, however, studying robotics early on can help the student to es-
tablish a foundation for a better understanding of science, technology, engi-
neering, and math (STEM) topics. This is why I would love to see all children
work with robots in school, starting in elementary school (5th grade is cer-
tainly old enough, and earlier is possible), and continuing throughout high
school. Getting hooked on creative robotics opens the students’ minds and
builds their confidence for a great variety of topics and eventually, careers
in STEM. Without such fun, mind-opening learning and discovery, students
often incorrectly judge STEM as either uncool or unattainable. The goal of
this book is to teach something real and true about robotics and its potential,
while staying interesting and engaging.
I’m amazed at how long it took to write this book. My son, who was about
to be born when the book idea occurred to me, just turned five. All kinds
of good things have happened in my career also, which slowed down the
writing. Eventually, I realized that life would just keep getting busier and
that there would never come "a great time to finish the book." Instead, time
had to be grabbed, borrowed, and stolen to get it done.
In the (long) course of writing this book, many wonderful K-12 and univer-
sity educators provided feedback on the drafts. I listened to all of them. The
students and instructors of the Intro to Robotics course I mentioned above
used drafts of the book, putting up with missing chapters and non-existent
figures. Marianna O’Brien, who teaches 8th grade science at Lincoln Middle
School in Santa Monica, was an angel: she spent over a year reading drafts,
making numerous notes and improving suggestions, and helping me to get
a glimpse at what it takes to be an effective pre-university teacher and how
to write a book that could help. I owe Marianna a huge thank you for her
tireless enthusiasm and detailed and encouraging comments! I am grateful
to Nate Koenig, a USC PhD student from my group who made the great
figures for the book, and then got pulled into doing the formatting, the copy-
right permissions, the index, and the references, as well as making what no
xvii
doubt seemed to be an endless stream of small fixes in the last few months.
The book owes its good looks to Nate.
I am grateful to Helen Grainer and Rodney Brooks of iRobot Corp., who
agreed to sponsor a free web-based robot programming workbook compan-
ion for this book, found at http://roboticsprimer.sourceforge.net/workbook.
The workbook, which I developed with my PhD students Nate Koenig and
David Feil-Seifer, provides step-by-step guidance, example exercises, and so-
lutions that should make real robot programming accessible to the readers of
this book and anybody else interested in robotics.
Role models are incredibly important. My mother, who has written nu-
merous short story, essay, and poetry books, was an inspiration; she made
the endeavor seem readily within my reach. One of the goals of this book is
to get people of all ages interested in robotics, so that we can create numer-
ous role models, and never have to see kids avoid robotics (and STEM topics
in general) from ignorance, peer pressure, or lack of confidence.
Above all, I am grateful to my family, who put up with some grabbed and
borrowed (but never stolen) time needed to write this book. My husband
provided moral support throughout and knew how to say "I don’t know
how you do it" in a way that inspired and encouraged me to figure out how
to do it. Finally, my two kids are the ones who motivated me the most to
write the book, since they will shortly be part of the intended audience. If
they and their peers think robotics is cool (and thereby Mom is cool as well),
and consider taking up creative endeavors in science and engineering, then
I’ve done something good.
South Pasadena, California, April 2007
1 What Is a Robot?
Defining Robotics
Welcome to The Robotics Primer! Congratulations, you have chosen a very

cool way to learn about a very cool topic: the art, science, and engineering
of robotics. You are about to embark on a (hopefully fun) ride that will end
with your being able to tell what is real and what is not real in movies and
articles, impress friends with cool facts about robots and animals, and much
more, most important of which is being able to better build and/or program
your own robot(s). Let’s get started!
What is a robot?
This is a good question to ask because, as with any interesting area of science
and technology, there is a lot of misunderstanding about what robots are and
are not, what they were and were not, and what they may or may not become
in the future. The definition of what a robot is, has been evolving over time,
as research has made strides and technology has advanced. In this chapter,
we will learn what a modern robot is.
The word ”robot” was popularized by the Czech playwright Karel Capek
(pronounced Kha-rel Cha-pek) in his 1921 play Rossum’s Universal Robots
(R.U.R.). Most dictionaries list Capek as the inventor of the word “robot,”
but more informal sources (such as the Web) say that it was actually his
brother, Josef, who made up the term. In either case, the word “robot” re-
sulted from combining the Czech words rabota, meaning “obligatory work”
and robotnik, meaning “serf.” Most robots today are indeed performing such
obligatory work, in the form of repetitive and fixed tasks, such as automobile
assembly and DNA sequencing. However, robotics is about much more than
obligatory labor, as you will see.
The idea of a robot, or some type of machine that can help people, is much
older than the Capek brothers. It is not possible to pin point where it orig-
2 1 What Is a Robot?
inated, because it is likely that many clever engineers of the past thought
of it in some form. The form changed over time, as science and technology
advanced, and made many of the previously unachievable dreams of robots
become a reality or at least enter the realm of possibility.
With the advancement of science and technology, the notion of a robot
has become more sophisticated. In the past, a robot was defined as a ma-
chine that was, basically, a clever mechanical device. Examples of such de-
vices, even extremely sophisticated ones, can be found throughout history,
and they go way back. As long as 3000 years ago, Egyptians used human-
controlled statues; and more recently, in the seventeenth and eighteenth cen-
tury Europe, various clockwork-based “lifelike” creatures were constructed
that could sign, play the piano, and even “breathe.” But, as we will see, those
were not really robots, not by our current definition and understanding of
what a robot is.
While original notions of robots were really of clever mechanical automa-
tons, as computational machinery developed (and in particular when it shrank
in size sufficiently to be imaginable within a robot’s body), the notions of
robots started to include thought, reasoning, problem-solving, and even emo-
tions and consciousness. In short, robots started to look more and more like
biological creatures, ranging from insects to humans.
These days we have (or should have) a very broad idea of what a robot can
be, and do not need to limit ourselves by what is currently possible mechani-
cally or computationally. However, it is still hard to anticipate how our ideas
of what a robot is and can be will evolve as science and technology advance.
So back to the question: what is a robot? What makes the machine in
figure 1.1 a robot, but those in figure 1.2 merely robot wannabes?
A robot is an autonomous system which exists in the physical
R OBOT world, can sense its environment, and can act on it to achieve
some goals.
This may seem like a very broad definition, but actually each of its parts is
important and necessary. Let’s take it apart to see why.
A robot is an AUTONOMOUS system
A UTONOMOUS An autonomous robot acts on the basis of its own decisions, and is not con-
trolled by a human.
There are, of course, plenty of examples of machines that are not autonomous,
T ELEOPERATED but are instead externally controlled by humans. They are said to be teleop-
3
Figure 1.1 An example of a robot.
erated; tele means “far” in Greek, so teleoperation means operating a system

from afar.
These machines, however, are not true robots. True robots act autonomous-
ly. They may be able to take input and advice from humans, but are not
completely controlled by them.
A robot is an autonomous system which exists in the PHYSICAL
WORLD
Existing in the physical world, the same world that people and animals
and objects and trees and and weather and many other things exist in, is a
fundamental property of robots. Having to deal with that physical world,
and its unbendable physical laws and challenges, is what makes robotics
what it is: a real challenge. Robots that exist on the computer are simula-
tions; they do not really have to deal with true properties of the physical
world, because simulations are never as complex as the real world. There-
fore, although there are a lot of simulated robots in cyberspace, a true robot
exists in the physical world.
world, can SENSE its environment
Figure 1.2 Examples of non-robots. On the left is a system that does not exist in the
physical world; on the right, a system that is not autonomous. These robot wannabes
are not the real thing.
S ENSORS Sensing the environment means the robot has sensors, some means of per-
ceiving (e.g., hearing, touching, seeing, smelling, etc.) in order to get infor-
mation from the world. A simulated robot, in contrast, can just be given the
information or knowledge, as if by magic. A true robot can sense its world
only through its sensors, just as people and other animals do. Thus, if a sys-
tem does not sense but is magically given information, we do not consider
it a true robot. Furthermore, if a system does not sense or get information,
then it is not a robot, because it cannot respond to what goes on around it.
world, can sense its environment, and can ACT ON IT
Taking actions to respond to sensory inputs and to achieve what is desired
is a necessary part of being a robot. A machine that does not act (i.e., does
not move, does not affect the world by doing/changing something) is not a
robot. As we will see, action in the world comes in very different forms, and
that is one reason why the field of robotics is so broad.
world, can sense its environment, and can act on it to ACHIEVE
SOME GOALS.
Now we finally come to the intelligence, or at least the usefulness, of a
robot. A system or machine that exists in the physical world and senses it,
5
but acts randomly or uselessly, is not much of a robot, because it does not use
the sensed information and its ability to act in order to do something useful
for itself and/or others. Thus, we expect a true robot to have one or more
goals and to act so as to achieve those goals. Goals can be very simple, such
as “Don’t get stuck” or quite complex, such as “Do whatever it takes to keep
your owner safe.”
Having defined what a robot is, we can now define what robotics is.
R OBOTICS Robotics is the study of robots, which means it is the study of their au-
tonomous and purposeful sensing and acting in the physical world.
It is said that Isaac Asimov, the amazingly productive science fiction writer,
was the first to use the term robotics, based on Capek’s term robot. If so, he is
the one who officially gave the name to the large and rapidly growing area
of science and technology that this book aims to introduce.
To Summarize
• Robotics is a fast-growing field whose definition has been evolving over

time, along with the field itself.
• Robotics involves autonomy, sensing, action, and achieving goals, all in

the physical world.
Food for Thought
• What else can you do from afar, by means of teleoperation? You can talk,
write, and see, as in telephone, telegraph, and television. There is more.
Can you think of it?
• Is a thermostat a robot?
• Is a toaster a robot?
• Some intelligent programs, also called software agents, such as Web crawle-
rs, are called “softbots.” Are they robots?
• Is HAL, from the movie 2001, the Space Odyssey, a robot?

Looking for More?
• This book comes with a free robot programming workbook you can down-
load from the World Wide Web at:
http://roboticsprimer.sourceforge.net/workbook/
The workbook follows the structure of this textbook with exercises and so-
lutions you can use to learn much more about robotics by trying it hands-
on.
• Here is a fun, short little introductory book about robotics: How to Build a
Robot by Clive Gifford. It’s very easy, has great cartoons, and touches on
a lot of concepts that you can then learn about in detail in this book.
• Robotics, The Marriage of Computers and Machines by Ellen Thro is another

short introductory book you might enjoy.
• For a truly comprehensive review of modern robotics, complete with a

great many robot photos, see Autonomous Robots by George Bekey.
• To learn more about various topics in this book and in general, look up
things on Wikipedia, an ever-growing free encyclopedia found on the Web
at http://en.wikipedia.org/wiki/.
Where Do Robots Come From?
2 A Brief but Gripping History of

Robotics
Have you wondered what the first robot ever built was like, who built it,
when it was built, and what it did? It turns out that it is not easy to an-
swer these questions precisely. Many machines have been built that could
be called robots, depending on how one defines a robot. Fortunately, we
have a modern definition of a robot from Chapter 1 and, using that defini-
tion, the first robot is typically considered to be W. Grey Walter’s Tortoise. In
this chapter, we will learn about the Tortoise, its creator, the related fields of
control theory, cybernetics, and artificial intelligence (AI), and how together
they played key parts in the history of robotics, from its humble origins to its
current state of the art.
2.1 Control Theory
C ONTROL T HEORY Control theory is the mathematical study of the properties of automated con-
trol systems ranging from steam engines to airplanes, and much in between.
Control theory is one of the foundations of engineering; it studies a vast
variety of mechanical systems that are a part of our everyday lives. Its for-
malisms help us understand the fundamental concepts which govern all me-
chanical systems. Since both the art of building automated systems and the
science of studying how they work date back to ancient times, control the-
ory can be said to have originated with the ancient Greeks, if not before.
It was largely studied as part of mechanical engineering, the branch of en-
gineering that focuses on designing and building machines, and on their
physical properties, and was used to study and further develop the control
of water systems (for irrigation and house-hold use), then temperature sys-
8 2 Where Do Robots Come From?
tems (for metalsmithing, animal care, etc.), then windmills, and eventually
steam engines (the machines that were part of the industrial revolution). By
the start of the twentieth century, classical mathematics, such as differential
equations, was applied to formalizing the understanding of such systems,
giving birth to what we think of as control theory today. The mathemat-
ics got more complicated as mechanisms began to incorporate electrical and
electronic components.
In Chapter 10, we will discuss feedback control, a concept from control the-
ory that plays an important role in robotics.
2.2 Cybernetics
While control theory was growing and maturing, another field relevant to
robotics emerged. It was pioneered in the 1940s, the years around the second
World War, by Norbert Wiener. Wiener originally studied control theory,
and then got interested in using its principles to better understand not just
artificial but also biological systems.
C YBERNETICS The field of study was named cybernetics, and its proponents studied bio-
logical systems, from the level of neurons (nerve cells) to the level of behav-
ior, and tried to implement similar principles in simple robots, using meth-
ods from control theory. Thus, cybernetics was a study and comparison of
communication and control processes in biological and artificial systems.
Cybernetics combined theories and principles from neuroscience and biol-
ogy with those from engineering, with the goal of finding common proper-
ties and principles in animals and machines. As we will see, W. Grey Walter’s
Tortoise is an excellent example of this approach.
The term “cybernetics” comes from the Greek word “kybernetes” which
means governor or steersman, the name for a central component of James
Watt’s steam engine, which was designed based on the ideas of windmill
control. In cybernetics, the idea was that machines would use a similar
steersman to produce sophisticated behavior similar to that found in nature.
A key concept of cybernetics focuses on the coupling, combining, and in-
teraction between the mechanism or organism and its environment. This
interaction is necessarily complex, as we will see, and is difficult to describe
formally. Yet this was the goal of cybernetics, and is still an important com-
ponent of robotics. It also led to the development of the Tortoise, which can
be considered to be the first robot, based on our definition in Chapter 1.
2.2 Cybernetics 9
Figure 2.1 One of the tortoise robots built by W. Grey Walter. (Photo courtesy of Dr.
Owen Holland)
2.2.1 Grey Walter’s Tortoise

William Grey Walter (1910-1977) was an innovative neurophysiologist inter-
ested in the way the brain works. He made several discoveries, including
the theta and delta brain waves, the activity your brain produces during
different stages of sleep. Besides pursuing neuroscience research, he was
interested in studying brain function by building and analyzing machines
with animal-like behavior.
Today, we call machines with properties similar to those of biological sys-
B IOMIMETIC tems biomimetic, meaning that they imitate biological systems in some way.
During his research, around the 1940s and thereafter, Grey Walter built a va-
riety of cleverly designed machines, which he called “turtles” or “tortoises,”
after the turtle character in Lewis Carroll’s Alice in Wonderland, and because
of their biomimetic behavior. The best known ones were named Elmer and
Elsie, loosely based on the acronyms for ELectro MEchanical Robots, and
Light Sensitive. The turtles were simple robots built with three wheels in
a tricycle-like design, using the front wheel for steering and the two back
wheels for driving. They were covered with a clear plastic shell, giving them
a life-like appearance, at least to a friendly and open-minded observer. Fig-
ure 2.1 shows what one of the turtles looked like.
W. Grey Walter (the first name, William, is rarely spelled out) gave his
turtles Latin names to describe their behaviors, such as Machina Speculatrix
and Machina Docilis. Machina Speculatrix means “machine that thinks” or
“speculates,” and Machina Docilis means “machine that can be tamed/taug-
ht,” by which he meant a machine that can learn, because it could be trained
with whistles.
Let’s look at what Machina Speculatrix consisted of:
• One photocell, a sensor that detects light levels (we will learn a great deal
about such sensors in Chapter 8)
• One bump sensor, a sensor that detects contact with objects (we will also
learn a great deal about such sensors in Chapter 8)
• One rechargeable battery
• Three motors, one for each wheel (we will learn about motors and other
actuators in Chapter 4)
• Three wheels (we will learn about wheels and other effectors in Chapter 4)
• One analog electronic circuit using two vacuum tubes (very old technol-
ogy!) serving as the “brain” and connecting the two sensors with the
A NALOG wheels. An analog signal is one that is continuous in time and amplitude
A NALOG E LECTRONIC (amount of wave oscillation); an analog electronic circuit processes analog
C IRCUIT signals.
From these simple components, Grey Walter endowed his Machina Spec-
ulatrix with the following behaviors or capabilities:
• Find the light
• Head toward the light
• Back away from bright light
• Turn and push to avoid obstacles
• Recharge battery.
As we will see later, this and other robots constructed by Grey Walter used
R EACTIVE C ONTROL reactive control, a means of controlling robots using a collection of prioritized
“reflexes.” When properly combined, these simple reflex-like rules resulted
in animal-like behavior. We will revisit reactive control in more detail in
Chapter 14, to learn what makes a robot turtle-like (besides being built with
a turtle-like shell).
The turtles manifested a variety of complex patterns of behaviors that were
E MERGENT B EHAVIOR not programmed in, a phenomenon we now call emergent behavior and will
2.2 Cybernetics 11
cover in detail in Chapter 18. Most importantly, they were excellent exam-
ples of how animal-like, realistic behavior can be achieved with very simple
artificial mechanisms.
Grey Walter believed that intelligence comparable with that found in ani-
mals can be created in machines through the combination of techniques from
robotics and artificial intelligence (AI). But since those two fields did not yet
exist in his day, cybernetics, a combination of the two, emerged under his in-
fluence and the influence of other prominent figures such as Norbert Weiner,
who was mentioned above. Cybernetics effectively combined “thinking,”
“acting,” and the interaction with the environment. As we will see below,
those fundamental components of robotics were subsequently split between
the fields of AI and robotics, with AI taking over thinking and robotics con-
cerning itself with action in the physical world. This separation and treat-
ment of each component in isolation worked to the detriment of progress
in both fields. In contrast, cybernetics was much more integrated. It took
robotics a while to come back to those good ideas, as you will see.
Interestingly, Grey Walter’s turtles also represent some of the earliest ex-
A RTIFICIAL L IFE amples of what is now called artificial alife or alife, another field of research
that is related to robotics but mostly focuses on computational, not physical,
systems that exist within a computer, not in the real world.
Grey Walter’s robots were shown around the world during his life-time,
and can now be seen in museums. (Interestingly, extinct robots such as Grey
Walter’s are often preserved much better than any extinct natural species.)
In the 1990s, one of the turtles was found (in a house, not freely roaming
around); it was refurbished and taken on a tour of various robotics confer-
ences. It gave no autographs or interviews, but it did inspire people once
again.
2.2.2 Braitenberg’s Vehicles

One of the people who was inspired by W. Grey Walter’s original work was
Valentino Braitenberg, who wrote a book, titled Vehicles, in 1984, well after
the field of cybernetics had died out. The book describes a series of ideas,
T HOUGHT or thought experiments or gedanken experiments (from the German word for
E XPERIMENTS thought), that show how one can design simple robots (which he called “ve-
G EDANKEN
hicles”) and get them to produce behaviors that appear very animal-like and
E XPERIMENTS
life-like. This book has been a source of inspiration for roboticists, even if
Braitenberg himself did not actually build any of the robots/vehicles he de-
scribed.
Figure 2.2 An example of Braitenberg’s vehicles, in simulation. (Photo courtesy of

Valention Braitenberg, "Vehicles: Experiments in Synthetic Psychology", published
by The MIT Press)
B RAITENBERG ’ S Braitenberg’s vehicles started out with a single motor and a single light sen-
V EHICLES sor, and gradually progressed to more motors and more sensors, and more
interesting connections between them, all using analog electronic circuits.
The sensors were directly connected to the motors, so that the sensory input
could drive the motor output. For example, a light sensor could be connected
directly to the wheels, so the stronger the light, the faster the robot moved,
P HOTOPHILIC as if it were attracted to the light; that’s called photophilic in Latin, literally
meaning “loving light.” Alternatively, in some vehicles the connection was
inverted, so that the stronger the light, the slower the robot would move, as
P HOTOPHOBIC if it were repulsed by the light or afraid of it; that’s called photophobic in Latin,
literally meaning “afraid of light.”
E XCITATORY A connection between the sensors and motors in which the stronger the
C ONNECTION sensory input, the stronger the motor output, is called an excitatory connec-
I NHIBITORY tion, because the input excites the output. Conversely, a connection in which
C ONNECTION the stronger the sensory input, the weaker the motor output, is called an in-
hibitory connection, because the input inhibits the output. The inspiration for
these connections comes from biology, since they are similar to (but crude
when compared with) neurons, which can have excitatory and inhibitory
connections with each other. By varying the connections and their strengths
(as with neurons in brains), numerous behaviors result, ranging from seek-
ing and avoiding light, much like Grey Walter’s turtles, to what looks like
social behavior, and even aggression and love.
Braitenberg’s book describes how such simple mechanisms can be em-
ployed to store information, build a memory, and even achieve robot learn-
ing. Some of his simpler designs have been successfully built by hobbyists
and fledgling roboticists (like yourself), and have also been a source of in-
spiration for more advanced robotics research. Like Grey Walter’s turtles,
Braitenberg’s vehicles were also reactive robots. We will learn more about
them, too, in Chapter 14.
While the field of cybernetics concerned itself with the behavior of robots
and their interaction with the environment, the new emerging field of artifi-
cial intelligence focused on, well, intelligence, naturally (no, artificially!).
2.3 Artificial Intelligence

Doesn’t AI sound cool? Or, alternatively, it does it seem scary as it is usually
portrayed in movies? In any case, chances are that it seems much more pow-
erful and complex than it really is. So let’s dispel the myths and see what AI
has to do with robotics.
A RTIFICIAL The field of artificial intelligence (AI) was officially “born” in 1956 at a con-
I NTELLIGENCE ference held at Dartmouth University, in Hanover, New Hampshire. This
meeting brought together the most prominent researchers of the day, in-
cluding Marvin Minsky, John McCarthy, Allan Newell, and Herbert Simon,
names you can’t miss if you read about the origins of AI, who are now con-
sidered founders of that field. The goal of the meeting was to discuss creating
intelligence in machines. The conclusions of the meeting can be summarized
as follows: in order for machines to be intelligent, they will have to do some
heavy-duty thinking; and to do that, they will need to use the following:
• Internal models of the world
• Search through possible solutions
• Planning and reasoning to solve problems
• Symbolic representation of information
• Hierarchical system organization
• Sequential program execution.
Don’t worry if much of the above list does not make sense right now. We
will learn more about these concepts in Chapters 12 and 13, and will see that
some of them may not play a fundamental role in aspects of robotics. The
important outcome of the Dartmouth conference was that as the field of AI
was being funded and shaped, it had a strong influence on robotics, or at
least on the branch we can call AI-inspired robotics.
Figure 2.3 Shakey, in its specially designed world. (Photo courtesy of SRI Interna-
tional, Menlo Park)
S HAKEY So what is AI-inspired robotics? Shakey, shown in figure 2.3, is a good

example of an early AI-inspired robot. Built at the Stanford Research Insti-
tute in Palo Alto, California, in the late 1960s, it used contact sensors and
a camera. As we will see in Chapter 9, cameras, which are vision sensors,
provide a great deal of information to a robot and are very complicated to
deal with. Because the new field of AI focused on reasoning, and vision re-
quired complex processing, those early roboticists were able to apply early
AI techniques to their robots. Shakey “lived” in a very special indoor world
consisting of a flat white floor and some large black objects, such as balls and
pyramids. It carefully (and slowly, given the state of computer technology at
that time) created plans for which way to move in that special world. People
stayed out of its way or did not move around it. Given the state of robotics
technology at that time, Shakey shook a bit when it tried to execute its plans:
hence its name.
Shakey is a popular and well-known (now retired) early robot. Its succes-
sor was called Flakey, and you can probably guess why. Let us consider some
other examples of early AI-inspired robots:
• HILARE: developed at the LAAS Lab (which stands for Laboratoire d’Ana-
lyse et D’Architecture des Systemes, loosely the “analysis and architec-
Figure 2.4 HILARE, another early robot. (Copyright LAAS-CNRS)
tures of systems”) in Toulouse, France in the late 1970s, used a video cam-
era, ultrasound sensors, and a laser rangefinder (we will learn about those
in Chapters 9). Unlike most other research robots, HILARE was used by
many generations of researchers and is one of the longest-lived robots to
date.
• CART: developed at Stanford University in Palo Alto, California, in 1977

by Hans Moravec, as part of his doctoral thesis (just a part of it; it takes
a lot to get a PhD in robotics, but it’s worth it). It was literally a cart
on bicycle wheels. Moravec is now considered one of the founders of
modern robotics, and his recent research has been on ultrasound sensing.
But back in the days of his doctoral thesis, he focused on vision sensors
(so to speak), and in particular on how to use vision to move the robot
V ISION -B ASED around, which we call vision-based robot navigation. (We’ll talk about nav-
N AVIGATION igation and its challenges in Chapter 19.) CART used vision to move
very slowly, not because it was slow-moving by design, but because it
was slow-thinking. The thinking took a long time, as with other early
AI-inspired robots (Shakey, HILARE), because of the difficulty of process-
ing data from vision cameras and the slowness of computer processors in
those days. It is said that one can see the shadows move as time passes
while CART is thinking about and planning what to do next.
• Rover: developed at Carnegie Mellon University (known as CMU), in

Pittsburgh, Pennsylvania, in 1983 also by Hans Moravec. This was the
robot he built after he got is doctoral degree and became a professor at
CMU. Rover used a camera and ultrasound sensing for navigating. (You
may be starting to notice that these old robots did nothing else but navi-
gate, and that was hard enough. Chapter 19 will explain why that is still
a hard, but much better-handled problem.) Rover was more advanced in
form than CART; it looked more like modern mobile robots do these days,
not like the CART did, but its thinking and acting were still very slow.
Since they thought too hard and too slowly, and failed to move very effec-
tively and responsively, these early robots of the 1970s and 1980s provided
important lessons for the fledgling field of robotics. The lessons, not surpris-
ingly, mostly had to do with the need to move faster and more robustly, and
to think in a way that would allow such action. In response, in the 1980s
robotics entered a phase of very rapid development. In that decade, several
new robotics directions emerged, shook things up, and eventually got orga-
T YPES OF R OBOT nized into the types of robot control that we use today: reactive control, hybrid
C ONTROL control, and behavior-based control. We will learn about them in Chapters 14,
15 and 16. These are all effective replacements for what early AI-inspired
robotics used, and what we now call purely deliberative control. We will learn
about it, and why it is not in use in today’s robotics, in Chapter 13.
If you are really paying attention, you have probably noticed that these
early AI-inspired robots were quite different in every way from Grey Wal-
ter’s tortoises and from Breitenberg’s vehicles. AI was (and is) very different
from cybernetics in its goals and approaches. But robotics needs to combine
both, in order to create robust, fast, and intelligent machines. What it takes
to do that, and how it is done, is what the rest of this book is about.
Recapping a brief but gripping history of robotics, it can be said that mod-
ern robotics emerged from the historic developments and interactions among
a handful of key research areas: control theory, cybernetics, and artificial
intelligence. These fields have had important and permanent impact on
robotics today, and at least two of them still continue as major areas of re-
search in their own right. Control theory thrives as a field of study applied
to various machines (though typically not intelligent ones); control-theoretic
ideas are at the core of low-level robot control, used for getting around. Cy-
bernetics no longer exists under that name, but research into biologically in-
spired biomimetic methods for robot control is very much alive and thriving,
and is being applied to mobile robots as well as humanoids. Finally, artifi-
cial intelligence is now a very large and diverse field of research that focuses
largely, but not exclusively, on non physical, unembodied cognition (this ba-
sically means “disembodied thinking”; we will talk about how the body im-
pacts thinking in Chapter 16). Robotics is growing by leaps and bounds. See
Chapter 22 to find out where all the robots are today, what they are up to,
and where the field is going next.
To Summarize
• Robotics grew out of the fields of control theory, cybernetics, and AI.
• The first modern robot, W. Grey Walter’s tortoise (one of them; there were
several), was constructed and controlled using the principles of cybernet-
ics.
• Braitenberg’s vehicles provided more examples of cybernetics and biomi-

metic principles.
• Early AI influenced early AI-inspired robotics which produced Shakey,

the Stanford CART, HILARE, and the CMU Rover, among others.
• AI-inspired early robots were very different from the tortoise and from
Braitenberg’s vehicles, and in general from biomimetic cybernetic ma-
chines.
• The current approaches to robot control, including reactive, hybrid, and

behavior-based control, emerged from the influences of and lessons learn-
ed from control theory (the mathematics of controlling machines), cyber-
netics (the integration of sensing, action, and the environment), and AI
(the mechanisms for planning and reasoning).
Food for Thought
How important is it for robots to be inspired by biological systems? Some

people argue that biology is our best and only model for robotics. Others
say that biology teaches us valuable lessons, but that engineering can come
up with its own, different solutions. The airplanes is a popular example
of non-biomimetic engineering; people’s first attempts to build flying ma-
chines were inspired by birds, but today’s planes and helicopters share little
in common with birds. Does it matter what kind of robot you are building
(biomimetic or not)? Does it matter if it is going to interact with people? For
more on this last topic, wait for Chapter 22.
Looking for More?
• The two original sources on cybernetics, written by the field’s founders,

are: Cybernetics or Control and Communication in the Animal and the Machine
by Norbert Wiener (1948) and Introduction to Cybernetics by W. R. Ashby
(1956).
• W. Grey Walter wrote a book and several articles about the brain. He
also published two Scientific American articles about his robotics ideas “An
imitation of life”, in 1950 (182(5): 42-45), and “A machine that learns”, in
1951 (185(2): 60-63).
• Hans Moravec wrote a great many articles about robotics. He also wrote
two books: Mind Children and Robot: Mere Machine to Transcendent Mind.
Both books make predictions about the future of robotics and artificial life;
they are so forward-looking that they sound quite like science fiction.
3 What’s in a Robot?
Robot Components
Have you wondered what makes up a robot, what makes it tick? (Actually,
it better not tick; things don’t tick these days, unless something is wrong!) In
this chapter, we will look inside a robot to see what it is made of, in terms
of both hardware and software. Peering inside will cause us to introduce
and briefly review many of the concepts which will be talked about in more
detail in the chapters to come. Here we will provide an overview of how
these many ideas and concepts come together to constitute a robot.
Let’s go back to our trusty definition of a robot:
world, can sense its environment, and can act on it to achieve
some goals.
The definition already gives us some hints as to what a robot consists of.
Specifically, it tells us that a robot’s main components are:
• A physical body, so it can exist and do work in the physical world
• Sensors, so it can sense/perceive its environment
• Effectors and actuators, so it can take actions
• A controller, so it can be autonomous.
Figure 3.1 shows these required components of a real robot, and the way in
which they interact with each other and with the robot’s environment. Let’s
talk a bit about each of these components: embodiment, sensing, action, and
autonomy.
20 3 What’s in a Robot?
Figure 3.1 The components of a robot.
3.1 Embodiment
Having a body is the first requirement for being a robot; it allows the robot to
do things: move and shake, go places, meet people, and get its work done. A
computer agent, no matter how realistically animated, is not a robot because
it does not share the physical universe with the rest of us physical creatures;
E MBODIMENT it is not physically embodied in the real world. Embodiment refers to having a
physical body. It is necessary for a robot, but it has a price:
• We are all in the same boat: An embodied robot must obey the same physical
laws that all physical creatures obey. It can’t be in more than one place at
a time; it can’t change shape and size arbitrarily (although some robots
can do a bit of shape-shifting, as you’ll see in Chapter 22); it must use
the effectors on its body to actively move itself around (more on that in
Chapter 4), it needs some source of energy to sense, think, and move,
and when moving, it takes some time to speed up and slow down; its
movement affects its environment in various ways; it cannot be invisible;
and so on.
• You are in my space: Having a body means having to be aware of other

bodies and objects around. All physical robots have to worry about not
3.2 Sensing 21
running into or hitting other things in their environment and, in some

cases, about not hitting themselves (as you will see in Chapter 6). This
seems easy, but it’s not, and so it’s typically the first thing you’ll program
a robot to do: avoid collisions.
• I can only do so much: Every body has its limitations. The shape of the
robot’s body, for example, has a lot to do with how it can move, what
it can sense (because sensors are attached to the body, in some way or
another), what work it can do, and how it can interact with other robots,
cars, lawnmowers, cats, stuff, and people in its environment.
• All in good time: Besides obviously influencing things that have to do

with space and movement, the body also influences things that have to
do with time. The body determines how fast the robot can move (on its
own power, at least) and react to its environment. It has often been said
that faster robots look smarter; that’s another example of how the robot’s
embodiment influences its “image.”
3.2 Sensing
S ENSORS Sensors are physical devices that enable a robot to perceive its physical envi-
ronment in order to get information about itself and its surroundings. The
S ENSING terms sensing and perception are treated as synonyms in robotics; both refer to
P ERCEPTION the process of receiving information about the world through sensors.
What does a robot need to sense? That depends on what the robot needs to
do, on what its task is. A good robot designer and programmer puts the right
kinds of sensors on the robot that allow it to perceive the information it needs
to do its job, to achieve its goals. Similarly, animals have evolved sensors that
N ICHE are well suited for their niche, namely their environment and position in their
ecosystem. Robots have a niche as well, consisting of their environment and
their task, and just like animals, the better they fit that niche, the longer they
will survive (in the lab, as a commercial product, etc.). We will talk more
about this in Chapter 7.
S TATE Sensing allows the robot to know its state. State is a general notion from
physics that is borrowed by robotics (and by computer science and AI, among
other fields) and refers to the description of a system. A robot’s state is a de-
scription of itself at any point in time. The robot (system) is said to be “in
a state” (kind of like being “in a state of disrepair”). The more detailed the
Figure 3.2 A robot’s sensors and its state space.
description is, the larger the state is, because it takes more bits or symbols to
write it down.
To a robot (actually, to any perceiving creature), state may be visible (for-
O BSERVABLE , mally called observable), partially hidden (partially observable), or hidden (un-
PARTIALLY observable). This means a robot may know very little or a great deal about
O BSERVABLE , AND itself and its world. For example, if the robot cannot “feel” or “see” one of
H IDDEN S TATE
its arms, the state of that arm is hidden from it. This can be a problem if
the robot needs to use the arm to reach for something, such as the hospital
patient it is trying to help. At the same time, the robot may have detailed in-
formation about the hospital patient, the patient’s state, through a machine
that senses and monitors the patient’s vital signs. We will talk more about
state and what to do with it in Chapter 12.
D ISCRETE AND State may be discrete (up, down, blue, red) or continuous (3.785 miles per
CONTINUOUS STATE hour). This refers to the type and amount of information used to describe the
system.
S TATE S PACE State space consists of all possible states a system can be in. For example,
if a light switch can be only on or off, its state space consists of two discrete
3.2 Sensing 23
states, on and off, and so is of size 2. If, on the other hand, the light switch
has a dimmer on it, and can be turned to a variety of light levels, then it
has many more states (possibly infinitely many continuous states). The term
S PACE space in this context refers to all possible values or variations of something.
Figure 3.2 shows an example of a robot that has two on/off bumpers and a
high/low battery, and the state space that results from those.
It is often useful for a robot to distinguish two types of state relative to
E XTERNAL AND itself: external and internal. External state refers to the state of the world as
I NTERNAL S TATE the robot can perceive it, while internal state refers to the state of the robot as
the robot can perceive it. For example, a robot may sense that it is dark and
bumpy around it (external state) and also that its battery level is low (internal
state).
Internal state can be used to remember information about the world (a
R EPRESENTATION path through a maze, a map, etc.). This is called a representation or an internal
model. Representations and models have a lot to do with how complex a
robot’s brain is; we will talk about this at length in Chapter 12.
In general, how intelligent the robot appears to be strongly depends on
how much and how quickly it can sense its environment and itself: its exter-
nal and internal state.
What can a robot sense? That depends on what sensors the robot has. All
of the robot’s sensors, put together, create the space of all possible sensory
S ENSOR S PACE readings, which is called the robot’s sensor space (also called perceptual space).
P ERCEPTUAL S PACE The sensor or perceptual space for a robot that has just a single on-off con-
tact sensor has two possible values: on and off. Now suppose the robot has
two such sensors; in that case its sensor space consists of four values: on+on,
on+off, off+on, off+off. That’s the basic idea, and you can see that a robot’s
sensor space grows pretty quickly as you add more sensors or more complex
sensors. That’s one of the reasons that robots need brains.
Sensors are the machine versions of eyes, ears, noses, tongues, hairs, and
various other sense organs in animals. But, as you will learn in Chapters 8
and 9, robot sensors, which include cameras, sonars, lasers, and switches,
among others, are quite different from biological sensors. A robot designer
and programmer needs to put her/his mind into the robot’s sensor space in
order to imagine how the robot perceives the world, and then how the robot
should react to it. This is not easy to do, precisely because while robots share
our physical world, they perceive it very differently than we do.
Figure 3.3 A robot that combines mobility and manipulation.
3.3 Action
E FFECTORS Effectors enable a robot to take action, to do physical things. They are the next
best thing to legs, flippers, wings, and various other body parts that allow
animals to move. Effectors use underlying mechanisms, such as muscles
A CTUATORS and motors, which are called actuators and which do the actual work for the
robot. As with sensors, robotic effectors and actuators are very different from
biological ones. They are used for two main activities:
1. Locomotion: moving around, going places
2. Manipulation: handling objects.
These correspond to the two major subfields of robotics:
1. Mobile robotics; which is concerned with robots that move around, mostly
on the ground, but also in the air and under water.
2. Manipulator robotics; which is concerned mostly with robotic arms of var-

ious types.
3.4 Brains and Brawn 25
Mobile robots use locomotion mechanisms such as wheels, tracks, or legs,

and usually move on the ground. Swimming and flying robots are also mo-
bile robots, but they usually move in higher dimensions (not just on the
ground), and are therefore even harder to control. Manipulators refer to vari-
D EGREES OF F REEDOM ous robot arms and grippers; they can move in one or more dimensions. The
dimensions in which a manipulator can move are called its degrees of freedom
(DOF). We will learn more about that in Chapter 6.
The separation between mobile and manipulator robotics is slowly dis-
appearing as more complex robots, such as humanoids which have to both
move around and manipulate objects, are being developed. Figure 3.3 shows
a robot that combines mobility and manipulation.
3.4 Brains and Brawn

Robots are different from animals in that biological brains take a great deal
of energy (power) relative to the rest of the body, especially in humans. In
robots, it’s the other way around, with the actuators taking more power than
the processor running the controller, the brain.
When it comes to computation, on the other hand, animals and robots are
similar; both need a brain in order to function properly. For example, cats
can walk without using their brain, purely using the signals from their spinal
cord, and people can survive for years in a coma. Similarly, well-designed
robots may move about randomly when their processor is reset. In all cases,
this is not a good thing, so make sure you properly protect your brain and
your robot’s brain, too.
Although we won’t talk much in this book about power issues, power is a
major problem in practical robotics. Power issues include:
• Providing enough power for a robot without loading it down with heavy
batteries
• Effectively keeping the electronics isolated from the sensors and effectors
• Preventing loss of performance as the power levels drop because batteries

wear down or there is a sudden high power demand (due to running
into the wall, as discussed in Chapter 4, or to a high-current sensor, as
discussed in Chapter 9)
• Replenishing power in an autonomous fashion, by the robot itself instead

of by people
• And many others.
As you dive into robotics, keep in mind that the brains and the brawn are
directly connected and related, and should be designed and treated together.
3.5 Autonomy
C ONTROLLERS Controllers provide the hardware and/or software that makes the robot au-
tonomous by using the sensor inputs and any other information (for instance,
whatever may be in the memory) to decide what to do (what actions to take),
and then to control the effectors to execute that action. Controllers play the
role of the brain and the nervous system. We say “controllers” and not “a
controller” on purpose; typically there is more than one, so that various parts
of the robot can be processed at the same time. For example, various sensors
may have separate controllers, as may various effectors. Do all things come
together in a central place? This turns out to be a major issue in robotics: they
may or may not, as they may or may not in biology. We will talk more about
that when we get to robot control in Chapters 13, 14, 15, and 16. Whatever
kind they may be, controllers enable robots to be autonomous.
You will see examples of controllers throughout this book. They will be
P SEUDOCODE written in pseudocode (from the Greek pseudo meaning “false” and code mean-
ing “program”), an intuitive way of describing the controller, not any partic-
ular robot programming language.
A UTONOMY Autonomy is the ability to make one’s own decisions and act on them. For
robots, autonomy means that decisions are made and executed by the robot
itself, not by human operators. Autonomy can be complete or partial. Com-
pletely autonomous robots, like animals, make their own decisions and act
on them. In contrast, teleoperated robots are partially or completely con-
trolled by a human and are more like complex puppets than like animals.
According to our definition, puppetlike robots that are not autonomous at
all are not really robots.
To Summarize
• The key components of a robot are sensors, effectors, and controllers.
• Sensors provide information about the world and the robot itself. They
define the sensory or perceptual space of the robot and allow it to know its
state, which may be discrete, continuous, observable, partially observable,
or hidden.
3.5 Autonomy 27
• Effectors and actuators provide the ability to take actions. They may pro-
vide locomotion or manipulation, depending on the type of robot: mobile
or manipulator.
• Controllers provide autonomy, which may be partial or complete.
Food for Thought
• What do you think is more difficult, manipulation or mobility? Think

about how those abilities develop in babies, children, and then adults.
• How large do you think your sensor space is?
• Can you think of things or information in your life that is observable, par-
tially observable, or hidden?
Looking for More?
The Robotics Primer Workbook exercises for this chapter are found here:
http://roboticsprimer.sourceforge.net/workbook/Robot_Components
Arms, Legs, Wheels, Tracks, and
4 What Really Drives Them

Effectors and Actuators
It should come as no surprise that while robotic bodies are inspired by bio-
logical ones, they are quite different from their biological counterparts in the
way they are constructed and the way they work. In this chapter, we will
learn about the components of a robot that enable it to take actions in order
to achieve its goals.
E FFECTOR An effector is any device on a robot that has an effect (impact or influence)
on the environment.
Effectors range from legs and wheels to arms and fingers. The robot’s con-
troller sends commands to the robot’s effectors to produce the desired effect
on the environment, based on the robot’s task. You can think of effectors
as equivalent to biological legs, arms, fingers, and even tongues, body parts
that can “do physical work” of some kind, from walking to talking. Just as
sensors must be well matched to the robot’s task, so must the effectors.
A CTUATOR All effectors have some mechanism that allows them to do their work. An
actuator is the mechanism that enables the effector to execute an action or
movement.
In animals, muscles and tendons are the actuators that make the arms and
legs and the backs do their jobs. In robots, actuators include electric mo-
tors, hydraulic or pneumatic cylinders, temperature-sensitive or chemically-
sensitive materials, and various other technologies. These mechanisms actu-
ate the wheels, tracks, arms, grippers, and all other effectors on robots.
30 4 Arms, Legs, Wheels, Tracks, and What Really Drives Them
Figure 4.1 A passive walker: a robot that uses gravity and clever mechanics to bal-
ance and walk without any motors. (Photo by Jan van Frankenhuyzen, courtesy of
Delft University of Technology)
4.1 Active vs. Passive Actuation
In all cases, the action of actuators and effectors requires some form of en-
PASSIVE A CTUATION ergy to provide power. Some clever designs use passive actuation, utilizing
potential energy in the mechanics of the effector and its interaction with the
environment instead of active power consumption.
Consider, for example, flying squirrels. They don’t actually fly at all, but
glide through the air very effectively, using the skin flaps between their front
leggs and torso. Glider planes and hang gliders use the same principle. Birds
use their wings for gliding as well, but since wings can be (and usually are)
used for flying, which involves active flapping, we don’t think of bird wings
as a very good example of a purely passive effector. Some plants have small
wings on their seeds which enable them to glide far away from the parent
plant in order to spread. Nature is full of examples of such effectors that
use clever designs to conserve energy and serve multiple uses, including the
inspiring of roboticists.
An example of an inspired robotics design for a passive effector, originally
developed by Tad McGeer, is shown in figure 4.1. McGeer developed a walk-
4.2 Types of Actuators 31
ing machine, called the Passive Walker, which had two legs with knees that
looked basically like a pair of human legs, but had no motors or muscles or
any active actuators at all. When put at the top of a downward-sloping ramp,
the Passive Walker would walk in a manner that looked very humanlike.
What do you think made it walk? It was actuated by gravity, and would
eventually come to rest at the bottom of the ramp. Of course, if put on a
flat surface, or worse yet at the bottom of an incline, it would not walk at
all, but simply fell over. It was not perfect, but it was a clever design that in-
spired many people. You’ll notice that simple clever solutions always inspire
people, not just in robotics but in any field of pursuit.
4.2 Types of Actuators
So, we have learned that there are passive and active actuators, and that
there are various clever designs for both. Now, let’s look into some options
for active actuators. As you might imagine, there are many different ways of
actuating a robotic effector. They include:
• Electric motors: These are the most common robotics actuators, the most
affordable and the simplest to use, based on electric current. We’ll learn
much more about them in this chapter.
• Hydraulics: These actuators are based on fluid pressure; as the pressure

changes, the actuator moves. They are quite powerful and precise, but are
also large, potentially dangerous, must be well-packaged, and of course
must be kept from leaking!
• Pneumatics: These actuators are based on air pressure; as the pressure

changes, the actuator moves. Much like hydraulic actuators, they are typ-
ically large, very powerful, potentially dangerous, and must also be kept
from leaking.
• Photo-reactive materials: Some materials perform physical work in re-

sponse to the amount of light around them; they are called photo-reactive.
Usually the amount of generated work (and thus movement) is quite small,
and therefore this type of actuator is currently used only for very small,
micro-scale robots. (Such photo-reactive materials are also used in tinted
glasses that automatically darken as needed. However, in those cases they
do not produce movement, and are therefore not actuators.)
• Chemically reactive materials: As the name implies, these materials re-

act to certain chemicals. A good example is a particular type of fiber
which contracts (shrinks) when put in an acidic solution and elongates
(gets longer) when put in a basic (alkaline) solution. Such materials can
L INEAR A CTUATOR be used as linear actuators; they provide linear movement (getting longer
or shorter), which is very different from the rotary movement provided
by motors.
• Thermally reactive materials: These materials react to changes in temper-

ature.
• Piezoelectric materials: These materials, usually types of crystals, create

electric charges when pushed/pressed.
This is not a complete list; engineers are developing new types of actuators
all the time. But let’s get back to the basics and look at the simplest and most
popular robotic actuators, the ones you are most likely to use: motors.
4.3 Motors
Motors are the most common actuators in robotics. They are well suited for
actuating wheels, since they provide rotational movement, thereby enabling
wheels to turn, and wheels, of course, are very popular effectors (in robotics
and on the ground in general). Motors are also very useful for driving other
types of effectors, besides wheels, as you will see next.
4.3.1 Direct-Current (DC) Motors

DC M OTOR Compared with all other types of actuators, direct current (DC) motors are
simple, inexpensive, easy to use, and easy to find. They can be purchased in
a great variety of sizes and packages, to accommodate different robots and
tasks. This is important, since you remember from earlier in this book that a
good robot designer matches all parts of the robot, including the actuators,
to the task.
You may know from physics that DC motors convert electrical energy into
mechanical energy. They use magnets, loops of wire, and current to generate
magnetic fields whose interaction turns the motor shaft. In this way, electro-
magnetic energy becomes kinetic energy producing movement. Figure 4.2
shows a typical DC motor; the components are hidden in the packaging, but
you can see the shaft on the left and the power wires on the right.
4.3 Motors 33
Figure 4.2 A standard direct-current (DC) motor, the most common actuator used
in robotics.
To make a motor run, you need to provide it with electrical power in the
right voltage range. If the voltage is low, but not too low, the motor will still
run, but will have less power. On the other hand, if the voltage is too high,
the power of the motor is increased, but the wear and tear will make it break
down sooner. It is much like revving a car engine; the more you do it, the
sooner the engine will die.
When a DC motor is provided with nice constant voltage in the right range,
it draws current in the amount proportional to the work it is doing. Work,
as defined in physics, is the product of force and distance, so when a robot
is pushing against a wall, the motors driving the wheels are drawing more
current, and draining more of its batteries, than when the robot is moving
freely, without obstacles in its way. The reason for the higher current is the
physical resistance of the stubborn wall, which keeps the distance traveled
small or zero, resulting in high force for the same amount of work being
done. If the resistance is very high (the wall just won’t move, no matter how
much the robot pushes against it), the motor draws the maximum amount
of power it can, and then, having run out of options, stalls. A car also stalls
under similar conditions, but it’s best not to try this at home.
The more current the motor uses, the more torque (rotational force) is pro-
duced at the motor shaft. This is important, since the amount of power a
motor can generate is proportional to its torque. The amount of power is also
proportional to the rotational velocity of the shaft. To be precise, the amount
of power is proportional to the product of those two values, the torque and
the rotational velocity.
When the motor is spinning freely, with nothing attached to its shaft, then
its rotational velocity is the highest but the torque is zero. The output power,
then, is also 0. In contrast, when the motor is stalled, its torque is the maxi-
mum it can produce but the rotational velocity is zero, so the output power is
again zero. Between these extremes of free spinning and stalling, the motor
actually does useful work and drives things efficiently.
How efficiently? you might ask. That depends on the quality of the motor.
Some are very efficient, while others waste up to half of their energy. It gets
worse with other types of motors, such as electrostatic micro-motors, which
are used for miniature robots.
How fast do motors turn?
Most DC motors have unloaded (free-spinning) speeds in the range of 3000

to 9000 revolutions per minute (rmp), which is the same as 50 to 150 revolu-
tions per second (rps). That means they produce high speed but low torque,
and are therefore well suited for driving light things that rotate very fast,
such as fan blades. But how often does a robot need to drive something like
that? Unfortunately, not often. Robots need to do work: pull the loads of
their bodies, turn their wheels, and lift their manipulators, all of which have
significant mass. This requires more torque and less speed than off-the-shelf
DC motors provide. So what can we do to make standard DC motors useful
to robots?
4.3.2 Gearing
G EARS We can use gears! Combining different gears is used to change the force and
torque output of motors.
The force generated at the edge of a gear is the ratio of the torque to the ra-
dius of the gear. By combining gears with different radii, we can manipulate
the amount of force and torque that gets generated.
Suppose that we have two gears, one of which is attached to the shaft of
the motor and is called the input gear, the other being the output gear. The
4.3 Motors 35
Figure 4.3 An example of a 3-to-1 (3:1) gear reduction.
torque generated at the output gear is proportional to the torque on the input
gear and the ratio of the radii of the two gears. Here is the general rule:
If the output gear is larger than the input gear, the torque increases. If
the output gear is smaller than the input gear, the torque decreases.
The torque is not the only thing that changes when gears are combined;
there is also a corresponding change in speed. If the circumference of the
input gear is twice that of the output gear then the output gear must turn
twice for each rotation of the input gear in order to keep up, since the two
are physically connected by means of their teeth. Here is the general rule:
If the output gear is larger than the input gear, the speed decreases. If
the output gear is smaller than the input gear, the speed increases.
Another way to think about it is:
When a small gear drives a large one, torque is increased and speed is
decreased. Analogously, when a large gear drives a small one, torque is
decreased and speed is increased.
This is how gears are used to trade off the extra speed of DC motors, which
is not often useful in robotics, for additional torque, which is.
Figure 4.4 An example of ganged gears; two 3:1 gears in series produce 9:1 gear
reduction.
Gears are combined by “the skin of” their teeth. Gear teeth require special
design so that they mesh properly. Any looseness between meshing gears,
B ACKLASH called backlash, makes the gear mechanism move sloppily back and forth be-
tween the teeth, without turning the whole gear. Backlash adds error to the
positioning of the gear mechanism, which is bad for the robot because it
won’t know exactly where it is positioned. Reducing backlash requires tight
meshing between the gear teeth, but that, in turn, increases friction between
the gears, which wastes energy and decreases the efficiency of the mecha-
nism. So, as you can imagine, proper gear design and manufacturing are
complicated, and small, high-precision, carefully machined gearboxes are
expensive.
Let’s get back to looking gears in the mouth and counting the teeth. To
achieve a particular gear reduction, we combine gears of different sizes (dif-
ferent numbers of teeth). For example, to achieve three-to-one or (3:1) gear
reduction, we mesh a small gear (say one with eight teeth) with a large one
(say one with 3 × 8 = 24 teeth). Figure 4.3 illustrates this; the large gear is
the output gear and the small gear is the input gear. As a result, we have
slowed down (divided) the large gear by a factor of three and at the same
time have tripled (multiplied) its torque, thereby slowing down and increas-
4.3 Motors 37
ing the torque of the motor. To achieve the opposite effect, we switch the
input and output gear.
Gears can be organized in a series or “ganged,” in order to multiply their
G EARS I N S ERIES , effect. For example, two 3:1 gears in series result in a 9:1 reduction. This
G ANGED G EARS requires a clever arrangement of gears. Three 3:1 gears in series can produce
a 27:1 reduction. Figure 4.4 shows an example of ganged gears with the
9:1 resulting ratio. This method of multiplying reduction is the underlying
mechanism that makes DC motors useful and ubiquitous (commonplace),
because they can be designed for various speeds and torques, depending on
the robot’s tasks.
4.3.3 Servo Motors

DC motors are great at rotating continuously in one direction. But it is often
useful for a robot to move a motor, and some effector attached to it, such as
an arm, to a particular position.
S ERVO M OTORS Motors that can turn their shaft to a specific position are called servo motors
or servos for short, because they can servo to a particular position. Servos are
used in toys a great deal, such as for adjusting steering in remote-controlled
(RC) cars and wing position in RC airplanes.
Would you guess that DC and servo motors are very different? Actually,
they are not. In fact, servo motors are made from DC motors by adding the
following components:
• Some gear reduction, for the reasons you saw above
• A position sensor for the motor shaft, to track how much the motor is
turning, and in what direction
• An electronic circuit that controls the motor, in order to tell it how much
to turn and in what direction.
Servo motor operation is all about getting the shaft of the motor in the
desired position. This position is somewhere along 180 degrees in either
direction from the reference point. Therefore, instead of turning the possible
360 degrees, the shaft of a servo motor is usually limited to only half of that,
180 degrees.
The amount of the turn within the zero to 180 degree range is specified
by an electronic signal. The signal is produced as a series of pulses; when
the pulse arrives, the shaft of the motor turns, and when no pulse arrives, it
stops. The up-and-down pattern of the pulses produces a wave like pattern,
WAVEFORM called the waveform.
The amount the motor shaft turns when the pulse arrives is determined
by the duration of the pulse; the longer the pulse, the larger the turn angle.
P ULSE -W IDTH This is called pulse-width modulation because the width (duration) of the pulse
M ODULATION modulates the signal. The exact width of the pulse is quite important, and
must not be sloppy. There are no milliseconds or even microseconds to be
wasted here, or the motor will behave badly, jitter, or try to turn beyond its
mechanical limit (those 180-degree stops). In contrast, the duration between
the pulses is not as important, since that is when the shaft is stopped.
Notice that both kinds of motors we have talked about so far, continuous
rotation and servo, control the position of the shaft. Most robot actuators use
P OSITION C ONTROL position control, in which the motor is driven so as to track the desired position
at all times. This makes motor-driven actuators very accurate, but also very
stiff. In practice this means the actuator locks onto the desired position and, if
disturbed, produces a large force in order to keep from losing that position.
T ORQUE C ONTROL This can be a bad thing for a robot. One alternative is to use torque control
instead of position control: the motor is driven so as to track the desired
torque at all times, regardless of its specific shaft position. The result is an
actuator that is much less stiff.
Considering the work they must do, motors, both DC and servo, require
more power (more current) to run than electronics do. For comparison, con-
sider that the 68HC11 processor, an 8-bit microprocessor used to control sim-
ple robots, draws 5 milliamps of current, while a small DC motor used on
the same type of robots draws from 100 milliamps to 1 amp of current. Typ-
ically, specialized circuitry is required for control of motors. Since this is not
an electronics textbook, we won’t go into more detail on this topic.
4.4 Degrees of Freedom
Now that we have spent some time talking about motors as the most popular
actuators, let’s get back to effectors and how they move. The wheel is a
pretty simple effector, whose job is to turn. A wheel can (typically) move
in only one way, by turning. (As we will see later, there are some cleverly
embellished wheel designs that provide sideways motion as well, but let’s
stick with simple turning for now.) Wheels can turn and motors make them
turn.
Figure 4.5 The six degrees of freedom (DOF) of a freely moving object (in this case
a helicopter) in space are x, y, z, roll, pitch, and yaw.
D EGREE OF F REEDOM A degree of freedom (DOF) is any of the minimum number of coordinates
required to completely specify the motion of a mechanical system. You can
think of it informally as a way in which the system (robot) can move.
How many degrees of freedom a robot has, is important in determining
how it can impact its world, and therefore if, and how well, it can accomplish
its task.
In general, a free (unattached, not tied, bolted, or otherwise constrained)
body (such as a robot) in 3D space (normal space around you has three di-
T RANSLATIONAL DOF mensions) has a total of six DOF. Three of those are called translational DOF,
as they allow the body to translate, meaning move without turning (rota-
tion). They are usually labeled x, y and z, by convention. The other three are
R OTATIONAL DOF called rotational DOF, as they allow the body to rotate (turn). They are called
roll, pitch, and yaw.
As an example, imagine a flying helicopter, such as the one shown in fig-
R OLL , P ITCH , AND
YAW
ure 4.5. Climbing up, diving down, and moving sideways correspond to
the translational DOF. Rolling from side to side, pitching up and down, and
yawing (turning) left or right correspond to the rotational DOF. The six DOF
together correspond to all the possible ways the helicopter could ever move.
While it can’t move in more than six DOF, it can move in fewer, as it turns
Figure 4.6 The degrees of freedom of cars and helicopters.
out. If you put the helicopter on the ground, and drive it there like a car, it
has fewer DOF. How many does it have? Hang on, we’ll get there.
An effector can have any number of DOF, starting with one. (Zero DOF
does not make much sense, as it can’t have an effect on the environment.) A
wheel has one DOF, as we saw, and it usually has one motor associated with
it, for actuating and controlling that DOF. But as you saw above, an effector
may be passive, which means it may have DOF that are not actively actuated
and controllable.
Most simple actuators, such as motors, control a single motion (up-down,
left-right, in-out, etc.) of an effector. The shaft on a motor, when attached
to a wheel (with or without all the gears in between), controls a single DOF,
which is enough for a wheel, which has only one. But more complex effec-
tors, such as robotic arms, have many more DOF (you’ll soon see how many
more), and thus require more actuators.
C ONTROLLABLE DOF If a robot has an actuator for every DOF, then all of the DOF are controllable.
This is the ideal scenario, but it is not always the case. The DOF that are not
U NCONTROLLABLE controllable are aptly called uncontrollable DOF. Let’s work with a concrete
DOF example see what that’s all about.
We agreed earlier that a helicopter has six DOF and moves in three dimen-
sions (3D). If we were to make it move in only two dimensions (2D), on the
ground, it would have fewer DOF. Consider the car, which moves on the
ground in 2D only, at least when moving safely and controllably. A car has
three DOF: position (x,y) and orientation (theta). This is because any body
that moves on a surface (such as a car on the road, a boat on the water, or a
person on the street) moves in 2D. In that flat 2D world, only three of the six
DOF are possible. In the flat plane, which is 2D, a body can translate along
two of the three translational DOF, but not the third, since it is the vertical
dimension, out of the plane. Similarly, the body can rotate in one dimension
(yaw, also called turn in 2D) but not the other two (roll and pitch), since they
are outside of the plane. So by staying on the ground (on the plane), we go
from 3D to 2D and from six DOF down to three.
But it gets worse. A car, like any other body in 2D, can potentially move
in three ways, but it can only do so if it has effectors and actuators that can
control those three DOF. And cars do not! The driver of the car can control
only two things: the forward/reverse direction (through the gas pedal and
the forward/reverse gear) and the rotation (through the steering wheel). So,
although a car has three DOF, only two of them are controllable. Since there
are more DOF than are controllable, there are motions that cannot be done
with a car, such as moving sideways. That’s why parallel parking is hard;
the car has to be moved to the desired location between two cars and next to
the curb, but the only way to get it there is through a series of incremental
back-and-forth moves, instead of one sideways move that would be possible
with an additional DOF. Figure 4.6 illustrates the degrees of freedom of cars
and helicopters.
Before you suddenly lose faith in cars, note that the two controllable DOF
of a car are quite effective: they can get the car to any position and orienta-
tion in 2D, which is about as much as you can ask for, but the car may have to
T RAJECTORY follow a very complicated path (also called trajectory) to get there, and may
have to stop and back up (as in parallel parking). In robotics, a formal way
to say it is that a car can get to any place in the plane by following a continu-
ous trajectory (meaning it does not have to fly) but with discontinuous velocity
(meaning it has to stop and go).
In summary, uncontrollable DOF create problems for the controller in that
they make movement more complicated. The ratio of the controllable DOF
(let’s call them CDOF) to total DOF (let’s call them TDOF) on a robot tells
us quite a bit about how easy it is to control its movement. There are three
possibilities:
1. CDOF = TDOF When the total number of controllable DOF is equal to

the total number of DOF on a robot (or actuator), the ratio is 1, and the
H OLONOMIC robot is said to be holonomic. A holonomic robot or actuator can control all
of its DOF.
2. CDOF < TDOF When the number of controllable DOF is smaller than
the total number of DOF, the ratio is less than 1, and the robot is said to be
N ONHOLONOMIC nonholonomic. A nonholonomic robot or actuator has more DOF than it can
control.
3. CDOF > TDOF When the number of controllable DOF is larger than the
total DOF, the ratio is larger than 1, and the robot is said to be redundant.
R EDUNDANT A redundant robot or actuator has more ways of control than the DOF it
has to control.
What does it all mean?
A good example of a holonomic mechanism is a helicopter, which has six

DOF, and can have all of those six DOF controlled. The same is true for au-
tonomous helicopters, which are robots that can fly under their own control.
A good example of a nonholonomic mechanism is of course the car, as we
saw above. The ratio of controllable to total DOF in a car is 2/3, less than 1,
and so we are stuck with the challenges of parallel parking. The same will
be true for autonomous cars, unless they have some fancy wheels.
Can you imagine what a redundant mechanism is like?
A great example is your arm. The human arm, not including the hand, has
seven DOF: three in the shoulder (up-down, side-to-side, rotation about the
axis of the arm), one in the elbow (open-close), and three in the wrist (up-
down, side-to-side, and again rotation). The shoulder is a little tricky, since it
B ALL - AND -S OCKET is based on a ball-and-socket joint, which is quite complicated and still some-
J OINT thing that has not been replicated in robotics. Also, the rotational DOF in the
wrist really comes from the muscles and ligaments in the forearm. In spite of
all that complication, in the end the human arm has seven DOF. If you have
fewer, go see a doctor. If you have more, that is worthy of scientific attention.
(Being double-jointed, as some people are in the elbow for example, does not
increase the total number of DOF; it just increases the range of motion in the
particular DOF, such as the elbow.)
All of the seven DOF in the human arm are controllable, as you could see
while you were counting them. That brings us to the puzzle: an object in 3D
Figure 4.7 A humanlike seven DOF arm and an exoskeleton for teleoperating it.
Both are made by Sarcos. (Photo courtesy of Sarcos Inc.)
space can have at most six DOF, but your arm has seven! How can your arm
have more? Well, it doesn’t really; the parts of the arm can still move only in
3D and within the six possible DOF, but the arm as a whole, because of the
DOF in its joints, allows for more than one way to move its parts (the wrist,
the fingertip, etc.) to a particular position in 3D space. This is what makes
it redundant, meaning that there are many (infinitely many, in fact) solutions
to the problem of how to move from one place to another. This is, it turns
out, why the control of complicated robotic arms, such as humanoid arms, is
a very hard problem that requires some rather fancy math, as discussed (but
not described in detail) in Chapter 6.
Figure 4.7 shows an example of a humanlike seven DOF robot arm, and
an exoskeleton that can be used to teleoperate it. The exoskeleton is called
the “master arm” and the robot arm is called the “slave arm,” because the
former can be used to teleoperate the latter. However, the robot arm can also
be programmed to be completely autonomous, and is then no slave at all,
but a proper robot.
In general, effectors are used for two basic purposes:
1. Locomotion: to move the robot around
2. Manipulation: to move other objects around.
We will learn about locomotion and manipulation in the next two chapters.
To Summarize
• Effectors and actuators work together to enable the robot to do its job.
Both are inspired by biological systems, but are quite different from them.
• There are several types of robot actuators, with motors being the most
popular.
• Gears are used to decrease motor speed and increase power, using simple
relationships of gear size, speed, and torque.
• Servo motors have particular properties that complement DC motors and

are useful in both toys and robots, among other places.
• Degrees of freedom (DOF) specify how a body, and thus a robot, can
move. The match between the robot’s DOF and its actuators (which de-
termine which DOF can be controlled) determines whether the robot is
holonomic, nonholonomic, or redundant, and has a most important im-
pact on what the robot can do.
Food for Thought
• How would you measure a motor’s torque? What about its speed? These
measurements are important for controlling robot movement. You’ll see
how in Chapter 7.
• How many DOF are there in the human hand?
Looking for More?
• The Robotics Primer Workbook exercises for this chapter are found here:
http://roboticsprimer.sourceforge.net/workbook/Robot_Components
• The Art of Electronics by Paul Horowitz and Winfield Hill, is the definitive
source for electronics for all, from tinkerers to researchers.
• Read Fred Martin’s Robotic Explorations: A Hands-on Introduction to Engi-

neering to learn more about motors, gears, and electronics, as well as ways
of putting those components together with LEGOs to create interesting
and fun robots.
• Sarcos creates the Disney animatronic characters, various robots you have
seen in movies, the Bellagio fountains in Las Vegas, and other sophisti-
cated robots.
• For a nice look at dimensionality, read the classic Flatland: A Romance of

Many Dimensions by Edwin A. Abbott.
5 Move It!
Locomotion
L OCOMOTION Locomotion refers to the way a body, in our case a robot, moves from place to
place. The term comes from the Latin locus for place and the English motion
for movement.
You might be surprised to know that moving around presents all kinds of
challenges. In fact, it is so hard that, in nature, movement requires a signifi-
cant increase in “brain power.” That is why creatures that move, and there-
fore have to contend with avoiding falling over/down/under, collisions, and
being run over, and often are chasing prey and/or running away from preda-
tors, are typically smarter than those that stay put. Compare plants with even
the simplest moving animals.
Most of this book will deal with robot brains, and you will see that moving
the robot around will be the first challenge for those brains. But first, in this
chapter we will talk about the bodies that make locomotion possible.
Many kinds of effectors and actuators can be used to move a robot around,
including:
• Legs, for walking/crawling/climbing/jumping/hopping, etc.
• Wheels, for rolling
• Arms, for swinging/crawling/climbing
• Wings, for flying
• Flippers, for swimming.
Can you think of any others?

48 5 Move It!
While most animals use legs to get around and go places, legged locomo-
tion is a more difficult robotic problem compared with wheeled locomotion.
The reasons for this include:
1. The comparatively larger number of degrees of freedom (DOF); as dis-

cussed in Chapter 4, the more DOF a robot has, the more complicated it is
to control.
2. The challenge of stability; it is harder to stay stable on legs than it is on

wheels, as we will discuss next.
5.1 Stability
S TABILITY Most robots need to be stable, meaning they do not wobble, lean, and fall over
easily, in order to get their job done. But there are different ways of being
S TATIC S TABILITY stable. In particular, there are two kinds of stability: static and dynamic. A
statically stable robot can stand still without falling over; it can be static and
stable. This is a useful feature, but it requires the robot’s body to have enough
legs or wheels to provide sufficient static points of support to keep the robot
stable.
Consider, for example, how you stand up. Did you know that you (along
with all other humans) are not statically stable? That means you don’t stay
upright and balanced without expending some effort and some active con-
trol by your brain. As everyone knows, if you faint, you fall down. Although
standing up may appear effortless, it takes a while to learn; babies take about
a year to master the task. Standing involves using active control of your
muscles and tendons in order to keep your body from falling over. This bal-
ancing act is largely unconscious, but certain injuries to the brain can make
it difficult or impossible, which shows that active control is required.
If instead of having two legs we had three, standing would actually be
much easier, as we could spread our legs out like a tripod. With four legs
it would be even easier, and so on. In general, with more legs (or ground
points), static stability becomes easier. Why do you think that is?
C ENTER OF G RAVITY Here is why: the center of gravity (COG) of any body needs to be above the
area that is covered by the ground points (legs or wheels). When it is, the
body is balanced and stays in the upright position; if the COG is not above
the ground points, it drags the body down, and it starts to fall over. The area
P OLYGON OF S UPPORT covered by the ground points is called the polygon of support. It is called that
because you can draw a projection of the body onto the surface and trace
5.1 Stability 49
Figure 5.1 Components of balancing and stability: the COG and the polygons of
support for a two-legged humanoid and a six-legged robot, shown on flat ground,
and on an incline, which makes the humanoid unbalanced.
the shape or contour of the support points, which will result in a polygon.
Figure 5.1 shows the COG and polygons of support for a humanoid robot
and a six-legged robot, demonstrating how the humanoid is less stable.
As you can see from the figure, the polygon of support for a standing per-
son is not very large. Our feet are relatively small compared with the rest of
our bodies (if they were not, movement would be really hard, as it is when
wearing huge clown shoes), and since we are relatively tall, our COG is quite
high on our bodies. Keeping that COG over the relatively small polygon of
support does not happen without active and trained effort. This is another
reason that it takes babies a long time to learn to walk: they have very large
heads compared with the rest of their bodies, and these heads bring their
COG very high relative to their bodies, making balancing particularly tricky.
No wonder crawling, which provides more points of support, usually comes
first.
In a two-legged robot, such as a humanoid, the polygon is small, and the
COG cannot be easily stably aligned to be within it and keep the robot up-
right. In contrast, in a three-legged robot, with its legs in a tripod organiza-
tion and its body above, static stability is easy. With four legs, it’s even easier,
50 5 Move It!
and a new feature is introduced: three legs can stay on the ground and keep
the robot stable while the fourth can be lifted up, so the body can move.
What happens when a statically stable robot lifts a leg and tries to move?
Does its COG stay within the polygon of support?
It depends on the geometry of the robot’s body and on the number of legs
S TATICALLY S TABLE that stay on the ground. If the robot can walk while staying balanced at all
WALKING times, we call this statically stable walking. A basic assumption of statically
stable walking is that the weight of a leg is negligible compared with that
of the body, so that the total COG of the robot is not much affected by the
movement of the leg. And of course there always need to be enough legs
to keep the body stable. A four-legged robot can only afford to lift only one
leg at a time, since it takes at least three legs on the ground to stay statically
stable. That results in a rather slow walking pace, and one that also takes
a lot of energy. In fact, it is generally the case that statically stable walking,
with any number of legs, while very safe, is also very slow and energy ineffi-
cient. Therefore, as a robot designer, you have to think about how important
continued static stability is to your robot and if it is worth the time and effort
it entails.
What is the alternative to static stability? Does the body have to be
stable and slow, or out of control and falling over?
D YNAMIC S TABILITY In Dynamic stability, the body must actively balance or move to remain sta-
ble; it is therefore called dynamically stable. For example, one-legged hopping
robots are dynamically stable: they can hop in place or to various destina-
tions and not fall over. These robots cannot stop and stay upright, just as
I NVERSE P ENDULUM you could not stop and stay standing on a pogo stick. Balancing one-legged
robots (and people and other objects) is formally called the inverse pendulum
problem, because it is the same as trying to balance a pendulum (or a stick) on
one’s finger, and is equally hard. Pendulums have been studied a great deal
in physics and there are well-understood solutions to the problem of balanc-
ing them. Your brain solves the inverse pendulum problem whenever you
stand, and so must your robot if it is dynamically stable.
Two-legged walking is only slightly easier than one-legged hopping, be-
cause it is also dynamically stable, with one leg being lifted and swung for-
ward while the other is on the ground. Trying to get a robot to walk on two
legs will really help you better understand and appreciate what your brain
is doing all the time without any conscious thought.
As we have seen, people are dynamically stable walkers, while some robots
5.2 Moving and Gaits 51
are statically stable, if they have enough legs. A statically stable robot may
not necessarily remain statically stable while walking. For example, no mat-
ter how many legs a robot has, if it lifts all of them, or almost all of them,
it will become unstable. Therefore, a statically stable robot can use dynam-
ically stable walking patterns, in order to be fast and efficient. In general,
there is a trade-off or compromise between stability and speed of movement,
as we will see next.
5.2 Moving and Gaits

G AIT A gait is the particular way a robot (or a legged animal) moves, including the
order in which it lifts and lowers its legs and places its feet on the ground.
Desirable robot gaits have the following properties:
• Stability: the robot does not fall over
• Speed: the robot can move quickly
• Energy efficiency: the robot does not use a great deal of energy to move
• Robustness: the gait can recover from some types of failures
• Simplicity: the controller for generating the gait is not unwieldy.
Not all of the above requirements can be achieved with all robots. Some-
times safety requirements compromise energy conservation, robustness re-
quirements compromise simplicity, and so on. As we saw in Chapter 4, it
is even possible to design a robot so that it can balance and walk, in certain
circumstances (usually with a push or down a hill) without any motors, as
shown in figure 4.1.
Some numbers of legs and gaits are particularly popular, and can be found
very commonly both in nature and in robotics. For example, most have six
legs and arthropods (invertebrates with segmented bodies) have six or more
(arachnids, better known as spiders, have eight), while the majority of ani-
mals have four. We humans, with our two legs, more complicated walking,
and slower running, are a minority relative to most animals. We are quite
slow.
Six-legged walking is highly robust and therefore common in nature,
and it has served as a popular model in robotics as well. Can you guess
why?
52 5 Move It!
Figure 5.2 Genghis, the popular six-legged walking robot. (Photo courtesy of Dr.
Rodney Brooks)
Figure 5.3 The alternating tripod gait (left) and ripple gait (right) of a six-legged
walking robot. Light circles indicate feet touching the ground.
Six legs allow for multiple walking gaits, both statically and dynamically
T RIPOD G AIT stable. The tripod gait is a statically stable gait in which, as the name implies,
three legs stay on the ground, forming a tripod, while the other three lift and
A LTERNATING T RIPOD move. If the sets of three legs are alternating, the gait is called the alternating
G AIT tripod gait and produces quite efficient walking that can be found in vari-
ous bugs (more specifically, insects and arthropods), including the common
cockroach.
5.3 Wheels and Steering 53
Numerous six-legged robots have been built. Figure 5.2 shows Genghis, a
well-known simple robot that was built over only a summer by two students
at the MIT Mobot Lab, part of what was then the MIT Artificial Intelligence
Lab. Genghis later became a commercial product and a model for NASA
rover prototypes. Its descendant Attila was a much more complicated (many
many more DOF, many more sensors, a bigger brain) and less robust robot.
Simplicity rules!
Almost all six-legged robots are endowed with the alternating tripod gait,
because it satisfies most of the desirable gait properties listed above. In this
gait, the middle leg on one side and two nonadjacent legs on the other side of
the body lift and move forward at the same time, while the other three legs
remain on the ground and keep the robot statically stable. Then the tripod
alternates to the other side of the body, and so on.
R IPPLE G AIT What happens if a robot has more than six legs? The alternating tripod
gait can still be used, in the form of the so-called ripple gait, because it ripples
down the length of the body. Arthropods such as centipedes and millipedes,
which have many more than six legs, use the ripple gait. Figure 5.3 shows
the alternating tripod and ripple gaits of a six-legged robot.
You may have noticed that we have been talking about statically stable
walking even though we said earlier that such walking is not the most ef-
ficient. Indeed, when bugs run, they usually use faster, dynamically stable
gaits. They may even become airborne at times, taking advantage of glid-
ing for brief periods of time. Such extreme examples of dynamic stability
demonstrate how speed and energy efficiency can be gained at the expense
of stability and controller simplicity.
Controlling a robot that can balance itself on one or two legs, or can glide
through the air and land safely, is a complicated control problem. Thus,
not surprisingly, balance and stability are complex problems in robotics, and
many robots are designed to avoid dealing with them if it is not necessary.
That is why most existing mobile robots have wheels or six legs: to sim-
plify locomotion. Robotics research, however, actively studies more complex
and interesting modes of locomotion, including hopping, crawling, climbing,
swimming, gliding, flying, and more.
5.3 Wheels and Steering
You might have heard the claim that wheels do not appear in nature, but that
is actually not quite correct. Wheellike structures do appear in certain bac-
54 5 Move It!
teria, but are definitely very rare in biological locomotion when compared
with legs, wings, and fins. Wheels are, however, more efficient than legs,
so you might wonder why animals do not have them. Evolution favors lat-
eral symmetry, meaning it evolves bodies that have two (or more) matched
sides/parts, using matched sets of genes, and so legs were easier to evolve,
and they work quite well. As a result, we find them everywhere in nature;
based on population size, bugs are the most abundant macro species (mean-
ing relatively “big”) and they start with six legs and go up from there. The
most abundant species of all are bacteria (which are comparatively small, so
micro), which dominate the living world by sheer numbers, but until micro-
robotics and nano-robotics really take off (see Chapter 22), we can put those
aside.
Because of their efficiency and their comparative simplicity of control, wh-
eels are the locomotion effector of choice in robotics. Wheeled robots (as
well as almost all wheeled mechanical devices, such as cars) are built to be
statically stable. Although most wheels do not stray too far from the basic
design, they can be constructed with as much variety and innovative flair
as legs. Wheels can vary in size and shape, can use simple tires or complex
tire patterns or tracks, and can even contain wheels within cylinders within
other wheels spinning in different directions to provide various types of lo-
comotion properties.
Robots with wheels are usually designed to be statically stable, which sim-
H OLONOMIC plifies control. However, they are not necessarily holonomic, meaning they
cannot control all of their available degrees of freedom (DOF), as we dis-
cussed in detail in Chapter 4. In particular, most simple mobile robots have
two or four wheels, and in both cases they are nonholonomic. A popular and
efficient design for wheeled robots involves two wheels and a passive caster
for balance, as shown in figure 5.4.
Having multiple wheels means there are multiple ways in which those
wheels can be controlled. Basically, multiple wheels can move either together
D IFFERENTIAL D RIVE or independently. The ability to drive wheels separately and independently,
D IFFERENTIAL through the use of separate motors, is called a differential drive. Analogously,
S TEERING being able to steer wheels independently is called differential steering.
Consider the benefits of a differential drive in the basic two-wheels-and-
caster robot design. If the two wheels are driven at the same speed, the robot
moves in a straight line. If one wheel (say the left) is driven faster than the
other, the robot turns (in this case to the right). Finally, if the wheels are
driven in the opposite directions of each other but at the same speed, the
Figure 5.4 A popular drive mechanism for simple robots, consisting of two differ-
entially steerable driven wheels and a passive caster for balance.
robot turns in place. These abilities allow the robot to maneuver along com-
plicated paths in order to do its job.
5.4 Staying on the Path vs. Getting There

In robot locomotion, we may be concerned with:
• Getting the robot to a particular location
• Having the robot follow a particular path (also called trajectory).
Following an arbitrary given path or trajectory is harder than having to

get to a particular destination by using any path. Some paths are impossible
to follow for some robots because of their holonomic constraints. For others,
some paths can be followed, but only if the robot is allowed to stop, change
directions (in place or otherwise), and then go again, just as we discussed
with parallel parking in Chapter 4.
A large subarea of robotics research deals with enabling robots to follow
arbitrary trajectories. Why?
56 5 Move It!
Because there are various applications where this ability is necessary, rang-
ing from autonomous car driving to brain surgery, and because it is not at all
T RAJECTORY AND easy to achieve. Trajectory planning, also called motion planning, is a com-
M OTION P LANNING putationally complex process which involves searching through all possible
trajectories and evaluating them, in order to find one that will satisfy the
requirements. Depending on the task, it may be necessary to find the very
O PTIMAL T RAJECTORY best (shortest, safest, most efficient, etc.), so-called optimal trajectory. Since
robots are not just points, their geometry (shape, turning radius) and steering
mechanism (holonomic properties) must be taken into account. Trajectory
planning is used in mobile robots, in two dimensions, and in robot arms,
in three dimensions, where the problem becomes even more complex. We
talked a bit about this in Chapter 4 and will go into more detail in the next
chapter.
Depending on their task, practical robots may not be so concerned with
following specific trajectories as with just getting to the goal location. The
ability to get to the goal is quite a different problem from planning a particu-
lar path, and is called navigation. We will learn more about it in Chapter 19.
To Summarize
• Moving around takes brains, or at least some nontrivial processing.
• Stability depends on the robot’s geometry, and in particular the position

of its COG and polygon of support. Stability can be static or dynamic.
• Static stability is safe but inefficient, dynamic stability requires computa-

tion.
• The number of legs is important. Two-legged walking is hard and slow;

walking start to get easier with four legs, and much more so with six or
higher.
• Alternating tripod and ripple gaits are popular when six or more legs are
available.
• Wheels are not boring, and drive control is not trivial. There are many
wheel designs and drive designs to choose from.
• Differential drive and steering are the preferred options in mobile robotics.
• Following a specific path or trajectory is hard, as is computing a specific

path/trajectory that has particular properties (shortest, safest, etc.).
• Getting to a destination is not the same as following a specific path.
Food for Thought
• How does an automated Web path planner, such as Mapquest or Google

Maps, which finds a route for, say, a trip from your house to the airport,
find the optimal path?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Locomotion
• For some really exciting and inspiring work about insect and animal lo-
comotion, including flying roaches and sticky geckos, see the research
of Prof. Robert J. (Bob) Full at University of California, Berkeley, in his
Polypedal Lab.
• For very interesting research on stick insects and the gaits they use nor-
mally or when their legs are out of order and/or when put on obstacle
courses, see the work of Prof. Holk Cruse at the University of Bielefeld in
Germany.
6 Grasping at Straws
Manipulation
In the last chapter we learned how robots get around and go places. In this
chapter we will learn what they can do once they get there.
M ANIPULATOR A robotic manipulator is an effector. It can be any type of gripper, hand,
arm, or body part that is used to affect and move objects in the environment.
M ANIPULATION Manipulation, then, refers to any goal-driven movement of any type of ma-
nipulator.
Manipulators typically consist of one or more links connected by joints,
M ANIPULATOR L INKS and the endeffector. Manipulator links are the individual components of the
manipulator which are controlled independently. If we take your arm as an
example, your upper arm would be one link and your lower arm/forearm
another. Links are connected by joints. We’ll talk about joints a bit later in
this chapter. First, let’s start from the end, at the endeffector.
6.1 Endeffectors
E NDEFFECTOR The endeffector is the part of the manipulator which affects the world. For
example, on a hand it may be the finger that pokes, on a pointer it may be
the tip that marks a spot, on a leg it may be a foot that kicks a ball. It is
usually the extreme, final, end part of the manipulator, the part that is used
to do the work (at least most of the time; the entire manipulator or large parts
of it may also be used, see Food for Thought below for an example).
In contrast to locomotion, where the body of the robot moves itself to get
to a particular position and orientation, a manipulator typically moves itself
in order to get the endeffector to the desired three-dimensional (3D) posi-
tion and orientation. Most traditional manipulation is concerned with the
problem of bringing the endeffector to a specific point in 3D space. This is a
surprisingly difficult problem. Why do you think that is?
60 6 Grasping at Straws
Because the endeffector is attached to the arm, and the arm is attached to
the body, manipulation involves not only moving the endeffector, but also
J OINT L IMIT considering the whole body. The arm must move so as to avoid its own joint
limit (the extreme of how far the joint can move), the body, and any other
obstacles in the environment. As a result, the path followed by the hand,
and thus the arm, and perhaps even the whole body, to achieve the desired
task (getting the endeffector to the desired position) may be quite complex.
More importantly, computing that path is a complex problem that involves
thinking about space with the robot’s body, manipulator, and task in mind.
F REE S PACE Specifically, it involves computing the free space of the particular manipulator,
body, and task (the space in which movement is possible), and then searching
that space for paths that get to the goal.
Because all of the above is far from easy, autonomous manipulation is very
challenging. The idea of autonomous manipulation was first used in teleop-
eration, where human operators moved artificial arms to handle hazardous
materials.
6.2 Teleoperation
T ELEOPERATION As we learned in Chapter 1, teleoperation means controlling a machine from a

R EMOTE C ONTROL distance. Teleoperation is related to remote control, but the two terms are not
used in the same way in robotics. In particular, teleoperation usually means
controlling a complex manipulator or rover (as in teleoperated NASA rovers
used for space exploration), while remote control is usually used for control
of simple mechanisms, such as toy cars. The term remote control is not used
in robotics, although its literal meaning is basically the same as teleoperation.
As was mentioned in Chapter 1, puppeteering is a form of teleoperation.
It’s no surprise that puppeteering is an art that requires a great deal of skill.
Similarly, teleoperating complicated manipulators and robots is also a com-
plex task that requires a great deal of skill. The challenge stems from the
following:
• Complexity of the manipulator: The more degrees of freedom (DOF) there

are in the manipulator being controlled, the harder it is to control.
• Constraints of the interface: Initially, joysticks were the most common

interface between a teleoperator and the arm being controlled. But using
a joystick to control a complicated mechanical arm, such as a seven DOF
humanlike one, requires a great deal of training and concentration.
6.2 Teleoperation 61
Figure 6.1 A complex teleoperated manipulator robot used in robot-assisted

surgery in some hospitals. (Photo ©[2007] Intuitive Surgical, Inc.)
• Limitations of sensing: It is difficult to control an arm without directly

seeing what it touches and where it moves, as well as feeling the touch
and the resistance, and so on.
Teleoperation has been used with great success in some very complex do-
R OBOT-A SSISTED mains, such as robot-assisted surgery, where it has been applied to hip surgery,
S URGERY cardiac (heart) surgery, and even brain surgery. In some cases, as in cardiac
surgery, the robot moves inside the body of the patient to cut and suture
(crudely put, to sew up), while the surgeon controls it from outside the body
with the movements of his or her fingers and hands, which are connected to
wires that transmit the signal to the robot’s manipulators. The major benefit
of such minimally invasive surgery, made possible by teleoperation, is that
instead of having to cut the patient’s chest open and crack some ribs to get to
the heart, only three small holes are all it takes to push in the manipulators
and perform the surgery. As a result, there are fewer post-operation prob-
lems and infections, and the patient recovers much faster. Figure 6.1 shows
a teleoperated robot used in robot-assisted surgery.
But even for well-designed teleoperation interfaces, it takes quite a bit of
training for the human operator to learn to control the manipulator(s), for the
reasons given above. One popular way of simplifying this problem (some-
what at least) is to use an exoskeleton.
E XOSKELETON An exoskeleton (from the Greek exo meaning “outside”), literally means an
outside skeleton. In biological systems, it is in the form of a hard outer struc-
ture, such as the shell of an insect, that provides protection or support.
In robotics, an exoskeleton is a structure that a human wears and controls;
it provides additional strength or, in teleoperation, allows for sensing the
movements and the forces the human produces in order to use those to con-
trol a robot. Exoskeletons do not qualify as robots according to our definition
(see Chapter 1) because they do not make their own decisions or act on them;
they just provide added strength or dexterity to human operators. Therefore,
they are merely shells, completely operated by a human and typically with-
out sensors or controllers.
Teleoperation can be used in combination with autonomy, allowing a hu-
man user to influence the robot without having to control it continuously.
In summary, teleoperation is one means of simplifying manipulator con-
trol, but it does not help solve the problem of autonomous manipulator con-
trol, so let us return to that.
6.3 Why Is Manipulation Hard?

Why is manipulation so hard? As was the case in locomotion, there is typi-
cally no direct and obvious connection between what the effector needs to do
in physical space and what the actuator does to move that effector in order
to get the job done.
K INEMATICS The correspondence between actuator motion and the resulting effector
motion is called kinematics. Kinematics consist of the rules about the structure
of the manipulator, describing what is attached to what.
J OINTS The various parts (links) of a manipulator are connected by joints. The
most common joint types are:
R OTARY J OINT • Rotary (ball-and-socket), providing rotational movement around a fixed

axis
P RISMATIC J OINT • Prismatic (like a piston), providing linear movement.
Robot manipulators can have one or more of each type of joint. Robotic
joints typically provide one controllable DOF, and in order to have maximum
generality of movement, a separate actuator (typically a motor, such as the
one you saw in figure 4.2 in Chapter 4) is used to move each of the DOF. This
Figure 6.2 An anthropomorphic (humanlike) robot arm.
immediately tell you that if you want to have many DOF in a manipulator,
you will need a lot of physical machinery to make up the mechanism and
move it in order to achieve the desired controlled movement. (And then
you’ll need batteries to power those motors, which are heavy, so the motors
now have to be stronger to lift the manipulators that carry them, which takes
power, and so on. Designing complicated robots is fun but not easy.)
Consider again the human arm (not including the hand) with its seven
DOF. Note that the arm itself has only three joints: the shoulder, the elbow,
and the wrist, and those three joints control the seven DOF. This means some
of the joints control more than one DOF; specifically, the shoulder joint has
three DOF, as does the wrist. How does that work? Looking more closely
at human anatomy, we find that the shoulder joint is a ball-and-socket joint
(just like the hip) and is actuated by large muscles. It turns out that ball-
and-socket joints are very difficult to create in artificial systems, not only
robots but also realistically modeled animated characters in computer games.
The wrist, unlike the shoulder, is a smaller joint controlled by a collection
of muscles and ligaments in the arm. It is also complex, but in an entirely
different way from the shoulder.
Trying to get a robot to look and move like an animal (any animal, from a
rat to a human) is very hard, at least in part because on robots we use motors
Figure 6.3 An anthropomorphic (humanlike) robot hand.
to drive mechanical links connected by rotary joints. In contrast, animals

use muscles, which are linear actuators, and are comparatively much lighter,
springier, more flexible, and stronger. There is ongoing research into artificial
muscles, but it is in early stages and far from ready for use in robots. Most
manipulator actuation is done with motors.
A NTHROPOMORPHIC Simulating the function of an anthropomorphic, meaning humanlike in shape
(from the Greek anthropo for “human,” and morph for “shape”), manipulator
in a robot is challenging. An anthropomorphic robot arm is shown in fig-
ure 6.2.
The human arm is complicated enough, but it is simple compared with the
human hand, an extremely complex and versatile manipulator. The hand has
a very large number of joints and, therefore, a very large number of DOF. It
also has a very compact design; if you were to construct a robotic hand with
the same number of DOF, you would need a lot of space to fit all the motors
in. Or you could move the motors away from the hand, but they would have
to go somewhere: up the arm or elsewhere on the body. An example of an
anthropomorphic robot hand is shown in figure 6.3.
Humans use their general-purpose hands to manipulate tools, such as knives
and forks and screwdrivers and hammers, for getting specific jobs done. In
contrast, robot manipulators are typically specialized, and may already have
Figure 6.4 NASA’s Robonaut humanoid torso robot. (Photo courtesy of NASA)
an attached tool at the endpoint, such as a welding gun, a paint sprayer, or

a screwdriver. More general-purpose robots, on the other hand, may have
multi-purpose grippers that allow them to grab and use various tools. Fi-
nally, some complex humanlike robotic hands have been developed for a
large variety of possible tasks. NASA’s Robonaut humanoid torso robot,
shown in figure 22.6, is a good example, with humanlike arms and hands
for fixing problems on the space shuttle and space station.
Regardless of the task at hand (pun intended), in order to control a robot
manipulator we have to know its kinematics: what is attached to what, how
many joints there are, how many DOF for each joint, and so on. All of these
properties can be stated formally, allowing us to use mathematics in order to
M ANIPULATION solve manipulation problems, problems about where the endpoint is relative to
P ROBLEMS the rest of the arm, and how to generate paths for the manipulator to follow
in order to get the job done.
To move the endeffector of a manipulator to a desired point, we typically
need to compute the angles for the joints of the whole manipulator. This con-
version from a Cartesian (x,y,z) position of the endpoint (e.g., a fingertip) and
I NVERSE K INEMATICS the angles of the whole manipulator (e.g., an arm) is called inverse kinematics.
The name refers to the fact that this is the opposite of the simpler process of
figuring out where the endpoint of the manipulator is given the joint angles
for all of the joints. That was kinematics, presented earlier in this chapter. The
process of computing inverse kinematics is expensive (computationally in-
tense). Why is that so? To answer that question we would need to get into
some math, and it may be best to read about it in one of the sources listed at
the end of this chapter. Suffice it to say that it is necessary, and important,
since enabling robots to move their arms to the right place at the right time
is necessary for a great many uses, from shaking hands to surgery (described
earlier).
After kinematics, the next aspect of control in general, and manipulation
D YNAMICS in particular, for us to consider is dynamics. Dynamics refers to the proper-
ties of motion and energy of a moving object. Since robots move around and
expend energy, they certainly have dynamics; and the faster they move, the
more significant their dynamics, and therefore the more impact dynamics
have on the robot’s behavior. For example, the behavior of a slow-moving
mouse-like maze-solving mobile robot is not very strongly impacted by its
dynamics, while the behavior of a fast-moving tennis-ball-juggling torso robot
certainly is. No surprise there.
Analogously to direct and inverse kinematics, manipulation also involves
the computation of direct and inverse dynamics. These computations are
even more complicated and computationally expensive. You can learn more
about this rather mathematical topic from the sources listed below (and may
be glad for not having the math included here).
Because the steps involved in robot manipulation are challenging, prob-
lems such as reaching and grasping constitute entire subareas of robotics
G RASP P OINTS research. These fields study topics that include finding grasp points (where
to put the fingers relative to the center of gravity, friction, obstacles, etc.),
the force/strength of grasps (how hard to squeeze so the robot neither drops
C OMPLIANCE nor crushes the object), compliance (yielding to the environment required for
tasks with close contacts, such as sliding along a surface), and performing
highly dynamic tasks (juggling, catching), to name just a few. Manipulation
is of particular interest as robots are poised to enter human environments in
order to assist people in a variety of tasks and activities. For this, they must
be able to interact effectively with and manipulate a variety of objects and
situations, a tall order indeed for robot designers and programmers.
To Summarize
• Robot manipulators consist of one or more links, joints connecting the

links, and an endeffector.
• There are two basic types of joints (rotary and prismatic) and a variety of
types of endeffectors (grippers, hands, pointers, tools, etc.).
• The process of manipulation involves trajectory planning, kinematics, and

dynamics, and it is quite computationally complex.
• Teleoperation can be used for manipulation, in order to avoid autonomous

control.
Food for Thought
• How many DOF are there in the human hand? Can you control each of
them independently?
• Which of the two joint types we discussed, rotational and prismatic, is

more commonly found in biological bodies? Can you think of examples
of specific animals?
• Are astronaut suits exoskeletons? What about a lever-controlled backhoe?
• Imagine an exoskeleton which has its own sensors, makes its own deci-
sions, and acts on them. This is definitely a robot, yet it is also controlled
by a human and attached to the human’s body. Can you imagine where
this could be useful?
• Robotic dogs have been used to play soccer. Because they are not designed
for that task, they move slowly and have trouble aiming and kicking the
ball (not to mention finding it, but that’s not part of manipulation; it be-
longs to sensing, discussed in the next few chapters). Some of the best
robot soccer teams have come up with a clever idea: instead of having
the dogs walk as they normally do, they lower them so that the two rear
legs are driving and steering the robot, while the two front legs are on
the ground, used to grasp the ball and kick it, which is much easier. This
turns out to work very well, even if it looks rather silly. What are the end-
effectors of the dogs in that case? Are the dogs now purely mobile robots?
Looking for More?
Here are some good resources for reading more about the topics we men-
tioned in this chapter:
• Robot Modeling and Control by Mark Spong
• Modeling and Control of Robot Manipulators by L. Sciavicco, B. Siciliano
• Principles of Robot Motion: Theory, Algorithms, and Implementations by H.

Choset, K. M. Lynch, S. Hutchinson, G. Kantor, W. Burgard, L. E. Kavraki
and S. Thrun
• Planning Algorithms by Steven M. LaValle

7 What’s Going On?
Sensors
Knowing what is going on is a requirement for survival, not to mention for

intelligent behavior. If a robot is to achieve anything at all, it must be able
to sense the state of its own body (its internal state; see Chapter 3) and the
state of its immediate environment (external state; also see Chapter 3). In
fact, as we learned in Chapter 1, for a robot to be a robot, it must be able
to sense. In this chapter you will find out how the robot’s ability to sense
directly influences its ability to react, achieve goals, and act intelligently.
A robot typically has two types of sensors based on the source of the infor-
mation it is sensing:
1. Proprioceptive sensors: These sensors perceive elements of the robot’s inter-

nal state, such as the positions of the wheels, the joint angles of the arms,
and the direction the head is facing. The term comes from the Latin word
proprius meaning “one’s own” (also featured in “proprietor,” “property,”
P ROPRIOCEPTION and “appropriate”). proprioception is the process of sensing the state of
one’s own body. It applies to animals as much as it does to robots.
2. Exteroceptive sensors: These sensors perceive elements of the state of the

external world around the robot, such as light levels, distances to ob-
jects, and sound. The term comes from the Latin word extra meaning
“outside” (also featured in “extrasensory,” “extraterrestrial,” and “extro-
E XTEROCEPTION verted”). Exteroception is the process of sensing the world around the
robot (including sensing the robot itself).
Proprioceptive sensors and exteroceptive sensors together constitute the

P ERCEPTUAL S YSTEM perceptual system of a robot, as shown in figure 7.1. However, one of the main
challenges of robotics is that sensors themselves do not provide convenient
state information to the robot. For example, sensors do not say: “There is
70 7 What’s Going On?
Figure 7.1 Proprioceptive and exteroceptive sensors on a simple robot.
a blue chair on your left, and your grandmother Zelda is sitting in it and
looking uncomfortable.” Instead, sensors may tell the robot the light levels
and colors in its field of view, whether it is touching something in a particular
area, whether there is a sound above some threshold, or how far away the
nearest object is, and so on.
Rather than being magic providers of all the information the robot could
S ENSORS possibly need, sensors are physical devices that measure physical quantities.
Table 7.1 considers some devices and the quantities they measure.
As table 7.1 illustrates, the same physical property may be measurable
with more than one type of sensor. This is very convenient, as we will find
out, since sensor information is prone to noise and errors, so acquiring infor-
mation from multiple sensors can provide improved accuracy.
Sensor noise and errors, which are inherent in physical measurement and
cannot be avoided, contribute to one of the major challenges of robotics: un-
U NCERTAINTY certainty. Uncertainty refers to the robot’s inability to be certain, to know for
sure, about the state of itself and its environment, in order to take absolutely
optimal actions at all times. Uncertainty in robotics comes from a variety of
sources, including:
Physical Property → Sensing Technology 71
Contact → bump, switch
Distance → ultrasound, radar, infra red
Light level → photocells, cameras
Sound level → microphones
Strain → strain gauges
Rotation → encoders and potentiometers
Acceleration → accelerometers and gyroscopes
Magnetism → compasses
Smell → chemical sensors
Temperature → thermal, infra red
Inclination → inclinometers, gyroscopes
Pressure → pressure gauges
Altitude → altimeters
Table 7.1 Some sensors and the information they measure.
• Sensor noise and errors
• Sensor limitations
• Effector and actuator noise and errors
• Hidden and partially observable state
• Lack of prior knowledge about the environment, or a dynamic and chang-

ing environment.
Fundamentally, uncertainty stems from the fact that robots are physical
mechanisms that operate in the physical world, the laws of which involve
unavoidable uncertainty and lack of absolute precision. Add to that imper-
fect sensor and effector mechanisms and the impossibility of having total
and perfect knowledge, and you can see why robotics is hard: robots must
survive and perform in a messy, noisy, challenging real world. Sensors are
the windows into that world, and in robotics those windows are, so far, quite
small and hard to see through, metaphorically speaking.
We can think of various robot sensors in terms of the amount of infor-
mation they provide. For example, a basic switch is a simple sensor that
B IT provides one bit of information, on or off. Bit, by the way, refers to the funda-
mental unit of information, which has two possible values, the binary digits
0 and 1; the word comes from “b(inary) (dig)it.” In contrast, a simple camera
lens (a vision sensor) is stunningly rich in information. Consider a standard
P IXEL camera, which has a 512 X 512 pixel lens. A pixel is the basic element of the
image on the camera lens, computer, or TV screen. Each of these 262,144 pix-
els may, in the simplest case, be either black or white, but for most cameras it
will provide much more range. In black-and-white cameras, the pixels pro-
vide a range of gray levels, and in color cameras they provide a spectrum of
R ETINA colors. If that sounds like a lot of information, consider the human retina, the
part of your eye that “sees” and passes information to the brain. The retina is
a light-sensitive membrane with many layers that lines the inner eyeball and
is connected to the brain via the optic nerve. That amazing structure consists
of more than a hundred million photosensitive elements; no wonder robots
are quite a way behind biological systems.
Although we just mentioned this, it is worth repeating:
Sensors do not provide state. They provide raw measured quantities,
which typically need to be processed in order to be useful to a robot.
The more information a sensor provides, the more processing is needed.
Consequently, it takes no brains at all to use a switch, but it takes a great deal
of brains (in fact, a large portion of the human brain) to process inputs from
vision sensors.
Simple sensors that provide simple information can be employed almost
directly. For example, consider a typical mobile robot that has a switch at
the front of the body. When the switch is pushed, the robot stops, because it
"knows" that it has hit something; when the switch is not pushed, the robot
keeps moving freely. Not all sensory data can be processed and used so
simply, which is why we have brains; otherwise, we would not need them.
In general, there are two ways in which sensory information can be treated:
1. We can ask the question: "Given that sensory reading, what should I do?"
or
2. We can ask the question: "Given that sensory reading, what was the world
like when the reading was taken?"
Simple sensors can be used to answer the first question, which is about
A CTION I N THE action in the world, but they do not provide enough information to answer
W ORLD the second question, which is about reconstruction of the world. If the switch
R ECONSTRUCTION
on the robot indicates that it has hit something, that is all the robot knows;
it cannot deduce anything more, such as the shape, color, size, or any other
information about the object it has contacted.
At the other extreme, complex sensors such as vision provide a great deal
more information (a lot of bits), but also require a great deal more processing
to make that information useful. They can be used to answer both ques-
tions asked above. By providing more information, a sensor can allow us to
attempt to reconstruct the world that produced the reading, at least with re-
spect to the particular measured quantity. In a camera image we can look for
lines, then objects, and finally try to identify a chair or even a grandmother
in the image. For an idea of what it takes to do that, see Chapter 9.
The problem of going from the output of a sensor to an intelligent response
S IGNAL - TO -S YMBOL is sometimes called the signal-to-symbol problem. The name comes from the
fact that sensors produce signals (such as voltage levels, current, resistance,
etc.), while an action is usually based on a decision involving symbols. For
example, the rule: "If grandmother is there and is smiling, approach her,
otherwise go away" can be easily programmed or encoded with symbols for
grandmother and smiling, but it is a lot more complicated to write the rule
using raw sensor data. By using symbols, we can make information abstract
and not sensor-specific. But getting from a sensory output (of any sensor)
to such an abstract symbolic form (or encoding) of information in order to
make intelligent decisions is a complex process. While it may seem obvious,
this is one of the most fundamental and enduring challenges in robotics.
Since sensors provide signals and not symbolic descriptions of the world,
they must be processed in order to extract the information the robot needs.
S ENSOR This is usually called sensor preprocessing because it comes before anything
P REPROCESSING else can be done in terms of using the data to make decisions and/or take
action. Sensor (pre)processing can be done in different ways and draws on
methods from signal processing (a branch of electrical engineering) and com-
putation.
7.1 Levels of Processing
Suppose that your robot has a switch sensor to detect bumping into obstacles,
as described above. To figure out if the switch is open or closed, you need to
E LECTRONICS measure the voltage going through the circuit. That is done with electronics.
Now suppose your robot has a microphone sensor for recognizing a voice.
Besides the electronic processing, it will need to separate the signal from any
background noise and then compare it with one or more stored voices in
S IGNAL P ROCESSING order to perform recognition. That is done with signal processing.
Next suppose your robot has a camera for finding your grandmother in
the room. Besides the electronics and signal processing, it will need to find
the objects in the room, then compare them against a large database in order
Figure 7.2 Levels of sensory processing.
C OMPUTATION to try to recognize the grandmother. That is done with computation.

As you can see, sensory information processing is challenging and can
be computationally intensive and time consuming. Of the above processes,
computation is the slowest but most general. For any specific problem, a sys-
tem can be designed to solve it at a lower, and thus faster, level. For example,
although computation is usually necessary for processing visual images, spe-
cialized microprocessors, so-called "vision chips" have been developed that
are designed for particular vision-based tasks, such as recognizing particular
faces or fruits or engine parts. They are fast but very specialized; they cannot
be used to recognize anything else.
Given that a great deal of processing can be required for perception (as
shown in figure 7.2), we can already see why a robot needs some type of
brain. Here is what a robot needs in order to process sensory inputs:
• Analog or digital processing capabilities (i.e., a computer)

• Wires to connect everything together
• Support electronics to go with the computer
• Batteries to provide power for the whole thing.

This means that perception requires:
• Sensors (power and electronics)
• Computation (more power and electronics)
• Connectors (to connect it all).
It is generally not a good idea to separate what the robot senses, how it
senses it, how it processes it, and how it uses it. If we do that, we end up with
a large, bulky, and ineffective robot. Historically, perception has been studied
and treated in isolation, and typically as a reconstruction problem, assuming
that a robot always has to answer the second question posed above. None
of these approaches has resulted in effective methods that robots can use to
better sense the world and go about their business.
Instead, it is best to think about the what, why, and how of sensing as a
single complete design, consisting of the following components:
• The task the robot has to perform
• The sensors best suited for the task
• The mechanical design most suited to enable the robot to get the sensory
information necessary to perform the task (e.g., the body shape of the
robot, the placement of the sensors, etc.).
Robotics researchers have figured out these important requirements of ef-

fective perception, and have been exploring various methods, including:
• Action-oriented perception (also called “active sensing”): instead of trying to

reconstruct the world in order to decide what to do, the robot can use the
knowledge about the task to look for particular stimuli in the environment
and respond accordingly. As an example, it is very hard for robots to
recognize grandmothers in general, but it is not nearly as hard to look
for a particular color pattern of a grandmother’s favorite dress, perhaps
combined with a particular size and shape, and speed of movement.
A clever psychologist named J. J. Gibson wrote (in 1979) about the idea
that perception is naturally biased by what the animal/human needs to
do (i.e., action), and is influenced by the interaction between the animal
and its environment; it is not a process that retrieves absolute “truth”
about the environment (i.e., reconstruction). This idea has been quite pop-
ular in action-oriented perception.
• Expectation-based perception: Use knowledge about the robot’s environ-

ment to help guide and constrain how sensor data can be interpreted. For
example, if only people can move in the given environment, we can use
motion by itself to detect people. This is how simple burglar alarms work:
they don’t recognize burglars, or even people, just detect movement.
• Task-driven attention: Direct perception where more information is needed

or likely to be provided. Instead of having the robot sense passively as
it moves, move the robot (or at least its sensors) to sense in the direction
where information is most needed or available. This seems very obvious
to people, who unconsciously turn their heads to see or hear better, but
most robots still use fixed cameras and microphones, missing the oppor-
tunity to perceive selectively and intelligently.
• Perceptual classes: Divide up (partition) the world into perceptual cate-

gories that are useful for getting the job done. That way, instead of being
confronted with an unwieldy amount of information and numerous pos-
sibilities, the robot can consider manageable categories it knows how to
handle. For example, instead of deciding what to do for every possible
distance to an obstacle, the robot may have only three zones: too close, ok
but be on the alert, and not to worry. See Chapter 14 for an example of
just such a robot.
The idea that sensor function (what is being perceived) should decide sen-
sor form (where the sensor should be, what its shape should be, if/how it
should move, etc.) is employed cleverly in all biological systems. Natu-
ral evolved sensors have special geometric and mechanical properties well
suited for the perceptual tasks of the creature that "wears" them. For exam-
ple, flies have complex compound, faceted eyes, some birds have polarized
light sensors, some bugs have horizon line sensors, humans have specially
shaped ears that help us figure out where the sound is coming from, and so
on. All of these, and all other biological sensors, are examples of clever me-
chanical designs that maximize the sensor’s perceptual properties, its range
and accuracy. These are very useful lessons for robot designers and program-
mers.
As a robot designer, you will not get the chance to make up new sensors,
but you will always have the chance (and indeed the need) to design inter-
esting ways of using the sensors you have at your disposal.
Here is an exercise: How would you detect people in an environment?

Remember the lessons from this chapter so far: Use the interaction with
the world and keep in mind the task.
The obvious answer is to use a camera, but that is the least direct solution
to the problem, as it involves a great deal of processing. You will learn more
about this in Chapter 9. Other ways of detecting people in the environment
include sensing:
• Temperature: Search for temperature ranges that correspond to human

body temperature
• Movement: If everything else is static, movement means people
• Color: Look for a particular range of colors corresponding to people’s skin

or their clothes or uniforms
• Distance: If an otherwise open distance range becomes blocked, there is

likely a moving human around.
The above are just some ways of detecting people by using sensors that
are simpler than vision and require less processing. They are not perfect,
but compared with vision, they are fast and easy. Often these alone, or in
combination, are enough to get the task done, and even when vision sensors
are available, other sensory modalities can improve their accuracy. Consider
burglar alarm sensors again: they sense movement through temperature
changes. While they could confuse a large dog with a human, in indoor en-
vironments canine (dog) burglars are rare, so the sensors tend to be perfectly
suited to the task.
Now let’s do another exercise: how would you measure distance to an
object?
Here are some options:
• Ultrasound sensors provide distance measurements directly (time of flight)
• Infra red sensors can provide it through return signal intensity
• Two cameras (i.e., stereo) can be used to compute distance/depth
• A camera can compute distance/depth by using perspective (and some

assumptions about the structure of the environment)
• A laser and a fixed camera can be used to triangulate distance

• A laser-based, structured light system, can overlay a grid pattern over the
camera image and the processing system can use the distortions in that
pattern to compute distance.
These are just some of the available means of measuring distance, and, as
you saw, distance is one measure that can be used to detect other objects and
people.
S ENSOR F USION Combining multiple sensors to get better information about the world is
called sensor fusion.
Sensor fusion is not a simple process. Consider the unavoidable fact that
every sensor has some noise and inaccuracy. Combining multiple noisy and
inaccurate sensors therefore results in more noise and inaccuracy and thus
more uncertainty about the world. This means some clever processing has
to be done to minimize the error and maximize accuracy. Furthermore, dif-
ferent sensors give different types of information, as you saw above in the
example of detecting people. Again, clever processing is necessary to put the
different types of information together in an intelligent and useful way.
As usual, nature has an excellent solution to this problem. The brain pro-
cesses information from all sensory modalities - vision, touch, smell, hearing,
sound - and a multitude of sensors. Eyes, ears, and nose are the obvious
ones, but consider also all of your skin, the hairs on it, the strain receptors
in your muscles, the stretch receptors in your stomach, and numerous oth-
ers sources of body awareness (proprioception) you have at your disposal,
consciously or otherwise. A great deal of our impressive brain power is in-
volved in processing sensory information. Therefore it is not surprising that
this is a challenging and important problem in robotics as well.
To Summarize
• Sensors are the robot’s window into the world (through exteroception) as
well as into its own body (through proprioception).
• Sensors do not provide state, but instead measure physical properties.
• Levels of sensor processing can include electronics, signal processing, and

computation. Simple sensors require less processing, and thus less asso-
ciated hardware and/or software.
• The what, why, and how of robot sensing, the form and the function,
should be considered as a single problem.
• Nature gives us ample examples of clever sensor form and function.
• Action-oriented perception, expectation-based perception, focus of atten-

tion, perceptual classes, and sensor fusion can all be used to improve the
availability and accuracy of sensor data.
• Uncertainty is a fundamental and unavoidable part of robotics.
Food for Thought
• Uncertainty is not much of a problem in computer simulations, which is

why simulated robots are not very close to the real, physical ones. Can
you figure out why?
• Some robotics engineers have argued that sensors are the main limiting
factor in robot intelligence: if only we had more, smaller, and better sen-
sors, we could have all kinds of amazing robots. Do you believe that is
all that’s missing? (Hint: If that were so, wouldn’t this book be much
thinner?)
• Being able to sense the self, being self-aware, is the foundation for con-
sciousness. Scientists today still argue about what animals are conscious,
and how that relates to their intelligence, because consciousness is a nec-
essary part of higher intelligence of the kind people have. What do you
think will it take to get robots to be self-aware and highly intelligent? And
if some day they are both, what will their intelligence be like, similar to
ours or completely different?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Sensors
• Jerome James Gibson (1904-1979) is considered one of the most important

contributors to the field of visual perception. The ideas we discussed in
the chapter come from his classic book The Perception of the Visual World,
written in 1950, in which he put forward the idea that animals "sampled"
information from their environment. In the same book he introduced the
notion of “affordance,” which is very important in machine vision (and
also, as it happens, in ergonomic design). To learn more, look for Gib-

son’s book or for books on machine vision. For suggestions for those, see
Chapter 9.
• You can learn about sensor fusion from the work of Prof. Robyn Murphy.
In fact, after reading this book, you should consider reading her Introduc-
tion to AI Robotics, which covers sensor fusion and many other topics at a
more advanced level.
8 Switch on the Light
Simple Sensors
As we learned in Chapter 7, we can consider a sensor simple if it does not

require a great deal of processing to yield useful information to the robot. In
this chapter we will take a closer look at several such simple sensors, includ-
ing switches, light sensors, position sensors, and potentiometers.
But first, let’s consider another way in which we can categorize all sensors,
both simple and complex. Sensors can be divided into two basic categories:
active and passive.
8.1 Passive vs. Active Sensors
PASSIVE S ENSORS Passive sensors measure a physical property from the environment. They con-
D ETECTOR sist of a detector, which perceives (detects) the property to be measured. In
A CTIVE S ENSORS contrast, active sensors provide their own signal/stimulus (and thus typically
require extra energy) and use the interaction of that signal with the environ-
ment as the property they measure. Active sensors consist of an emitter and
E MITTER a detector. The emitter produces (emits) the signal, and the detector perceives
(detects) it.
Passive sensors can be simple or complex. In this chapter we will learn
about some simple passive sensors, including switches and resistive light
sensors, and in Chapter 9 we will learn about cameras, currently the most
complex passive sensors. Analogously, active sensors are not necessarily
complex. In this chapter we will learn about reflectance and break beam
sensors, which are simple and active, and in the next chapter we will learn
about ultrasound (sonar) and laser sensors, both of which are complex and
active.
82 8 Switch on the Light
Figure 8.1 A basic switch, used for interacting with lights and robots, among other
uses.
Remember, whether a sensor is complex is determined by the amount of

processing its data require, while whether a sensors is active is determined by
its mechanism.
Let’s start by taking a look at the simplest passive sensors: switches.
8.2 Switches
Switches, such as the one shown in figure 8.1, are perhaps the simplest sen-
sors of all. They provide useful information at the electronics (circuit) level,
since they are based on the principle of an open vs. a closed circuit. If a
switch is open, no current can flow though the circuit; if it is closed, current
can flow through. By measuring the amount of current flowing through the
circuit, we can tell if the switch is open or closed. So, switches measure the
change in current resulting from a closed circuit, which in turn results from
physical contact of an object with the switch.
Depending on how you wire a switch to a circuit, it can be normally open
or normally closed. Either way, the measurement of current is all you need
in order to use the switch as a sensor. This simple principle is applied in a
8.2 Switches 83
wide variety of ways to create switches, and switches are, in turn, used in a
variety of clever ways for sensing, such as:
C ONTACT S ENSORS • Contact sensors detect when the sensor has contacted another object (e.g.,
they trigger when a robot hits a wall or grabs an object).
L IMIT S ENSORS • Limit sensors detect when a mechanism has moved to the end of its range
(e.g., they trigger when a gripper is wide open).
S HAFT E NCODERS • Shaft encoder sensors detects how many times a motor shaft turns by having
a switch click (open/close) every time the shaft turns.
You use many kinds of switches in your everyday life: light switches, com-
puter mouse buttons, keys on keyboards (computers and electronic pianos),
buttons on the phone, and others.
The simplest yet extremely useful sensor for a robot is a bump switch, also
called a contact switch, which tells the robot when it has bumped into some-
thing. Knowing this, the robot can back away and get around the obstacle,
or stick to it, or keep bumping it, whatever is consistent with its goals. You’ll
discover that even for such a simple sensor as a bump switch, there are many
different ways of implementation.
As we know from numerous examples in biology, building a clever body
structure around a sensor can make that sensor much more sensitive and
accurate. Therefore, switches can be attached to a great variety of places and
components on a robot. For instance, a switch can be attached to a large, rigid
(for example, plastic) surface so that when any part of the surface contacts an
object, the switch closes. This is a good way to find out if any part of a robot’s
front, say, or side, has hit an obstacle. Another clever way to use a switch is
in the form of a whisker, as found on many animals.
Can you think of how to build a (simple) whisker from the principles of
a switch?
Here is one way: Attach a long wire to a switch; whenever the wire is bent
enough, the switch will close and thus indicate contact. But this is not very
sensitive, as the whisker has to bend quite a lot to close the switch. What can
you do to fix this? Well, you can use a rigid material instead of a flexible wire
for the whisker; this will make it more sensitive, and in fact more similar to
the bump sensor we talked about above, for the chassis of the robot’s body.
But the whisker could also break if the robot does not stop. What else can
you do?
A more effective way to build a whisker sensor is as follows: Use a metal

(conductive) wire placed in a metal (conductive) tube. When the whisker
bends, it contacts the tube, and thus closes the circuit. By adjusting the length
and width of the tube, the whisker sensitivity can be tuned for the particular
task and environment of the robot.
The examples above just scratch the surface of the numerous ways to clev-
erly design and place switches to make a robot aware of contact with objects
in its environment.
Touch/contact sensors abound in biology, in much more sophisticated for-
ms than the simple switches we have discussed so far. Whiskers and anten-
nae are the biological inspiration for the sensors we have talked about in this
section. You can consider any part of your skin to be a sensor for contact, as
well as for pressure and heat, and every hair on your body as a whisker. Such
abundance of sensory input is not yet available to machines; it’s no wonder
we are far from having truly “sensitive” robots.
8.3 Light Sensors
Besides being able to detect contact with objects, it is often useful for a robot
to be able to detect areas of darkness and light in its environment. Why?
Because a light can then be used to mark a special area, such as the battery
recharging station or the end of a maze. With a light sensor, a robot can also
find dark places and hide there.
What other uses for light sensors can you think of?
P HOTOCELL Light sensors measure the amount of light impacting a photocell. Photo-
cells, as the name indicates (photo means “light” in Greek), are sensitive to
light, and this sensitivity is reflected in the resistance in the circuit they are
attached to. The resistance of a photocell is low when it is illuminated, sens-
ing a bright light; it is high when it is dark. In that sense, a light sensor is
really a “dark” sensor. This can be made simpler and more intuitive; by sim-
ply inverting the output in the circuit, you can make it so that low means
dark and high means light.
Figure 8.2 shows a photocell; the squiggly line is the photoresitive part
that senses/responds to the light in the environment. Basic household night
lights use the same photocells as some robots, in order to detect and respond
to particular light levels. In night lights, low light causes the bulb to light up,
while in the robot it may result in changing speed or direction, or some other
Figure 8.2 An example of a photocell.
appropriate action. Remember Braitenberg’s vehicles in Chapter 2? They

used this simple principle to produce all sorts of interesting robot behaviors,
such as following, chasing, avoiding, and oscillating, which resulted in com-
plex interpretations by observers, that included repulsion, aggression, fear,
and even love.
Light sensors are simple, but they can detect a wide range of wavelenghts,
much broader than the human eye can see. For example, they can be used
to detect ultraviolet and infra red light, and can be tuned to be sensitive to a
particular wavelength. This is very useful for designing specialized sensors;
you will see how later in this chapter.
Just as we saw with switches, light sensors can be cleverly positioned, ori-
ented, and shielded in order to improve their accuracy and range properties.
They can be used as passive or active sensors in a variety of ways, and they
can measure the following properties:
• Light intensity: how light/dark it is
• Differential intensity: difference between photocells
• Break in continuity: “break beam,” change/drop in intensity.

We will see examples of all of those uses in this chapter. Another property
of light that can be used for sensing is polarization.
8.3.1 Polarized Light

“Normal” light emanating from a light source consists of light waves that
P OLARIZING F ILTER travel in all directions relative to the horizon. But if we put a polarizing filter
in front of the light source, only the light waves with the direction of the filter
will pass through it and travel away. This direction is called the “characteris-
tic plane” of the filter; “characteristic” because it is special for that filter, and
P OLARIZED L IGHT “plane” because it is planar (two-dimensional). Polarized light is light whose
waves travel only in a particular direction, along a particular plane.
Why is this useful? Because just as we can use filters to polarize the light,
we can use filters to detect light with a particular polarization. Therefore we
can design polarized light sensors.
In fact, we don’t have to limit ourselves to just one filter and thus one
characteristic plane of light. We can combine polarizing filters. How does
that work? Consider a light source (such as a lightbulb) covered with a po-
larizing filter. The resulting polarized light is only in the characteristic plane
of the filter. What happens if we put another, identical filter in the path of the
resulting polarized light? All of the light gets through. But what if we use a
filter with a 90-degree angle of polarization instead? Now none of the light
gets through, because none of it is in the characteristic plane of the second
filter.
By playing around with photocells and filters and their arrangement, you
can use polarized light to make specialized sensors that cleverly manipulate
what and how much light is detected. These are active sensors, since they
consist not only of a photocell (for detecting the light level) but also of one
(or more) light source(s) (for emitting the light) and one (or more) filter(s) for
polarizing the light. The general idea is that the filtering happens between
the emitter and the receiver; exactly where – whether closer to the emitter, to
the receiver, or both – depends on the robot, its environment, and its task.
Polarized light sensing exists in nature as well; many insects and birds use
polarized light to navigate more effectively.
8.3.2 Reflective Optosensors

Reflective optosensors, as you can guess by the name, operate on the princi-
ple of reflected light. These are active sensors; they consist of an emitter and
Figure 8.3 The two types of optosensor configuration: reflectance (on the left) and
break beam (on the right).
L IGHT-E MITTING a detector. The emitter is usually made with a light-emitting diode (LED), and
D IODE (LED) the detector is usually a photodiode/phototransistor. You can learn about
the details of what these electronic components consist of as suggested in the
Looking for More section at the end of this chapter.
Reflective optosensors do not use the same technology as the resistive pho-
tocells we learned about in the last section. Resistive photocells are nice and
simple, but their resistive properties make them slow, because change in re-
sistance takes time. Photodiodes and phototransistors, on the other hand,
are much faster, and are therefore preferred for use in robotics.
There are two basic ways in which reflective optosensors can be arranged,
based on the relative positions of the emitter and the detector:
R EFLECTANCE 1. Reflectance sensors: The emitter and the detector are side by side, separated
S ENSORS by a barrier; the presence of an object is detected when the light reflects
from it and back into the detector.
B REAK B EAM S ENSORS 2. Break beam sensors: The emitter and the detector face one another; the pres-
ence of an object is detected if the beam of light between the emitter and
the detector is interrupted or broken (thus the name) .
8.3.3 Reflectance Sensors
What can you do with this simple idea of measuring light re-
flectance/reflectivity?
Here are just a few of the many useful things you can do:
• Detect the presence of an object: Is there a wall in front of the robot?
• Detect the distance to an object: How far is the object in front of the robot?
• Detect some surface feature: Find the line on the floor (or the wall) and
follow it.
• Decode a bar code: Recognize a coded object/room/location/beacon.
• Track the rotations of a wheel: Use shaft encoding; we will learn more
about this later in this chapter.
Although the general idea of using reflected light is simple, the exact prop-
erties of the process are anything but. For example, light reflectivity is af-
fected by the color, texture (smoothness or roughness), and other properties
of the surface it hits. A light-colored surface reflects light better than a dark-
colored one; a matte (non-shiny) black surface reflects little if any light, and
so is invisible to a reflective sensor. Therefore, it may be harder and less
reliable to detect dark objects using reflectance than light ones. In the case
of determining the distance to an object, the same principle makes lighter
objects that are far away seem closer than dark objects that are close.
This gives you part of the idea why sensing in the physical world is chal-
lenging. No sensor is perfect, and all sensors are prone to error and noise
(interference from the environment). Therefore, even though we have useful sen-
sors, we cannot have complete and completely accurate information. These intrin-
sic, unavoidable limitations of sensors are rooted in the physical properties
of the sensor mechanism, and have an impact on the resulting accuracy or
the sensor. Therefore, these fundamental limitations are part of uncertainty in
robotics.
Let’s talk about sensor noise as it relates to reflective light sensors. A light
sensor has to operate in the presence of the light existing in the environment,
A MBIENT L IGHT which is called ambient light. The reflectance sensor must ignore ambient
light in order to be sensitive only to its own emitter’s reflected light. This
is difficult if the wavelength of the ambient light is the same as that of the
emitter. The sensor’s mechanism has to somehow subtract or cancel out the
ambient light from the detector reading, so that it can accurately measure
only the light coming from the emitter, which is what it needs.
How is that done? How does the detector know the amount of ambient
light?
It has to sense it. The ambient light level is measured by taking a sensor
reading with the emitter off. To measure only the emitter’s reflected light, the
detector takes two (or more, for higher accuracy) readings of the sensor level,
one with the emitter on and one with it off. When one is subtracted from the
other (and the signs are adjusted properly so we don’t end up with negative
light), the difference produces the amount of the emitter’s light reflected back
to the sensor. This is an example of sensor calibration.
C ALIBRATION Calibration is the process of adjusting a mechanism so as to maximize its
performance (accuracy, range, etc.). Sensors require calibration, some only
initially, and others continually, in order to operate effectively. Calibration
may be performed by the designer, the user, or the sensor mechanism itself.
Going back to our reflective optosensor, we’ve just seen how to calibrate
it in order to subtract ambient noise. But if we do this calibration only once,
when the sensor is first used, we may find that the sensor becomes very in-
accurate over time. Why is that? Because ambient light levels change during
the course of the day, lights get turned on and off, bright or dark objects may
be present in the area, and so on. Therefore, if the environment can change,
the sensor has to be calibrated repeatedly to stay accurate and useful. There
is no time to sit back and relax when it comes to sensing.
As we mentioned above, ambient light is a problem if it is the same wave-
length as the emitter light, and therefore interferes with it and is hard to
cancel out. The general way to avoid interference is to encode the signal
from the sensor in a way that the detector can easily separate it from ambi-
ent light. One way to do this is through the use of polarization filters, as we
saw earlier. Another way is through adjusting the wavelength of the emitted
light. Here is how.
8.3.4 Infra Red Light

V ISIBLE L IGHT Visible light is light in the frequency band of the electromagnetic spectrum
that human eyes can perceive.1 Infra red (IR) light has a wavelength different
1. If you remember the electromagnetic spectrum from physics, that helps in this chapter, but is
not necessary.
from visible light and is not in the visible spectrum. IR sensors are a type of
light sensors which function in the infra red part of the frequency spectrum.
They are used in the same ways that visible light sensors are: as reflectance
sensors or as break beams.
Both of those uses are forms of active sensors. IRs are not usually used
as passive sensors in robotics. Why is that? Because robots do not usually
need to detect ambient IR. But other types of sensors, such as IR goggles,
better known as “night vision” glasses, do detect ambient IR passively. They
collect and enhance the light in the IR spectrum and transform it into the
visible spectrum. When tuned to the frequency of human body heat, they
can be (and are) used to detect human movement in the dark.
IR is very useful in robotics because it can be easily modulated, and in that
way made less prone to interference. Such modulated IR can also be used for
communication (i.e., for transmitting messages), which is how IR modems
work. So let’s learn about modulation.
8.3.5 Modulation and Demodulation of Light

M ODULATED L IGHT Light is modulated by rapidly turning the emitter on and off, pulsing it. The
D EMODULATOR resulting pulsed signal is then detected by a demodulator, a mechanism that
is tuned to the particular frequency of the modulation, so it can can be de-
coded. The detector has to sense several “on” flashes in a row in order for
the demodulator to determine its frequency and decode it.
Strobe light is a type of modulated visible light, and given how hard it is to
look at a strobe light, you can see (so to speaks) why visible light is not usu-
ally used in modulated form: it’s too hard on the eyes. However, modulated
IR is commonly used, since IR is not in the visible spectrum. Most house-
hold remote controls are based on modulated IR, including the fought-over
TV channel changer.
8.3.6 Break Beam Sensors

You probably have a good intuitive idea of how a break beam sensor works.
In general, any pair of compatible emitter-detector devices can be used to
produce break beam sensors, including:
• An incandescent flashlight bulb and a photocell
• Red LEDs and visible-light-sensitive phototransistors
• Infra red IR emitters and detectors.
Figure 8.4 Break beam shaft encoder mechanism.
Where have you seen break beam sensors? Images from movies might
come to mind, with crisscrossed laser beams and clever burglars making
their way through them. More realistically, one of the most common uses
of break beam sensing is not in plain sight, because it is packed inside motor
mechanisms and used to keep track of shaft rotation. Here is how it works.
8.3.7 Shaft Encoders

S HAFT E NCODER Shaft encoders measure the angular rotation of a shaft or an axle. They provide
position and/or velocity information about the shaft they are attached to. For
example, the speedometer measures how fast the wheels of the car are turn-
ing, and the odometer measures the number of rotations of the wheels. Both
speedometers and odometers use shaft encoding as the underlying sensing
mechanism.
In order to detect a turn, or a part of a turn, we have to somehow mark the
thing that is turning. This is usually done by attaching a round notched disk
to the shaft. If the shaft encoder uses a switch, the switch clicks every time
the shaft completes a full rotation. More commonly, a light sensor is used:
a light emitter is placed on one side of the disk, and a detector on the other,
in a break beam configuration. As the disk spins, the light from the emitter
Figure 8.5 Reflectance-based shaft encoder mechanism.
reaches the detector only when the notched part of the disk passes in front of
it.
If there is only one notch in the disk, then every time the notch passes
between the emitter and the detector, this means the disk has completed one
full rotation. This is useful, but it allows for measuring with only a low level
of precision. If any noise or error is present, one or more turns might be
missed, and the encoder will thus be quite inaccurate.
To make the encoder more accurate as well as more precise, many notches
are cut into the disk. The break beam principle is still the same: whenever
the light gets through, it is sensed by the detector and counted. Figure 8.4
shows what the mechanism looks like. You can see that it is important to
have a fast sensor if the shaft turns very quickly. That is why a resistive sen-
sor would not be appropriate; it is comparatively slow, while an optosensor
works well for this propose, as we discussed earlier in this chapter. (If you
forgot, just remember that optosensors use light, which travels faster than
anything else.)
An alternative to cutting notches in the disk is to paint the disk with wedges
of alternating, contrasting colors. The best color choices are black (absorbing,
nonreflecting) and white (highly reflecting), as they provide the highest con-
trast and the best reflective properties. But in this case, since there are no
notches in the disk, how does the break beam sensor work? It doesn’t. This
is not a break beam sensor, but a reflectance sensor. Instead of putting the
sensor in the break beam configuration, in this case the emitter and the de-
tector are placed on the same side of the disk, side by side, in a reflectance
configuration. Figure 8.5 shows what that looks like.
Regardless of whether the shaft encoding sensor is a break beam or a re-
flectance sensor, the detector will output a wave function (a sequence of the
signal being on and off, up and down) of the sensed light intensity of the
emitter. This output is then processed, using signal processing, with hard-
ware or a simple processor, to calculate the position and the speed by count-
ing the peaks of the waves.
We can use encoders in at least two ways to measure the speed of a robot:
• Encode and measure the speed of a driven wheel
• Encode and measure the speed of a passive wheel (caster) that is dragged
by the robot.
Why might we want to use the second of the above? One use is for robots
that have legs instead of wheels. For example, the designers of Genghis, the
six-legged robot mentioned in Chapter 5, had it drag a little wheel in some
of the experiments, to measure the distance it had traveled, especially when
it was learning how to walk (see Chapter 21 to learn about that).
We can combine the position and velocity information the encoder pro-
vides to have the robot do more sophisticated things, such as move in a
straight line or turn at by exact angle. However, doing such precise move-
ments is quite difficult, because wheels tend to slip and slide, and there is
usually some backlash in the gearing mechanism (recall Chapter 4). Shaft en-
coders can provide feedback to correct some of the errors, but having some
error remain is unavoidable. There is no perfect sensor, since uncertainty is
a fact of life.
So far, we’ve talked about detecting position and velocity, but did not talk
about direction of rotation. Suppose the robot’s wheel suddenly changes the
direction of rotation; it would be useful for the robot to be aware of it. If the
change is intentional, the encoder can tell the robot how accurate the turn
was; and if the change is unintentional, the encoder may be the first or only
way the robot will know it has turned.
The mechanism for detecting and measuring direction of rotation is called
Q UADRATURE S HAFT quadrature shaft encoding. One place where you might have used it is inside an
E NCODING old-fashioned computer mouse, the kind that has a ball inside (not the more
recent, optical type). In such a mouse, the direction of rotation of the ball,
and thus the direction of the movement of the mouse itself, is determined
through quadrature shaft encoding.
Quadrature shaft encoding is an elaboration of the basic break beam idea:
instead of using only one sensor, use two. The two encoders are aligned so
that their two inputs coming from the detectors are 90 degrees (one quarter
of a full circle, thus the name quadrature) out of phase. By comparing the
outputs of the two encoders at each time step with the output of the pre-
vious time step, we can tell if there is a direction change. Since they are
out of phase, only one of them can change its state (i.e., go from on to off
or vice versa) at a time. Which one does it determines in which direction
the shaft is rotating. Whenever a shaft is moving in one direction, a counter
is incremented in that encoder, and when it turns in the opposite direction,
the counter is decremented, thus keeping track of the overall position of the
mechanism.
In robotics, quadrature shaft encoding is used in robot arms with complex
joints, such as the ball-and-socket joints we discussed in Chapter 4. It is also
C ARTESIAN R OBOTS used in Cartesian robots, which are similar in principle to Cartesian plotter
printers, and are usually employed for high-precision assembly tasks. In
those, an arm moves back and forth along an axis or gear.
We have seen that switches and light sensors can be used in a variety of
different ways, and in some cases in the same ways (as in shaft encoding).
Let’s talk about one more type of simple sensor.
8.4 Resistive Position Sensors
As we just learned, photocells are resistive devices that sense resistance in

response to the light. Resistance of a material, it turns out, can change in
response to other physical properties besides light. One such property is
tension: the resistance of some devices increases as they are bent. These
passive “bend sensors” were originally developed for video game controls,
but have since been adopted for other uses as well.
As you might expect, repeated bending fatigues and eventually wears out
the sensor. Not surprisingly, bend sensors are much less robust than light
sensors, although the two use the same underlying principle of responding
to resistance.
8.4 Resistive Position Sensors 95
By the way, your muscles are full of biological bend sensors. These are
proprioceptive sensors that help the body be aware of its position and the
work it is doing.
8.4.1 Potentiometers
Potentiometers, popularly known as “pots,” are commonly used for manual
tuning of analog devices: they are behind every knob or slider you use on
a stereo system, volume control, or light dimmer. These days, it’s getting
D IGITAL harder to find such knobs, since most devices are digital, meaning using dis-
crete (from the Latin discretus meaning “separate”) values. The word “digi-
tal” comes from the Latin digitus meaning finger, and most electronic devices
today are tuned digitally, by pushing buttons (with fingers), and thus they
use switches rather than pots. But not so long ago, tuning to radio stations
and adjusting volume, among other things, were done with pots.
Potentiometers are resistive sensors; turning the knob or pushing a slider
effectively alters the resistance of the sensor. The basic design of potentiome-
ters involves a tab that slides along a slot with fixed ends. As the tab is
moved, the resistance between it and each of the ends of the slot is altered,
but the resistance between the two ends remains fixed.
In robotics, potentiometers are used to tune the sensitivity of sliding and
rotating mechanisms, as well as to adjust the properties of other sensors. For
example, a distance sensor on a robot may have a potentiometer attached
to it which allows you to tune the distance and/or sensitivity of that sensor
manually.
You might be thinking that these simple sensors are not so very simple af-
ter all. And you are right; sensor mechanisms can get pretty complex pretty
fast, but they are not much compared with biological sensors. Still, keep in
mind that while these are nontrivial physical mechanisms, the resulting sen-
sor data are quite simple and require little processing. In the next chapter
we’ll move on to some complex sensors that spew out much more data and
give us more processing work to do.
To Summarize
• Sensors can be classified into active and passive, simple and complex.
• Switches may be the simplest sensors, but they provide plenty of vari-
ety and have a plethora of uses, including detecting contact, limits, and
turning of a shaft.
• Light sensors come in a variety of forms, frequencies, and uses, includ-

ing simple photocells, reflective sensors, polarized light and infra red (IR)
sensors.
• Modulation of light makes it easier to deal with ambient light and to de-
sign special-purpose sensors.
• There are various ways to set up a break beam sensor, but they are most
commonly used inside motor shaft encoders.
• Resistive position sensors can detect bending and are used in a variety of
analog tuning devices.
Food for Thought
• Why might you prefer a passive to an active sensor?
• Are potentiometers active or passive sensors?
• Our stomach muscles have stretch receptors, which let our brains know
how stretched our stomach are, and keep us from eating endlessly. What
robot sensors would you say are most similar to such stretch receptors?
Are they similar in form (mechanism of how they detect) or function
(what they detect)? Why might stretch receptors be useful to robots, even
without stomachs and eating?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Sensors
• The all-around best text for learning about electronics as well as debug-
ging hardware problems is The Art of Electronics by Paul Horowitz and
Winfield Hill (Cambridge University Press). Every robot lab worth its
beans has at least one well-worn copy.
• Sensors for Mobile Robots: Theory and Applications by H. R. (Bart) Everett is

a comprehensive and reader-friendly textbook covering all of the sensors
we have overviewed in this chapter, and quite a few more.
9 Sonars, Lasers, and Cameras
Complex Sensors
Congratulations, you’ve graduated from simple to complex sensors, and

now you are in for it! The sensors we have seen so far, passive or active,
do not require a great deal of processing or computation in order to provide
information directly useful to a robot. However, the information they pro-
vide is itself simple and limited: light levels, presence or absence of objects,
distance to objects, and so on. Not only are complex processing and com-
putation not necessary, but they would be of little use. Not so with complex
sensors. In contrast to simple ones, complex sensors provide much (much,
much) more information, that can yield quite a bit more useful fodder for the
robot, but also requires sophisticated processing.
In this chapter, we will learn about ultrasound, laser, and vision sensors,
some of the most commonly used complex sensors in robotics. But don’t as-
sume that they are the only complex sensors available; there are others (such
as radar, laser radar, GPS, etc.), and new ones are always being developed.
9.1 Ultrasonic or Sonar Sensing
U LTRASOUND Ultrasound literally means “beyond sound,” from the Latin ultra for “be-
yond” (used here in the same way as in “ultraviolet” and “ultraconserva-
tive”) and refers to a range of frequencies of sound that are beyond human
S ONAR hearing. It is also called sonar, from so(und) na(vigation and) r(anging). Fig-
ure 9.1 shows a mobile robot equipped with sonar sensors.
The process of finding your (or a robot’s) location based on sonar is called
E CHOLOCATION echolocation. Echolocation works just the way the name sounds (no pun in-
tended): sound bounces off objects and forms echoes that are used to find
one’s place in the environment. That’s the basic principle. But before we get
into the details, let’s first consider some examples.
98 9 Sonars, Lasers, and Cameras
Figure 9.1 A mobile robot with sonar sensors.
The principle of echolocation comes from nature, where it is used by sev-

eral species of animals. Bats are the most famous for using sophisticated
echolocation; cave bats that dwell in nearly total darkness do not use vision
(it would not do them much good), but rely entirely on ultrasound. They
emit and detect different frequencies of ultrasound, which allows them to fly
effectively in very crowded caves that have complicated structures and are
packed with hundreds of other flying and hanging bats. They do all this very
quickly and without any collisions. Besides flying around in the dark, bats
also use echolocation to catch tiny insects and find mates. Dolphins are an-
other species known for sophisticated echolocation. What used to be secret
research is now standard in aquarium shows: blindfolded dolphins can find
small fish and swim through hoops and mazes by using echolocation.
As usual, biological sensors are vastly more sophisticated than current ar-
tificial ones (also called “synthetic,” because they have been “synthesized,”
not because they are made out of synthetics); bat and dolphin sonars are
much more complex than artificial/synthetic sonars used in robotics and
other applications. Still, synthetic sonars are quite useful, as you will see.
So how do they work?
T IME - OF -F LIGHT Artificial ultrasound sensors, or sonars, are based on the time-of-flight prin-
ciple, meaning they measure the time it takes something (in this case sound)
to travel (“fly”). Sonars are active sensors consisting of an emitter and a de-
tector. The emitter produces a chirp or ping of ultrasound frequency. The
sound travels away from the source and, if it encounters a barrier, bounces
off it (i.e., reflects from it), and perhaps returns to the receiver (microphone).
If there is no barrier, the sound does not return; the sound wave weakens
(attenuates) with distance and eventually breaks down.
If the sound does come back, the amount of time it takes for it to return can
be used to calculate the distance between the emitter and the barrier that the
sound encountered. Here is how it works: a timer is started when the chirp
is emitted, and is stopped when the reflected sound returns. The resulting
time is then multiplied by the speed of sound and divided by two. Why?
Because the sound traveled to the barrier and back, and we are only trying
to determine how far away the barrier is, its one-way distance.
S PEED OF S OUND This computation is very simple, and relies only on knowing the speed of
sound, which is a constant that varies only slightly due to ambient tempera-
ture. At room temperature, sound travels 1.12 feet per millisecond. Another
way to put it is that sound takes 0.89 milliseconds to travel the distance of 1
foot. This is a useful constant to remember.
The hardware for sonar sensing most commonly used in robotics is the
Polaroid Ultrasound Sensor, initially designed for instant cameras. (Instant
cameras were popular before digital cameras were invented, since they pro-
vided instant photos; otherwise people had to wait for film to be developed,
which took at least a day, unless they a personal film development lab.) The
physical sensor is a round transducer, approximately 1 inch in diameter, that
T RANSDUCER emits the chirp/ping and receives the sound (echo) that comes back. A trans-
ducer is a device that transforms one form of energy into another. In the case
of the Polaroid (or other ultrasound) transducers, mechanical energy is con-
verted into sound as the membrane of the transducer flexes to produce a ping
that sends out a sound wave that is inaudible to humans. You can actually
hear most robot sonars clicking but what you hear is the movement of the
emitter (the membrane), not the sound being sent out.
The hardware (electronics) of ultrasound sensors involves relatively high
power, because significant current is needed for emitting each ping. Impor-
tantly, the amount of current required is much larger than what computer
processors use. This is just one of many practical examples showing why
it is a good idea to separate the power electronics of a robot’s sensing and
actuation mechanisms from those of its controller processor. Otherwise, the

robot’s brain might have to literally slow down in order for the body to sense
or move.
The Polaroid ultrasound sensor emits sound that spreads from a 30-degree
sound cone in all directions, and at about 32 feet attenuate to a point that they
do not return to the receiver, giving the sensor a 32-foot range. The range
of an ultrasound sensor is determined by the signal strength of the emitter,
which is designed based on the intended use of the sensor. For robots (and
for instant cameras, as it happens), the range of 32 feet is typically sufficient,
especially for indoor environments. Some other uses of sonar require quite a
bit less or more range, as you will see in the next section.
9.1.1 Sonar Before and Beyond Robotics

Ultrasound is used in a variety of applications besides (and before) robotics,
from checking on babies in utero (inside the mother’s womb) to detecting ob-
jects and attackers in submarines. When sonar is used to look into people’s
bodies, the result is called a sonogram, echogram, or ultrasonogram, coming
from the Greek gram meaning “letter” and referring to writing and drawing.
Sound travels well through air and water, and since the human body consists
largely of water (over 90 percent by weight), ultrasound is a good technol-
ogy for seeing what’s going on inside. One of the most common uses of
sonograms is for range (distance) measurements, just like its use in robotics.
However, the Polaroid and other sonars used in robotics operate at about 50
KHz, while the medical ones operate in the higher frequency range, around
H ERTZ (H Z ) 3.5 to 7 MHz. Just in case you did not know or you forgot, a Hertz (Hz) is a
unit of frequency. One Hertz means once per second, a Kilohertz (KHz) is
1000 Hz and a Megahertz (MHz) is 1,000,000 Hz.
Simply sensing the distances between the emitter and the environment
may be sufficient for many robotics applications, but in medical sonar imag-
ing, much more complex postprocessing is involved in order to create a com-
posite image of the body part. This image is not static, but is updated in real
time, allowing for what looks like real-time video of, for example, the beating
heart of a baby in utero.
Since sound travels well through water, while vision is almost useless in
the underwater environment, it is no surprise that sonar is the favored sen-
sor for underwater navigation, specifically for helping submarines detect
and avoid any unexpected obstacles, such as other submarines. You have
no doubt seen images of large, looming, menacing, and constantly ping-
ing submarines in the movies. Sonars used in submarines have long ranges,
through the use of stronger signal intensity and narrower cones. As in depth
sounders, these sonars send an intense beam of sound into the ocean (or
any body of water), and wait for it to return, in order to see how deep the
water is or, in other words, how far the nearest object/surface is. As you
can imagine, these sonars need to reach much farther than 32 feet; thus their
use of a narrower cone and stronger signal intensity. While such uses of ul-
trasound are inaudible and very useful to people, they are audible and, it
turns out, dangerous to marine animals, such as whales and dolphins. High-
strength and long-range ultrasound emissions have been shown to confuse
whales and cause them to beach and die. The exact causes of this behavior
are not yet understood, but the power of ultrasound should not be underes-
timated. When properly controlled and directed, it can be used to break up
objects, such as kidney stones. Between the body and the ocean, there are
other, more mundane uses of sonar. They include automated tape measures,
height measures, and burglar alarms.
The principle of time-of-flight underlies all uses of sonar as a ranging and
imaging device. In almost all applications, multiple sensor units are em-
ployed for increased coverage and accuracy. Most robots using sonars are
equipped with several, usually a full ring, covering a cross-section of the
robot’s body.
Here is an easy question: what is the smallest number of standard Po-
laroid sonar sensors needed to cover the cross-section of a robot?
As we will see later, a dozen sonars cover the circle around a robot (unless
the robot has a very wide berth, in which case more are used). Whether a
dozen or more are used, they cannot all be pinged/emitted at the same time.
Can you guess why?
Sonar sensors, Polaroid or others, are inexpensive and easy to incorpo-
rate into robot hardware. If only sonars always returned accurate distance
readings! But they do not, since things are never that simple in the physical
world. Sonar data can be tricky, and here is the reason why.
9.1.2 Specular Reflection

As we saw, sonar sensing is based on the emitted sound wave reflecting from
surfaces and returning to the receiver. But the sound wave does not necessar-
ily bounce off the nearest surface and come right back, as we might hope it
would. Instead, the direction of reflection depends on several factors, includ-
Figure 9.2 Specular reflection of an ultrasound signal.
ing the properties of the surface (how smooth it is) and the incident angle of
the sound beam and the surface (how sharp it is).
A major disadvantage of ultrasound sensing is its susceptibility to specular
S PECULAR R EFLECTION reflection. Specular reflection is the reflection from the outer surface of the
object; this means the sound wave that travels from the emitter bounces off
multiple surfaces in the environment before returning to the detector. This is
likely if the surface it encounters is smooth and if the angle between the beam
and the surface is small. The smaller the angle, the higher the probability that
the sound will merely graze the surface and bounce off, thus not returning
to the emitter but moving on to other surfaces (and potentially even more
grazing bounces) before returning to the detector, if it returns at all. This
bouncing sound generates a false faraway reading, one that is much longer
than the straight-line distance between the robot (its sonar sensor) and the
surface. The smoother the surface is, the more likely the sound is to bounce
off. In contrast, rough surfaces produce more irregular reflections, which
are more likely to return to the emitter. Think of it this way: as the sound
hits a rough surface, it scatters, bouncing back at various angles relative to
the various facets and grooves and features on the surface. At least some of
the reflections are likely to go back to the emitter, and thus provide a rather
accurate distance measure. In contrast, as the sound hits a uniformly smooth
surface (a specular one), it may graze or bounce off it uniformly in a direction

away from the detector. In figure 9.2 you can see an illustration of specular
reflection. Specularity is a property of light as well as of sound, which adds to
the challenges of machine vision; we’ll worry about that in the next section.
In the worst case of specular reflection, the sound bounces around the en-
vironment and does not return to the detector, thus fooling the sensor into
detecting no object/barrier at all, or one that is far away instead of nearby.
This could happen in a room full of mirrors, as at a carnival, or full of glass
cases, as at a museum. A crude but effective way to combat specular re-
flection is to alter the environment in which the sonar-based robot has to
navigate by making the surfaces less reflective. How can we do that? Well,
we can rub all smooth surfaces with sandpaper, or we can use rough wall-
paper, or we can put little wood slats in the walls, basically anything to
introduce features on the surfaces. Fortunately, since sonar beams are rel-
atively focused, especially at short ranges, only the surfaces that the cone of
the sonars is most likely to sense need be altered, not entire walls. For exam-
ple, in research labs, roboticists have lined experimental areas with bands of
corrugated cardboard, because its ridges have much better sonar reflectance
properties than the smooth walls. In general, altering the environment to suit
the robot is not a great idea and it is often not even possible (such as under
the ocean or in space).
How else can we get around the specular reflection problem?
One solution is to use phased arrays of sensors to gain more accuracy. The
basic idea is to use multiple sensors covering the same physical area, but
activated out of phase. This is exactly what is used in automatic tape measure
devices: these contraptions, when pointed at an object, give the distance to
that object, by using multiple carefully arranged and timed sonar sensors.
You can think of this as a hardware-level solution to the problem.
Can you think of some software/processing-level computational solu-
tions?
Remember that the problem is that long sonar readings can be very in-
accurate, as they may result from false reflected readings from nearby ob-
jects rather than accurate faraway readings. We can use this fact to embed
some intelligence into the robot, in order to make it accept short readings
but do more processing on the long ones. One idea is to keep a history of
past readings, and see if they get longer or shorter over time in a reasonable,
continuous way. If they do, the robot can trust them, and if not, the robot
assumes they are due to specular effects. This approach is effective in some
environments but not in all, and is especially challenging in unknown struc-
D ISCONTINUITY tures. After all, the environment may have discontinuities, sudden and large
changes in its features. Those cannot be anticipated by the robot, so it is left
having to trust its sensor readings, however unpredicatable they may be.
Besides postprocessing, the robot can also use action-oriented perception,
which we discussed in Chapter 7. Whenever it receives an unexpected long
sonar reading that seems strange/unlikely, it can turn and/or move so as to
change the angle between its sonar sensor and the environment, and then
take another reading, or many more readings, to maximize accuracy. This is
a good example of a general principle:
Using action to improve sensory information is a powerful method of
dealing with uncertainty in robotics.
Ultrasound sensors have been successfully used for very sophisticated
robotics applications, including mapping complex outdoor terrain and in-
door structures. Sonars remain a very popular, affordable ranging sensor
choice in mobile robotics.
9.2 Laser Sensing
Sonars would be great sensors if they were not so susceptible to specular

reflection. Fortunately, there is a sensor that largely avoids the problem, but
at some cost and trade-offs: the laser.
L ASER Lasers emit highly amplified and coherent radiation at one or more fre-
quencies. The radiation may be in the visible spectrum or not, depending on
the application. For example, when laser sensors are used as burglar detec-
tors, they are typically not visible. In the movies, we are often shown visible
laser grids, but in reality making them visible makes the burglar’s job easier
By the way, what principle are such sensors based on? They are break beam
sensors: when a burglar (or a roaming cat, say) breaks the laser beam, the
alarm goes off. When laser sensors are used for range measurements, they
also are typically not visible, since it is usually not desirable to have a visible
and distracting beam of light scanning the environment as the robot moves
around.
Laser range sensors can be used through the time-of-flight principle, just
like sonars. You can immediately guess that they are much faster, since the
9.2 Laser Sensing 105
Figure 9.3 A mobile robot sporting a laser sensor. The laser is the thing that resem-
bles a coffee maker and is labeled with the name of the manufacturer, SICK.
speed of light is quite a bit greater than the speed of sound. This actually
causes a bit of a problem when lasers are used for measuring short distances:
the light travels so fast that it comes back more quickly than it can be mea-
sured. The time intervals for short distances are on the order of nanoseconds
and can’t be measured with currently available electronics. As an alternative,
phase-shift measurements, rather than time-of-flight, are used to compute
the distance.
Robots that use lasers for indoor navigation and mapping, such as the one
shown in figure 9.3, typically operate in relatively short-range environments
by laser standards. Therefore, lasers used on mobile robots use phase-shift
rather than time-of-flight. In both cases the processing is typically performed
within the laser sensor itself, which is equipped with electronics, and so the
sensor package produces clean range measurements for the robot to use.
Lasers are different from sonars in many other ways stemming from the
differences in the physical properties of sound and light. Lasers involve
higher-power electronics, which means they are larger and more expensive.
They are also much (much, much) more accurate. For example, a popular
laser sensor, the SICK LMS200 scanning laser rangefinder, has a range of 8m
and a 180-degree field of view. The range and bearing of objects can be de-
termined to within 5mm and 0.5 degrees. The laser can also be used in a
long-range mode (up to 80m), which results in a reduction in accuracy of
only about 10cm.
Another key distinction is that the emitted laser light is projected in a beam
rather than a cone; the spot is small, about 3mm in diameter. Because lasers
use light, they can take many more measurements than sonar can, thereby
R ESOLUTION providing data with a higher resolution. Resolution refers to the process of
separating or breaking something into its constituent parts. When something
has high resolution, it means it has many parts. The more parts there are to
the whole, the more information there is. That’s why “high res” is a good
thing.
So what is the resolution of the SICK sensor? Well, it makes 361 readings
over a 180-degree arc at a rate of 10Hz. The true rate is much higher (again,
because of the speed of light and the speed of today’s electronics), but the
resulting rate is imposed by the serial link for getting the data out. Ironically,
the serial link is the bottleneck, but it’s good enough, since a real robot does
not need laser data at any rate higher than what this sensor provides via the
port; it could not physically react any faster anyway.
Laser sensors sound like a dream come true, don’t they? They have high
resolution and high accuracy, and do not suffer nearly as much from specular
effects. Of course they are not totally immune to specularities, since they are
light-based and light is a wave that reflects, but specularities are not much
of a problem due to the resolution of the sensor, especially compared with
ultrasound sensing. So what is the downside?
Well, first of all, they are large, about the size of an electric coffee maker
(and sort of look like one, too); it takes some space to house all those high-
power electronics. Next, and most impressively, laser sensors are very ex-
pensive. A single SICK laser costs two orders of magnitude more than a
Polaroid sonar sensor. Fortunately, price is bound to come down over time,
and packaging is bound to become smaller. However, the high resolution
of the laser range sensor has its downside. While the narrow beam is ideal
for detecting the distance of a particular point, to cover an area, the laser
needs to sweep or scan. The planar (2D) SICK laser mentioned above scans
horizontally over a 180-degree range, providing a highly accurate slice of
distance readings. However, if the robot needs to know about distances out-
side of the plane, more lasers and/or more readings are needed. This is quite
doable, since lasers can scan quickly, but it does take additional sensing and
processing time. 3D laser scanners also exist, but are even larger and more
expensive. They are ideal, however, for accurately mapping out distances to
objects, and thus the space around the sensor (or robot using the sensor).
In mobile robotics, simple sensors are usually most popular, and lasers,
even planar ones, are not sufficiently affordable or portable for some appli-
cations. Can you guess which ones? For example, any that include small
robots, or robots interacting with children or people who might look in the
direction of the laser, which is too high-power to be considered completely
safe.
Let’s go back to the laser grid for detecting intruders and burglars in movies.
Projecting a visible grid of laser light on the environment is part of another
approach to sensing. The distortions of the grid represent the shapes of ob-
jects in the environment. But to detect the pattern and its distortions, we need
another sensor: a camera. That brings us to the most complex and versatile
sensory modality of all: vision.
9.3 Visual Sensing
Seeing requires a visual sensor, something akin to biological eyes. Cameras

are the closest thing to natural eyes that we have available among synthetic
sensors. Needless to say, any/all biological eyes are more complex than any
cameras we have today. But to be fair, seeing is not done by eyes only, but
largely by the brain. Eyes provide the information about the incoming light
patterns, and the brain processes that information in complex ways to an-
swer questions such as “Where are my car keys?” and make decisions such
as “Stop right now; there they are, between the sofa cushions!”
Although cameras and computers are a far cry from, and much less com-
plex than, biological eyes and brains, you will soon see that the information
provided by cameras is not simple in the least, and that the associated pro-
cessing is some of the most complex in robotics. The research field that deals
M ACHINE V ISION with vision in machines, including robots, is called, appropriately, machine
vision. Robots have particular perceptual needs related to their tasks and
environments, and so some parts of machine vision research are relevant
and useful for robotics, while others have proven not to be, even if they are
very useful for other applications. Therefore, machine vision and robotics
are two separate fields of research with some overlap of interests, problems,
and uses.
Traditionally, machine vision has concerned itself with answering the ques-
tions “What is that?”, “Who is that?” and “Where is that?” To answer these,
Figure 9.4 Components of biological and synthetic vision systems.
the approach has been to reconstruct what the world was like when the cam-
era took its picture, in order to understand the picture, so that the question
can be answered. We already talked about vision as reconstruction in Chap-
ter 7.
After decades of research in machine vision, we know that visual recon-
struction is an extremely difficult problem. Fortunately for robotics, it is not
the problem that robots really need to solve. Instead, robots are typically
concerned with acting so as to achieve their goals or, put more simply, with
doing the right thing. Instead of answering the machine vision questions
above, they need to answer action-related questions of the type: “Should I
keep going or turn or stop?”, “Should I grab or let go?”, “Where do I turn?”,
and so on. It is usually not necessary to reconstruct the world in order to
answer those questions. To begin to understand how robots use vision, let’s
first see (so to speak) how cameras, the basic visual sensors, do their work.
9.3.1 Cameras
B IOMIMETIC Cameras are biomimetic meaning they imitate biology, in that they work some-
what the way eyes do. But as usual, synthetic and natural vision sensors are
quite different. Figure 9.4 shows a simple comparison, marking the key com-
ponents of the vision system.
Here are those components. Light, scattered from objects in the environ-
S CENE ment (which are collectively called the scene), goes through an opening (the
I RIS iris, which is in the simplest case a pinhole, but usually is a lens) and hits the
I MAGE P LANE image plane. The image plane corresponds to the retina of the biological eye,
P HOTOSENSITIVE which is attached to numerous light-sensitive (photosensitive) elements called
E ARLY V ISION rods and cones. These in turn are attached to nerves that perform early vision,
the first stages of visual image processing, and then pass information on to
H IGH -L EVEL V ISION other parts of the brain to perform high-level vision processing, everything
else that is done with the visual input. As we have mentioned before, a very
large portion of the human (and other animal) brain activity is dedicated to
visual processing, so this is a highly complex endeavor. Instead of rods and
cones, film cameras use silver halides on photographic film, and digital cam-
eras use silicon circuits in charge-coupled devices (CCD). In all cases, some
information about the incoming light (e.g., its intensity, color) is detected by
the photosensitive elements on the image plane.
In machine vision, the computer needs to make sense out of the informa-
tion on the image plane. If the camera is very simple, and uses a tiny pinhole,
then some computation is required to determine the projection of the objects
from the environment onto the image plane (note that they will be inverted).
L ENS If a lens is involved (as in vertebrate eyes and real cameras), then more light
can get in, but at the price of being focused; only objects a particular range of
F OCUS distances from the lens will be in focus. This range of distances is called the
D EPTH OF F IELD camera’s depth of field.
P IXEL The image plane is usually subdivided into equal parts called pixels, typ-
ically arranged in a rectangular grid. As we saw in Chapter 7, in a typical
camera there are 512 × 512 pixels on the image plane. For comparison, there
are 120 × 106 rods and 6 × 106 cones in the human eye.
The projection of the scene on the image plane is called, not surprisingly,
I MAGE the image. The brightness of each pixel in the image is proportional to the
amount of light that was reflected into the camera by the part of the object or
S URFACE PATCH surface that projects to that pixel, called the surface patch.
Since you already know that reflectance properties vary, you can guess
that the particular reflectance properties of the surface patch, along with the
number and positions of light sources in the environment, and the amount
of light reflected from other objects in the scene onto the surface patch all
have a strong impact on what the pixel brightness value ends up being. All
those influences that affect the brightness of the patch can be lumped into
two kinds of reflections: specular (off the surface, as we saw before) and
D IFFUSE R EFLECTION diffuse. Diffuse reflection consists of the light that penetrates into the object, is
absorbed, and then comes back out. To correctly model light reflection and
reconstruct the scene, all these properties are necessary. No wonder visual
reconstruction is hard to do. It’s a good thing that robots usually don’t need
to do it.
We need to step back for just a second here and remember that the cam-
era, just like the human eye, observes the world continually. This means
T IME S ERIES it captures a video, a series of images over time. Processing any time series
information over time is pretty complicated. In the case of machine vision,
F RAME each individual snapshot in time is called a frame, and getting frames out of
a time series is not simple. In fact, it involves specialized hardware, called
F RAME G RABBER a frame grabber, a device that captures a single frame from a camera’s analog
D IGITAL I MAGE video signal and stores it as a digital image. Now we are ready to proceed
I MAGE P ROCESSING with the next step of visual processing, called image processing.
9.3.2 Edge Detection

E DGE D ETECTION The typical first step (early vision) in image processing is to perform edge
detection, to find all the edges in the image.
How do we recognize edges? What are edges, really?
E DGE In machine vision, an edge is defined as a curve in the image plane across
which there is a significant change in the brightness. More intuitively, finding
edges is about finding sharp changes in pixel brightness. Finding changes
mathematically is done by taking derivatives. (Calculus has its uses.) A
simple approach to finding edges, then, is to differentiate the image and look
for areas where the magnitude of the derivative is large, which indicates that
the difference in the local brightness values is also large, likely due to an
edge. This does find edges, but it also finds all sorts of other things that
produce large changes, such as shadows and noise. Since it is impossible
to distinguish “real” edges from those resulting from shadows simply by
looking at pixel brightness/intensity in the image, some other method has to
be used to do better.
What about noise – how do we deal with it?
Unlike shadows, noise produces sudden and spurious intensity changes

that do not have any meaningful structure. This is actually a good thing,
Figure 9.5 A camera image that has been processed through edge detection and
then segmentation.
since noise appears as peaks in the intensities, and those peaks can be taken
away by a process called “smoothing.”
How is smoothing done automatically?
S MOOTHING Again, math comes to the rescue. To perform smoothing, we apply a math-
ematical procedure called convolution, which finds and eliminates the iso-
C ONVOLUTION lated peaks. Convolution applies a filter to the image; this is called convolving
the image. This type of mathematical filter is really the same, in principle,
as a physical filter, in that the idea is to filter out the unwanted things (in
this case the spurious peaks that come from visual noise) and let through
the good stuff (in this case the real edges). The process of finding real edges
involves convolving the image with many filters with different orientations.
Think back to the previous chapter, when we talked about polarized light fil-
ters. Those were physical filters, and here we are talking about mathematical
filters, but both have the same function: to separate out a particular part of
the signal. In the case of the polarized filters, we were looking for a particu-
lar frequency of light; in the case of edge detection filters, we are looking for
intensities with particular orientations.
Edge detection used to be a very popular difficult problem in machine vi-

sion, and many algorithms were written for it, tested, and published. Even-
tually, researchers developed the best algorithms possible for the task, so
now edge detection is not considered an interesting research problem, but it
is still a very real practical problem in machine vision. Whenever one has to
do edge detection, there is an “off-the-shelf” algorithm to use, and in some
cases specialized hardware, such as edge detection processors, can be used
to speed up the visual processing.
Once we have edges, the next thing to do is try to find objects among all
S EGMENTATION those edges. Segmentation is the process of dividing or organizing the image
into parts that correspond to continuous objects. Figure 9.5 shows an image
that has been processed by edge detection and segmentation.
But how do we know which lines correspond to which objects? And
what makes an object?
The next few sections describe several cues and approaches we can use to
detect objects.
9.3.3 Model-Based Vision

Suppose that your robot has a bunch of line drawings of chairs in its memory.
Whenever it sees an object in the environment, it performs edge detection,
which produces something like a very bad line drawing, and compares the
outcome with those stored drawings to see if any of them match what it saw
in the environment, which would indicate it saw a chair.
M ODEL -B ASED V ISION Those stored drawings are called models and the process is called model-
based vision. It is a part of a philosophy about how the brain may recognize
familiar objects, or how we might get robots to do it effectively. Model-based
vision uses models of objects, and some prior information or knowledge
about those objects, represented and stored in a way that can be used for
comparison and recognition.
Models can be stored in a variety of forms; line drawings are just one
form. Even though 2D line drawings are relatively simple and intuitive, us-
ing model matching for recognizing them is still a complex process. Here is
why: it is not enough just to compare what the robot sees (even after edge
detection) with the stored model. The robot may be looking at the object
from any angle and any distance. Therefore, to compare it effectively with
the model, it has to properly scale (change the size of) the model, and rotate
the model to try different orientations. Also, since any edge in the image
may correspond to any edge in the model, all those combinations have to be
considered and evaluated. Finally, since the robot does not know what it is
looking at, it needs to consider all models it has stored in its memory, un-
less it can somehow cleverly eliminate some of them as unlikely. All of this
is computationally intensive, taking plenty of memory (to store the models)
and processor power (to do scaling, rotation, and comparisons).
Models can vary from simple 2D line drawings to weirdly processed, math-
ematically distorted images that combine all the various views of the object
to be recognized in a mathematical way. For example, some very success-
ful face recognition systems use only a few views of the person’s face, and
then do some interesting math to produce a model that can then recognize
that person from many more points of view. Face recognition is a very popu-
lar problem in machine vision, and model-based approaches seem very well
suited for it, since faces do have repeatable features, such as two eyes, a nose,
and a mouth, with relatively constant ratios of distances in between those
features (for most people, at least). Nevertheless, we are still far from be-
ing able to reliably and efficiently recognize a particular “face in the crowd,”
whether it be your mother who has come to pick you up at the airport or a
known criminal trying to flee the country.
Face recognition is one of the important things that your brain performs
very effectively, and is in fact fine-tuned through evolution to do very well.
This is so because humans are such social animals, for whom it is very im-
portant to know who is who in order to establish and maintain social order.
Imagine if you could not recognize faces, what would your life be like? There
is a neurological disorder that produces that very deficit in some people; it
P ROSOPAGNOSIA is called prosopagnosia, from the Greek prosop for “face” and agnosia for “not
knowing.” It is rare and, so far, incurable.
Face recognition is likely to be useful in robotics, especially for robots
that are going to interact with people as their companions, helpers, nurses,
coaches, teachers, or pets. But there are many faces to learn and recognize, so
this will remain an interesting research challenge, not just for machine vision
H UMAN -R OBOT but also for the area of robotics that deals with human-robot interaction and
I NTERACTION which we will discuss in Chapter 22.
9.3.4 Motion Vision

Visual systems are often, though not always, attached to things that move
(such as people and robots, for example). The movement of the body and
M OTION V ISION the camera on it makes vision processing more challenging and motion vision
is a set of machine vision approaches that uses motion to facilitate visual

processing.
If the vision system is trying to recognize static objects, it can take advan-
tage of its own motion. By looking at an image at two consecutive time steps,
and moving the camera in between, continuous solid objects (at least those
that obey physical laws we know about) will move as one, and their bright-
ness properties will be unchanged. Therefore, if we subtract two consecutive
images from one another, what we get is the “movement” between the two,
while the objects stay the same. Notice that this depends on knowing exactly
how we moved the camera relative to the scene (the direction and distance
of the movement), and on not having anything else moving in the scene.
As we mentioned in Chapter 7, by using active perception a robot can use
movement to get a better view of something. However, moving about to
go places and moving to see better are not necessarily the same, and may
be in conflict in some cases. Therefore, a mobile robot using vision has to
make some clever decisions about how it moves, and has to subtract its own
movement from the visual image in order to see what it can see.
If other objects also move in the environment, such as other robots and
people, the vision problem becomes much harder. We will not go into any
more details about it here, but you can pursue it in the readings at the end of
this chapter.
9.3.5 Stereo Vision
Everything we have discussed so far about vision has assumed a single cam-
B INOCULAR V ISION era. Yet in nature, creatures have two eyes, giving them binocular vision. The
S TEREO V ISION main advantage of having two eyes is the ability to see in stereo. Stereo vision,
formally called binocular stereopsis, is the ability to use the combined points
of view from the two eyes or camers to reconstruct three-dimensional solid
objects and to perceive depth. The term stereo comes from the Greek stereos
meaning “solid,” and so it applies to any process of reconstructing the solid
from multiple signals.
In stereo vision, just as in motion vision (but without having to actually
move), we get two images, which we can subtract from one another, as long
as we know how the two cameras or eyes are positioned relative to each
other. The human brain “knows” how the eyes are positioned, and similarly
we, as robot designers, have control over how the cameras on the robot are
positioned as well and can reconstruct depth from the two images. So if you
can afford two cameras, you can get depth perception and reconstruct solid
objects.
This is the way 3D glasses, which make images in movies look solid, work.
In normal movies, the images come from a single projector and both of your
eyes see the same image. In 3D movies, however, there are two different
images from two different projectors. That’s why when you try to watch a
3D movie without the special glasses, it looks blurry. The two images do
not come together, on the screen or in your brain. But when you put on the
glasses, the two come together in your brain and look 3D. How? The special
glasses let only one of the projected images into each of your eyes, and your
brain fuses the images from the two eyes, as it does for everything you look
at. This does not work with normal movies because the images in your two
eyes are the same, and when they are fused, they still look the same. But the
3D movie images are different, by clever design, and when brought together,
they look better.
You might wonder how it is that those simple glasses with colored foil
lenses manage to make just one of the projected images go into each of your
eyes. That’s simple: one of the projected images is blue (or green) and one
is red, and the glasses let in one color each. Pretty easy and neat, isn’t it?
The first 3D movie using this method was made in 1922, and you can still
see the same technology used for TV, movies, and books. Today there are
more sophisticated ways of achieving the same result, with better color and
sharpness. These involve the use of something else you already know about:
polarized light, which we learned about in Chapter 8. The movies are shown
from two projectors that use different polarization, and the 3D glasses use
polarizing filters instead of the simpler color filters.
The ability to perceive in 3D using stereo is fundamental to realistic hu-
man/animal vision, and so it is involved in a variety of applications from
video games to teleoperated surgery. If you lose the use of one of your eyes,
you will lose the ability to see depth and 3D objects. To see how important
depth perception is, try catching a ball with one eye closed. It’s hard, but not
impossible. That is because your brain compensates for the loss of depth for
a little while. If you put a patch over one eye for several hours or longer, you
will start to trip and fall and reach incorrectly toward objects. Fortunately,
once you take the patch off, your brain will readjust to seeing in 3D. But if
you lose an eye permanently, it won’t. So eyes and cameras are to be treated
with care.
9.3.6 Texture, Shading, Contours
What other properties of the image can we find and use to help in object
detection?
Consider texture. Sandpaper looks quite a bit different from fur, which
looks quite a bit different from feathers, which look quite a bit different from
a smooth mirror, and so on, because all reflect the light in very different ways.
Surface patches that have uniform texture have consistent and almost iden-
tical brightness in the image, so we can assume they come from the same
object. By extracting and combining patches with uniform and consistent
texture, we can get a hint about what parts of the image may belong to the
same object in the scene.
Somewhat similarly, shading, contours, and object shape can also be used
to help simplify vision. In fact, anything at all that can be reliably extracted
from a visual image has been used to help deal with the object recognition
problem. This is true not only for machines but also (and first) for biological
vision systems, so let’s consider those now.
9.3.7 Biological Vision
The brain does an excellent job of quickly extracting the information we need
from the scene. We use model-based vision to recognize objects and people we
know. Without it, we find it hard to recognize entirely unexpected objects
or novel ones, or to orient ourselves, as in the typical example of waking up
and not knowing where you are. Biological model-based vision is of course
different from machine vision, and it is still poorly understood, but it works
remarkably well, as you can tell when you effortlessly recognize a face in the
crowd or find a lost object in a pile of other stuff.
We use motion vision in a variety of ways in order to better understand the
world around us, as well as to be able to move around while looking and
V ESTIBULAR O CULAR not having it all result in a big blur. The latter is done through the vestibular
R EFLEX (VOR) ocular reflex (VOR, in which your eyes stay fixed even though your head is
moving, in order to stabilize the image. (Go ahead and try it, move your
head around as you read the rest of this paragraph.) There has been a great
deal of research on VOR in neuroscience and machine vision and robotics,
since that ability would be very useful for robots, but it is not quite simple to
implement.
We are “hard-wired” to be sensitive to movement at the periphery of our

field of view, as well as to looming objects, because they both indicate poten-
tial danger. Like all carnivores, we have stereo vision, because it helps to find
and track prey. In contrast, herbivores have eyes on the sides of their heads,
pointing in different directions, which are effective at scanning for (carniv-
orous) predators, but whose images do not overlap or get fused together in
the way carnivore stereo vision works.
We are very good at recognizing shadows, textures, contours, and vari-
ous other shapes. In a famous experiment performed by a scientist named
Johansson in the 1970s, a few dots of light were attached to people’s clothes
and the people were videotaped as they moved in the dark, so only the move-
ment of the dots was visible. Any person watching the dots could immedi-
ately tell that they were attached to moving humans, even if only very few
light dots were used. This tells us that our brains are wired to recognize hu-
man motion, even with very little information. Incidentally, this depends on
seeing the dots/people from the side. If the view is from the top, we are not
able to readily recognize the activity. This is because our brains are not wired
to observe and recognize from a top-down view. Birds’ brains probably are,
so somebody should do that experiment.
We have gone from machine vision to human vision to bird vision, so let’s
return to robot vision now.
9.3.8 Vision for Robots

Robot vision has more stringent requirements than some other applications
of machine vision, and only slightly less demanding requirements than bi-
ological vision. Robot vision needs to inform the robot about important
things: if it’s about to fall down the stairs, if there is a human around to
help/track/avoid, if it has finished its job, and so on. Since vision process-
ing can be a very complex problem, responding quickly to the demands of
the real world based on vision information is very difficult. It is not only
impractical to try to perform all the above steps of image processing before
the robot gets run over by a truck or falls down the stairs, but fortunately it
may be unnecessary. There are good ways of simplifying the problem. Here
are some of them:
1. Use color; look for specifically and uniquely colored objects, and recog-
nize them that way (such as stop signs, human skin, etc.).
B LOB T RACKING 2. Use the combination of color and movement; this is called color blob track-
ing and is quite popular in mobile robotics. By marking important ob-

S ALIENT jects (people, other robots, doors, etc.) with salient (meaning “noticeable,”
“attention-getting”), or at least recognizable colors, and using movement
to track them, robots can effectively get their work done without having
to actually recognize objects.
3. Use a small image plane; instead of a full 512 × 512 pixel array, we can
reduce our view to much less, for example, just a line (as is used in linear
CCD cameras). Of course there is much less information in such a reduced
image, but if we are clever and know what to expect, we can process what
we see quickly and usefully.
4. Combine other, simpler and faster sensors with vision. For example, IR
cameras isolate people by using body temperature, after which vision can
be applied to try to recognize the person. Grippers allow us to touch and
move objects to help the camera get a better view. Possibilities are endless.
5. Use knowledge about the environment; if the robot is driving on a road
marked with white or yellow lines, it can look specifically for those lines
in the appropriate places in the image. This greatly simplifies following a
road, and is in fact how the first, and still some of the fastest, robot road
and highway driving is done.
Those and many other clever techniques are used in robot vision to make
it possible for robots to see what they need to see quickly enough for doing
their task.
Consider the task of autonomous or at least semiautonomous driving. This
robotics problem is gaining popularity with the auto industry, as a potential
means of decreasing the number of auto accidents. Automakers would be
glad to have cars that can make sure the driver does not swerve off the road
or into oncoming traffic. But in that task environment, everything is moving
very quickly, and there is no time for slow vision processing. This is in fact
a very exciting area of machine vision and robotics research. In 2006, several
robot cars (regular cars and vans with robotic control of the steering) man-
aged to drive completely autonomously from Los Angeles to Las Vegas, a
very long trip. The cars used vision and laser sensing to obtain the informa-
tion needed for performing the task. The next challenge being pursued is to
do something similar in urban environments.
Complex sensors imply complex processing, so they should be used se-
lectively, for tasks where they are required or truly useful. As we have said
before (and will say again):
To design an effective robot, it is necessary to have a good match between

the robot’s sensors, task, and environment.
To Summarize
• Sensor complexity is based on the amount of processing the data require.

Sensors may also have complex mechanisms, but that is not what we are
as concerned with in robotics.
• Ultrasound (sonar) sensing uses the time-of-flight principle to measure

the distance between the transducer and the nearest object(s).
• Ultrasound sensing is relatively high-power and is sensitive to specular

reflections.
• Ultrasound is used not only by robots and other machines (from sub-
marines to medical imaging equipment) but also by animals (dolphins,
whales).
• Lasers are used in ways similar to sonars, but are much faster and more
accurate, as well as much more expensive.
• Vision is the most complex and sophisticated sensory modality, both in bi-
ology and in robotics. It requires by far the most processing and provides
by far the most useful information.
• Machine vision has traditionally concerned itself with questions of recog-

nition such as “Who is that?” and “What is that?”, while robot vision has
concerned itself with questions related to action, such as “Where do I go?”
and “Can I grab that?”
• Object recognition is a complex problem. Fortunately, it can often be

avoided in robot vision.
• Motion vision, stereo vision, model-based vision, active vision, and other
strategies are employed to simplify the vision problem.
Food for Thought
• What is the speed of sound in metric units?

• How much greater is the speed of light than the speed of sound? What
does this tell you about sensors that use one or the other?
• What happens when multiple robots need to work together and all have
sonar sensors? How might you deal with their sensor interference? In
Chapter 20 we will learn about coordinating teams of robots.
• Besides using time-of-flight, the other way to use sonars is to employ the
Doppler shift. This involves examining the shift in frequency between the
sent and reflected sound waves. By examining this shift, one can very ac-
curately estimate the velocity of an object. In medical applications, sonars
are used in this way to measure blood flow, among other things. Why
don’t we use this in robotics?
• Since two eyes are much better than one, are three eyes much better, or
even any better, than two?
Looking for More?
• Check out Directed Sonar Sensing for Mobile Robot Navigation and other
work by Hugh Durrant-Whyte (Australian Center for Field Robotics) and
John Leonard (MIT) on complex sonar processing for navigation, in the
lab, outdoors, and even underwater.
• You can learn more about 3D glasses here:

http://science.howstuffworks.com/3-d-glasses2.htm.
• If you to discover the nitty gritty of the mathematics behind robot vision,
read Robot Vision by Berthold Klaus Paul Horn. A good book on machine
vision to check out is Computer Vision by Dana Ballard and Christopher
Brown. There are quite a few other books on machine/robot/computer
vision out there, so do a search on the Internet and browse.
10 Stay in Control
Feedback Control
So far, we have talked about robot bodies, including their sensors and effec-
tors. Now it’s time to turn to robot brains, the controllers that make decisions
and command the robot’s actions.
From Chapter 2 you remember that control theory was one of the founding
areas of robotics, and that feedback control is a key component of every real
robot. In this chapter we will learn the principles of feedback control and
a bit of math that is used to implement feedback controllers on any system,
from a steam engine to a modern robot.
10.1 Feedback or Closed Loop Control
F EEDBACK C ONTROL Feedback control is a means of getting a system (a robot) to achieve and main-
S ET P OINT tain a desired state, usually called the set point, by continuously comparing
its current state with its desired state.
F EEDBACK Feedback refers to the information that is sent back, literally “fed back,” into
the system’s controller.
The most popular example of a control system is the thermostat. It is a
challenge to describe concepts in control theory without resorting to ther-
mostats, but let’s try, since we already know that thermostats are not robots.
D ESIRED S TATE The desired state of the system, also called the goal state, is where the system
G OAL S TATE wants to be. Not surprisingly, the notion of goal state is quite fundamental to
goal-driven systems, and so it is used both in control theory and in AI, two
very different fields, as we saw in Chapter 2. In AI, goals are divided into
two types: achievement and maintenance.
A CHIEVEMENT G OAL Achievement goals are states the system tries to reach, such as a particular
location, perhaps the end of a maze. Once the system is there, it is done,
122 10 Stay in Control
and need not do any more work. AI has traditionally (but not exclusively)
concerned itself with achievement goals.
M AINTENANCE G OAL Maintenance goals, on the other hand, require ongoing active work on the
part of the system. Keeping a biped robot balanced and walking, for exam-
ple, is a maintenance goal. If the robot stops, it falls over and is no longer
maintaining its goal of walking while balanced. Similarly, following a wall
is a maintenance goal. If the robot stops, it is no longer maintaining its goal
of following a wall. Control theory has traditionally (but not exclusively)
concerned itself with maintenance goals.
The goal state of a system may be related to internal or external state, or a
combination of both. For example, the robot’s internal goal may be to keep
its battery power level in some desired range of values, and to recharge the
battery whenever it starts to get too low. In contrast, the robot’s external goal
state may be to get to a particular destination, such as the kitchen. Some
goal states are combinations of both, such as the goal state that requires the
robot to keep its arm extended and balance a pole. The arm state is internal
(although externally observable), and the state of the pole is external. The
goal state can be arbitrarily complex and consist of a variety of requirements
and constraints. Anything the robot is capable of achieving and maintaining
is fair game as a goal. Even what is not doable for a robot can be used as a
goal, albeit an unreachable one. The robot may just keep trying and never
get there. It’s good to strive.
So, if the system’s current and desired states are the same, it does not need
to do anything. But if they are not, which is the case most of the time, how
does the system decide what to do? That is what the design of the controller
is all about.
10.2 The Many Faces of Error
The difference between the current and desired states of a system is called
E RROR the error, and a goal of any control system is to minimize that error. Feedback
control explicitly computes and tells the system the error in order to help it
reach the goal. When the error is zero (or small enough), the goal state is
reached. Figure 10.1 shows a typical schematic of a feedback control system.
As a real world example of error and feedback, let’s consider the game
that’s sometimes called “hot and cold,” in which you have to find or guess
some hidden object, and your friends help you by saying things like “You’re
getting warmer, hotter, colder, freezing” and so on. (Ok, we are talking about
10.2 The Many Faces of Error 123
Figure 10.1 A diagram of a typical feedback controller.
temperature now, but we are still not talking about thermostats.) What they
are really doing is computing the error and giving you feedback about it.
Imagine a cruel version of the same game, in which your friends tell you
only “You are there, you win!” or “Nope, you are not there.” In that case,
what they are telling you is only if the error is zero or non-zero, if you are at
the goal state or not. This is not very much information, since it does not help
you figure out which way to go in order to get closer to the goal, to minimize
the error.
Knowing if the error is zero or non-zero is better than knowing nothing,
but not by much. In the normal version of the game, when you are told
D IRECTION OF E RROR “warm” or “cool,” you are being given the direction of the error, which allows
for minimizing the error and getting closer to the goal.
By the way, in Chapter 21 we will learn how the notion of feedback is
related to the basic principles of reinforcement learning. Replace “warm”
with “good robot” and “cool” with “bad robot,” and you will start to get the
idea of how some forms of learning work. But we have a few chapters to go
before we get there.
M AGNITUDE OF E RROR When the system knows how far off it is from the goal, it knows the mag-
nitude of error, the distance to the goal state. In the “hot and cold” game, the
gradations of freezing, chilled, cool, warm, and so on are used to indicate the
distance from (or closeness to) the goal object.
We can do better still by telling more than just the magnitude of the error.
If the system is given the precise direction and magnitude of the error, it is
told exactly what to do to get to the goal. In the game above, if you told
the person who was guessing: “Ok, turn toward the door, go through to the
kitchen, open the freezer, and look under the ice tray,” it would not be much
of a game. For robots, on the other hand, getting that much information still
does not make control or learning trivial, but it does make it easier.
As you can see, control is made easier if the robot is given plenty of feed-
back information about its error and if that information is provided accu-
rately and frequently. Let’s work this out through an example of a wall-
following robot.
10.3 An Example of a Feedback Control Robot
How would you write a controller for a wall-following robot using feed-
back control?
The first step is to consider the goal of the task. In wall-following, the goal
state is a particular distance, or range of distances, from a wall. This is a
maintenance goal, since wall-following involves keeping that distance over
time.
Given the goal, it is simple to work out the error. In the case of wall-
following, the error is the difference between the desired distance from the
wall and the actual distance at any point in time. Whenever the robot is at
the desired distance (or range of distances), it is in the goal state. Otherwise,
it is not.
We are now ready to write the controller, but before we do that, we should
consider sensors, since they will have to provide the information from which
both state and error will be computed.
What sensor(s) would you use for a wall-following robot and what in-
formation would they provide?
Would they provide magnitude and direction of the error, or just magni-
tude, or neither? For example, a contact bump sensor would provide the
least information (it’s a simple sensor, after all). Such a sensor would tell
10.3 An Example of a Feedback Control Robot 125
the robot only whether it has hit a wall; the robot could sense the wall only
through contact, not at a distance. An infra red sensor would provide infor-
mation about a possible wall, but not the exact distance to it. A sonar would
provide distance, as would a laser. A stereo vision system could also pro-
vide distance and even allow for reconstructing more about the wall, but that
would most definitely be overkill for the task. As you can see, the sensors
determine what type of feedback can be available to the robot.
Whatever sensor is used, assume that it provides the information to com-
pute distance-to-wall. The robot’s goal is to keep distance-to-wall at a par-
ticular value or, more realistically, in a particular range of values (let’s say
between 2 and 3 feet). Now we can write the robot’s feedback controller in
the form of standard if-then-else conditional statements used in program-
ming, like this one:
If distance-to-wall is in the right range,

then keep going.
If distance-to-wall is larger than desired,
then turn toward the wall,
else turn away from the wall.
Given the above controller algorithm, what will the robot’s behavior look
like? It will keep moving and switch/wiggle back and forth as it moves
along. How much switching back and forth will it do? That depends on two
parameters: how often the error is computed and how much of a correction
(turn) is made each time.
Consider the following controller:
If distance-to-wall is exactly as desired,

then keep going.
If distance-to-wall is larger than desired,
then turn by 45 degrees toward the wall,
else turn by 45 degrees away from the wall.
The above controller is not very smart. Why? To visualize it, draw a wall
and a robot and follow the rules of the above controller through a few iter-
ations to see what the robot’s path looks like. It oscillates a great deal and
rarely if ever reaches the desired distance before getting too close to or too
far from the wall.
In general, the behavior of any simple feedback system oscillates around
the desired state. (Yes, even in the case of thermostats.) Therefore, the robot
oscillates around the desired distance from the wall; most of the time it is
either too close or too far away.
How can we decrease this oscillation?
There are a few things we can do. The first is to compute the error often, so
the robot can turn often rather than rarely. Another is to adjust the turning
angle so the robot turns by small rather than large angles. Still another is to
find just the right range of distances that defines the robot’s goal. Deciding
how often to compute the error, how large a turning angle to use, and how
to define the range of distances all depend on the specific parameters of the
robot system: the robot’s speed of movement, the range of the sensor(s), and
the rate with which new distance-to-wall is sensed and computed, called the
S AMPLING R ATE sampling rate.
The calibration of the control parameters is a necessary, very important,

and time-consuming part of designing robot controllers.
Control theory provides a formal framework for how to properly use pa-
rameters in order to make controllers effective. Let’s learn how.
10.4 Types of Feedback Control

The three most used types feedback control are proportional control (P), pro-
portional derivative control (PD), and proportional integral derivative con-
trol (PID). These are commonly referred to as P , P D, and P ID control. Let’s
learn about each of them and use our wall-following robot as an example
system.
10.4.1 Proportional Control

P ROPORTIONAL The basic idea of proportional control is to have the system respond in propor-
C ONTROL tion to the error, using both the direction and the magnitude of the error. A
proportional controller produces an output o proportional to its input i, and
is formally written as:
o = Kp i
Kp is a proportionality constant, usually a constant that makes things work,
and is specific to a particular control system. Such parameters are abundant
in control; typically you have to figure them out by trial and error and cali-
bration.
What would a proportional controller for our wall fallowing robot look
like?
It would use distance-to-wall as a parameter to determine the angle and
distance and/or speed with which the robot would turn. The larger the error,
the larger the turning angle and speed and/or distance; the smaller the error,
the smaller the turning angle and speed and/or distance.
In control theory, the parameters that determine the magnitude of the sys-
G AIN tem’s response are called gains. Determining the right gains is typically very
difficult, and requires trial and error, testing and calibrating the system re-
peatedly. In some cases, if the system is very well understood, gains can be
computed mathematically, but that is rare.
If the value of a particular gain is proportional to that of the error, it is
P ROPORTIONAL G AIN called proportional gain. As we saw in the case of our wall-following robot, in-
correct gain values cause the system to undershoot or overshoot the desired
state. The gain values determine whether the robot will keep oscillating or
will eventually settle on the desired state.
D AMPING Damping refers to the process of systematically decreasing oscillations. A
system is properly damped if it does not oscillate out of control, meaning its
oscillations are either completely avoided (which is very rare) or, more prac-
tically, the oscillations gradually decrease toward the desired state within a
reasonable time period. Gains have to be adjusted in order to make a system
properly damped. This is a tuning process that is specific to the particular
control system (robot or otherwise).
When adjusting gains, you have to keep in mind both the physical and
the computational properties of the system. For example, how the motor
responds to speed commands plays a key part in control, as do the back-
lash and friction in the gears (remember Chapter 4?), and so on. Physical
properties of the robot influence the exact values of the gains, because they
A CTUATOR constrain what the system actually does in response to a command. Actuator
U NCERTAINTY uncertainty makes it impossible for a robot (or a human, for that matter) to
know the exact outcome of an action ahead of time, even for a simple action
such as “Go forward three feet.” While actuator uncertainty keeps us from
predicting the exact outcome of actions, we can use probability to estimate
and make a pretty good guess, assuming we know enough about the sys-
tem to set up the probabilities correctly. See the end of the chapter for more
reading on probabilistic robotics.
Figure 10.2 Different trajectories produced by P, PD, and PID feedback controllers.
10.4.2 Derivative Control

As we have seen, setting gains is difficult, and simply increasing the pro-
portional gain does not remove oscillatory problems from a control system.
While this may work for small gains (called low gains; gain is referred to
as high or low), as the gain increases, the system’s oscillations increase along
with it. The basic problem has to do with the distance from the set point/des-
ired state: When the system is close to the desired state, it needs to be controlled
differently than when it is far from it. Otherwise, the momentum generated
by the controller’s response to the error, its own correction, carries the sys-
tem beyond the desired state and causes oscillations. One solution to this
problem is to correct the momentum as the system approaches the desired
state.
First, let’s remember what momentum is:
M omentum = mass ∗ velocity
Since momentum and velocity are directly proportional (the faster you
move and/or the bigger you are, the more momentum you have), we can
control momentum by controlling the velocity of the system. As the system
nears the desired state, we subtract an amount proportional to the velocity:
−(gain ∗ velocity)
D ERIVATIVE T ERM This is called the derivative term, because velocity is the derivative (the rate
of change) of position. Thus, a controller that has a derivative term is called
a D controller.
A derivative controller produces an output o proportional to the derivative
of its input i:
o = Kd dt
di
Kd is a proportionality constant, as before, but this time with a different

name, so don’t assume you can use the same number in both equations.
The intuition behind derivative control is that the controller corrects for
the momentum of the system as it approaches the desired state. Let’s apply
this idea to our wall-following robot. A derivative controller would slow the
robot down and decrease the angle of its turning as its distance from the wall
gets closer to the desired state, the optimal distance to the wall.
10.4.3 Integral Control

Yet another level of improvement can be made to a control system by intro-
I NTEGRAL T ERM ducing the so-called integral term, or I. The idea is that the system keeps
track of its own errors, in particular of the repeatable, fixed errors that are
S TEADY S TATE E RROR called steady state errors. The system integrates (sums up) these incremental
errors over time, and once they reach some predetermined threshold (once
the cumulative error gets large enough), the system does something to com-
pensate/correct.
An integral controller produces an output o proportional to the integral of
its input i:

o = Kf i(t)dt
Kf is a proportionality constant. You get the pattern.
How do we apply integral control to our wall-following robot?
Actually, it’s not too easy, because there is no obvious place for steady state
error to build up in the simple controller we came up with. That’s a good
thing about our controller, so we can pat ourselves on the back and take
another example. Consider a lawn-mowing robot which carefully covers the
complete lawn by going from one side of the yard to the other, and moving
over by a little bit every time to cover the next strip of grass in the yard. Now
suppose that the robot has some consistent error in its turning mechanism, so
whenever it tries to turn by a 90-degree angle to move over to the next strip
on the grass, it actually turns by a smaller angle. This makes it fail to cover
the yard completely, and the longer the robot runs, the worse its coverage of
the yard gets. But if it has a way of measuring its error, even only once it gets
large enough (for example, by being able to detect that it is mowing already
mowed grass areas a lot), it can apply integral control to recalibrate.
Now you know about P , I, and D, the basic types of feedback control.
Most real-world systems actually use combinations of those three basic types;
P D and P ID controllers are especially prevalent, and are most commonly
used in industrial applications. Let’s see what they are.
10.4.4 PD and PID Control

P D control is a combination, actually simply a sum, of proportional (P ) and
derivative (D) control terms:
o = Kp i + Kd dt
di
P D control is extremely useful and applied in most industrial plants for

process control.
P ID control is a combination (yes, again a sum) of proportional P , integral
I, and derivative D control terms:

o = K p i + Kf i(t)dt + Kd dt
di
Figure 10.2 shows example trajectories that would be produced by our

wall-following robot if it were controlled by P, PD, and PID feedback con-
trollers.
You can learn a great deal more about feedback control and the larger topic
of control theory from the sources given at the end of this chapter. It’s a large
field studied in electrical engineering and mechanical engineering.
Before we move on, we should put control theory into perspective rel-
ative to robotics. As you have been reading and will read much more in
the chapter to come, getting robots to do something useful requires many
components. Feedback control plays a role at the low level, for example,
for controlling the wheels or other continuously moving actuators. But for
other aspects of robot control, in particular achieving higher-level goals (nav-
igation, coordination, interaction, collaboration, human-robot interaction),
Figure 10.3 A diagram of a typical feedforward controller.
other approaches must come into play that are better suited for representing
and handling those challenges. Those levels of robot control use techniques
that come from the field of artificial intelligence, but that have come a long
way from what AI used to be about in its early days, as we learned in Chap-
ter 2. That’s coming up, so stick around.
10.5 Feedforward or Open Loop Control
C LOSED L OOP Feedback control is also called closed loop control because it closes the loop
C ONTROL between the input and the output, and provides the system with the error as
feedback.
What is an alternative to closed loop control?
You guessed it (or you read the section title): the alternative to feedback or
closed loop control is called feedforward or open loop control. As the name
O PEN L OOP C ONTROL implies, open loop control or feedforward control, does not use sensory feedback,
F EEDFORWARD and state is not fed back into the system. Thus the loop between the input
C ONTROL
and output is open, and not really a loop at all. In open loop control, the
system executes the command that is given, based on what was predicted,
instead of looking at the state of the system and updating it as it goes along,
as in feedback control. In order to decide how to act in advance, the controller
determines set points or sub-goals for itself ahead of time. This requires look-
ing ahead (or looking forward) and predicting the state of the system, which
is why the approach is called feedforward. Figure 10.3 shows a schematic of
a feedforward open loop controller.
Open loop or feedforward control systems can operate effectively if they
are well calibrated and their environment is predictable and does not change
in a way that affects their performance. Therefore they are well suited for
repetitive, state-independent tasks. As you have probably guessed, those are
not very common in robotics.
To Summarize
• Feedback/closed loop control and feedforward/open loop control are im-

portant aspects of robotics.
• Feedback control aims to minimize system error, the difference between

the current state and the desired state.
• The desired state or goal state is a concept used in AI as well as control

theory, and can come in the form of achievement or maintenance.
• Error is a complex concept with direction and magnitude, and cumulative

properties.
• Proportional, derivative, and integral control are the basic forms of feed-
back control, which are usually used in combination in real-world control
systems (robotic and otherwise).
Food for Thought
• What happens when you have sensor error in your system? What if your
sensor incorrectly tells you that the robot is far from a wall but in fact it is
not? What about vice versa? How might you address these issues?
• What can you do with open loop control in your robot? When might it be
useful?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Feedback_Control
• Here are some textbooks on control theory, if you want to learn (much)
more:
– Signals and Systems by Simon Haykin and Barry Van Veen, John Wiley.
– Intelligent Control Systems, Theory and Applications edited by Madan M.

Gupta and Naresh K. Sinha.
– Linear Control System Analysis and Design: Conventional and Modern by
J. J. D’Azzo and C. Houpis.
– Automatic Control Systems by B. C. Kuo.
11 The Building Blocks of Control
Control Architectures
The job of the controller is to provide the brains for the robot and so it can
be autonomous and achieve goals. We have seen that feedback control is
a very good way to write controllers for making a robot perform a single
behavior, such as following a wall, avoiding obstacles, and so on. Those
behaviors do not require much thinking, though. Most robots have more
things to do than just follow a wall or avoid obstacles, ranging from simple
survival (not running into things or running out of power) to complex task
achievement (whatever it may be). Having to do multiple things at once and
deciding what to do at any point in time is not simple, even for people, much
less for robots. Therefore, putting together controllers that will get the robot
to produce the desired overall behavior is not simple, but it is what robot
control is really about.
So how would you put multiple feedback controllers together? What if
you need more than feedback control? How would you decide what is
needed, which part of the control system to use in a given situation and
for how long, and what priority to assign to it?
I hope you had no hard-and-fast answers to the above questions, because
there aren’t any, in general. These are some of the biggest challenges in robot
control, and we are going to learn how to deal with them in this and the next
few chapters.
11.1 Who Needs Control Architectures?
Just putting rules or programs together does not result in well-behaved robots,
through it may be a lot of fun, as long as the robots are small and not danger-
ous. While there are numerous different ways in which a robot’s controller
136 11 The Building Blocks of Control
can be programmed (and there are infinitely many possible robot control
programs), most of them are pretty bad, ranging from completely incorrect
to merely inefficient. To find a good (correct, efficient, even optimal) way to
control a given robot for a given task, we need to know some guiding prin-
ciples of robot control and the fundamentally different ways in which robots
can be programmed. These are captured in robot control architectures.
C ONTROL A robot control architecture provides guiding principles and constraints for
A RCHITECTURE organizing a robot’s control system (its brain). It helps the designer to pro-
gram the robot in a way that will produce the desired overall output behav-
ior.
The term architecture is used here in the same way as in “computer archi-
tecture”, where it means the set of principles for designing computers out of
a collection of well-understood building blocks. Similarly, in robot architec-
tures, there is a set of building blocks or tools at your disposal to make the
job of robot control design easier. Robot architectures, like computer architec-
tures, and of course “real” building architectures, where the term originally
comes from, use specific styles, tools, constraints, and rules.
To be completely honest, you don’t have to know anything about control
architectures to get a robot to do something. So why should we bother with
this and the next few chapters? Because there is more to robotics than getting
simple robots to do simple things. If you are interested in how you can get
a complex robot (or a team of robots) to do something useful and robust in a
complicated environment, knowing about control architectures is necessary
to help you get you there. Trial and error and intuition go only so far, and
we’d like to take robotics very far indeed, far beyond that.
You might be wondering exactly what we mean by robot control, since so
far we have seen that it involves hardware, signal processing, and computa-
tion. That’s true in general: the robot’s “brain” can be implemented with a
conventional program running on a microprocessor, or it may be embedded
in hardware, or it may be a combination of the two. The robot’s controller
does not need to be a single program on a single processor. In most robots
it is far from that, since, as we have seen, there is a lot going on in a robot.
Sensors, actuators, and decisions need to interact in an effective way to get
the robot to do its job, but there is typically no good reason to have all those
elements controlled from a single centralized program, and there are many
reasons not to.
Can you think of some of those reasons?

11.2 Languages for Programming Robots 137
Robustness to failure, for one thing; if the robot is controlled in a central-

ized fashion, the failure of that one processor makes the whole robot stop
working. But before we think about how to spread out the different parts of
the robot’s brain, let’s get back to designing robot control programs, wher-
ever on the robot they may be running.
Robot control can take place in hardware and in software, but the more
complex the controller is, the more likely it is to be implemented in soft-
ware. Why?
Hardware is good for fast and specialized uses, and software is good for
flexible, more general programs. This means that complicated robot brains
typically involve computer programs of some type or another, running on
the robot in real time. The brain should be physically on the robot, but that
may be just prejudice based on how biological systems do it. If wireless radio
communication is reliable enough, some or all of the processing could reside
off the robot. The trouble is that communication is never perfectly reliable,
so it is much safer to keep your brain with you at all times (keep your head
on your shoulders), for robots and people.
Brains, robotic or natural, use their programs to solve problems that stand
in the way of achieving their goals and getting their jobs done. The process of
solving a problem using a finite (not endless) step-by-step procedure is called
A LGORITHM an algorithm, and is named for the Iranian mathematician, Al-Khawarizmi (if
that does not seem similar at all, it may be because we are not pronouncing
his name correctly). The field of computer science devotes a great deal of
research to developing and analyzing algorithms for all kinds of uses, from
sorting numbers to creating, managing, and sustaining the Internet. Robotics
also concerns itself with the development and analysis of algorithms (among
many other things, as we will learn) for uses that are relevant to robots, such
as navigation (see Chapter 19), manipulation (see Chapter 6), learning (see
Chapter 21), and many others (see Chapter 22). You can think of algorithms
as the structure on which computer programs are based.
11.2 Languages for Programming Robots
So robot brains are computer programs and they are written using program-
ming languages. You may be wondering what the best robot programming
language is. Don’t waste your time: there is no “best” language. Robot pro-
grammers use a variety of languages, depending on what they are trying
to get the robot to do, what they are used to, what hardware comes with
the robot, and so on. The important thing is that robot controllers can be
implemented in various languages.
Any programming language worth its beans is called “Turing universal,”
which means that, in theory at least, it can be used to write any program.
This concept was named after Alan Turing, a famous computer scientist from
England who did a great deal of foundational work in the early days of com-
puter science, around World War II. To be Turing universal, a programming
language has to have the following capabilities: sequencing (a then b then
c), conditional branching (if a then b else c), and iteration (for a=1 to 10 do
something). Amazingly, with just those, any language can compute every-
thing that is computable. Proving that, and explaining what it means for
something to be computable, involves some very nice formal computer sci-
ence theory that you can learn about from the references at the end of the
chapter.
The good news from all that theory is that programming languages are
just tools; you can use them to program various things, from calculators to
airline scheduling to robot behaviors, and everything in between. You can
use any programming language to write any program, at least in theory. In
practice, of course, you should be smart and picky, and choose the language
that is well suited to the programming task. (You have no doubt noticed this
repeating theme about finding suitable designs for robot sensors and bodies
and controllers, and suitable languages for programming the robots, and so
on. It’s a basic principle of good engineering of any kind.)
If you have programmed even a little bit, you know that programming
languages come in a great variety and are usually specific for particular uses.
Some are good for programming Web pages, others for games, and still oth-
ers for programming robots, which is what we really care about in this book.
So although there is no best language for programming robots, some lan-
guages are better for it than others. As robotics is growing and maturing
as a field, there are more and more specialized programming languages and
tools.
Various programming languages have been used for robot control, rang-
ing from general-purpose to specially designed ones. Some languages have
been designed specifically to make the job of programming particular robot
control architectures easier. Remember, architectures exist to provide guid-
ing principles for good robot control design, so it is particularly convenient
if the programming language can make following those principles easy and
failing to follow them hard.
Figure 11.1 A way to visualize the relationship among control architectures, con-
trollers, and programming languages.
Figure 11.1 shows a way in which you can think about the relationship
among control architectures, robot controllers, and programming languages.
They are all different, yet are needed to make a robot do what you’d like it to
do.
11.3 And the Architectures are...
Regardless of which language is used to program a robot, what matters is the

control architecture used to implement the controller, because not all archi-
tectures are the same. On the contrary, as you will see, architectures impose
strong rules and constraints on how robot programs are structured, and the
resulting control software ends up looking very different.
We have already agreed that there are numerous ways you can program a
robot, and there are fewer, but still many, ways in which you can program a
robot well. Conveniently, all those effective robot programs fall into one of
the known types of control architectures. Better still, there are very few types
of control (that we know of so far). They are:
1. Deliberative control
2. Reactive control
3. Hybrid control
4. Behavior-based control.
In the next few chapters we will study each of these architectures in detail,
so you will know what’s good and bad about them and which one to choose
for a given robot and task.
Because robotics is a field that encompasses both engineering and science,
there is a great deal of ongoing research that brings new results and discover-
ies. At robotics conferences, there are presentations, full of drawings of boxes
and arrows, describing architectures that may be particularly well suited for
some robot control problem. That’s not about to go away, since there are so
many things robots could do, and so much yet to be discovered. But even
though people may keep discovering new architectures, all of them, and all
the robot programs, fit into one of the categories listed above, even if the
programmer may not realize it.
In most cases, it is impossible to tell, just by observing a robot’s behavior,
what control architecture it is using. That is because multiple architectures
can get the job done, especially for simple robots. As we have said before,
when it comes to more complex robots, the control architecture becomes very
important.
Before we launch into the details of the different types of control architec-
tures and what they are good for, let’s see what are some important questions
to consider that can help us decide which architecture to use. For any robot,
task, and environment, many things need to be considered, including these:
• Is there a lot of sensor noise?
• Does the environment change or stay static?
• Can the robot sense all the information it needs? If not, how much can it
sense?
• How quickly can the robot sense?
• How quickly can the robot act?
• Is there a lot of actuator noise?
• Does the robot need to remember the past in order to get the job done?
• Does the robot need to think into the future and predict in order to get the
job done?
• Does the robot need to improve its behavior over time and be able to learn
new things?
Control architectures differ fundamentally in the ways they treat the fol-
lowing important issues:
• Time: How fast do things happen? Do all components of the controller

run at the same speed?
• Modularity: What are the components of the control system? What can
talk to what?
• Representation: What does the robot know and keep in its brain?
Let’s talk about each of these briefly. We will spend more time understand-
ing them in detail in the subsequent chapters.
11.3.1 Time
T IME -S CALE Time, usually called time-scale, refers to how quickly the robot has to respond
to the environment compared with how quickly it can sense and think. This
is a key aspect of control and therefore a major influence on the choice of
what architecture should be used.
The four basic architecture types differ significantly in how they treat time.
D ELIBERATIVE Deliberative control looks into the future, so it works on a long time-scale
C ONTROL (how long depends on how far into the future it looks). In contrast, reactive
R EACTIVE C ONTROL
control responds to the immediate, real-time demands of the environment
without looking into the past or the future, so it works on a short time-scale.
H YBRID C ONTROL Hybrid control combines the long time-scale of deliberative control and the
short time-scale of reactive control, with some cleverness in between. Finally,
B EHAVIOR -B ASED behavior-based control works to bring the time-scales together. All of this will
C ONTROL make more sense as we spend time getting to know each type of control and
the associated architectures.
11.3.2 Modularity
M ODULARITY Modularity refers to the way the control system (the robot’s program) is bro-
ken into pieces or components, called modules, and how those modules in-
teract with each other to produce the robot’s overall behavior. In deliberative
control, the control system consists of multiple modules, including sensing

(perception), planning, and acting, and the modules do their work in se-
quence, with the output of one providing the input for the next. Things
happen one at a time, not at the same time, as we will see in Chapter 13.
In reactive control, things happen at the same time, not one at a time. Multi-
ple modules are all active in parallel and can send messages to each other in
various ways, as we will learn in Chapter 14. In hybrid control, there are three
main modules to the system: the deliberative part, the reactive part, and the
part in between. The three work in parallel, at the same time, but also talk
to each other, as we will learn in Chapter 15. In behavior-based control, there
are usually more than three main modules that also work in parallel and
talk to each other, but in a different way than in hybrid systems, as we will
cover in Chapter 16. So as you can see, how many modules there are, what is
in each module, whether the modules work sequentially or in parallel, and
which modules can talk to which others are distinguishing features of control
architectures.
11.3.3 Representation
Finally, representation. That’s a tough one to summarize, so we’ll give it a
whole chapter, coming up next, to do it justice.
So how do we know which architecture to use for programming the
controller of a particular robot?
Let’s learn more about the different architectures so we can answer that
question. After talking about representation, in the next few chapters we
will study each of the four basic architecture types: deliberative, reactive,
hybrid, and behavior-based.
To Summarize
• Robot control can be done in hardware and/or in software. The more

complex the robot and its task, the more software control is needed.
• Control architectures provide guiding principles for designing robot pro-

grams and control algorithms.
• There is no one best robot programming language. Robots can be pro-

grammed with a variety of languages, ranging from general-purpose to
special-purpose.
• Special-purpose robot programming languages can be written to facilitate

programming robots in particular control architectures.
• Robot control architectures differ substantially in how they handle time,

modularity, and representation.
• The main robot control architectures are deliberative (not used), reactive,
hybrid, and behavior-based.
Food for Thought
• How important is the programming language? Could it make or break a

particular gadget, device, or robot?
• With the constant development of new technologies that use computa-

tion, do you think there will be increasingly more or increasingly fewer
programming languages?
Looking for More?
• Introduction to the Theory of Computation by Michael Sipser is an excellent

textbook for learning about the very abstract but very elegant topics only
touched on in this chapter.
• One of the most widely used textbooks for learning about programming
languages is Structure and Interpretation of Computer Programs by Harold
Abelson and Gerald Jay Sussman. It’s actually fun to read, although not
necessarily easy to understand.
12 What’s in Your Head?
Representation
In many tasks and environments, the robot cannot immediately sense every-
thing it needs to know. It is therefore sometimes useful to remember what
happened in the past, or to try to predict what will happen in the future. It is
also sometimes useful to store maps of the environment, images of people or
places, and various other information that will be useful for getting the job
done.
R EPRESENTATION Representation is the form in which information is stored or encoded in the
robot.
Representation is more than memory. In computer science and robotics,
M EMORY we think of memory as the storage device used to keep information. Just
referring to memory does not say anything about what is stored and how it
is encoded: is it in the form of numbers, names, probabilities, x,y locations,
distances, colors? Representation is what encodes those important features
of what is in the memory.
What is represented and how it is represented has a major impact on robot
control. This is not surprising, as it is really the same as saying “What is in
your brain influences what you can do.” In this chapter we will learn about
what representation is all about and why it has such an important role in
robot control.
How does internal state, the information a robot system keeps around
(remember Chapter 3), relate to representation? Is it the same thing? If
not, what is it?
In principle, any internal state is a form of representation. In practice,
what matters is the form and function of that representation, how it is stored
146 12 What’s in Your Head?
Figure 12.1 A few possible representation options for a maze-navigating robot.
and how it is used. “Internal state” usually refers to the “status” of the sys-
tem itself, whereas “representation” refers to arbitrary information about the
world that the robot stores.
12.1 The Many Ways to Make a Map
W ORLD M ODEL Representation of the world is typically called a world model. A map (as used
in navigation; see Chapter 19) is the most commonly used example of a world
model. To give an illustration of how the representation of a particular world,
its map, can vary in form, consider the problem of exploring a maze.
What can the robot store/remember to help it navigate a maze?
• The robot may remember the exact path it has taken to get to the end of
the maze (e.g., “Go straight 3.4 cm, turn left 90 degrees, go straight 12
cm, turn right 90 degrees.”). This remembered path is a type of map for
getting through the maze. This is an odometric path.
• The robot may remember a sequence of moves it made at particular land-

marks in the environment (e.g., “Left at the first junction, right at the sec-
12.1 The Many Ways to Make a Map 147
ond junction, straight at the third.”). This is another way to store a path
through the maze. This is a landmark-based path.
• The robot may remember what to do at each landmark in the maze (e.g.,
“At the green/red junction go left, at the red/blue junction go right’, at
the blue/orange junction go straight.”). This is a landmark-based map;
it is more than a path since it tells the robot what to do at each junction,
no matter in what order it reaches the junction. A collection of landmarks
T OPOLOGICAL M AP connected with links is called a topological map because it describes the
topology, the connections, among the landmarks. Topological maps are
very useful. (Don’t confuse them with topographic maps, which are some-
thing completely different, having to do with representing the elevation
levels of the terrain.)
• The robot may remember a map of the maze by “drawing it” using exact
lengths of corridors and distances between walls it sees. This is a metric
map of the maze, and it is also very useful.
The above are not nearly all the ways in which the robot can construct and
store a model of the maze. However, the four types of models above already
show you some important differences in how representations can be used.
The first model, the odometric path, is very specific and detailed, and that
it is useful only if the maze never changes, no junctions become blocked or
opened, and if the robot is able to keep track of distances and turns very ac-
curately. The second approach also depends on the map not changing, but it
does not require the robot to be as specific about measurements, because it
relies on finding landmarks (in this case, junctions). The third model is simi-
lar to the second, but connects the various paths into a landmark-based map,
a network of stored landmarks and their connections to one another. Finally,
the fourth approach is the most complicated because the robot has to mea-
sure much more about the environment and store much more information.
On the other hand, it is also the most generally useful, since with it the robot
can now use its map to think about other possible paths in case any junctions
become blocked.
Figure 12.1 illustrates a few more possible representations that could be
used by a maze-following robot: no representation at all, a typical 2D map, a
visual image of the environment, and a graph of the maze structure.
12.2 What Can the Robot Represent?

Maps are just one of the many things a robot may want to represent or store
or model in its “brain.” For example, the robot may want to remember how
long its batteries last, and to remind itself to recharge them before it is too
late. This is a type of a self-model. Also, the robot may want to remember
that traffic is heavy at particular times of day or in particular places in the
environment, and to avoid traveling at those times and in those places. Or
it may store facts about other robots, such as which ones tend to be slow or
fast, and so on.
There are numerous aspects of the world that a robot can represent and
model, and numerous ways in which it can do it. The robot can represent
information about:
• Self: stored proprioception, self-limitations, goals, intentions, plans
• Environment: navigable spaces, structures
• Objects, people, other robots: detectable things in the world
• Actions: outcomes of specific actions in the environment
• Task: what needs to be done, where, in what order, how fast, etc.
You can already see that it could take quite a bit of sensing, computation,
and memory to acquire and store some types of world models. Moreover,
it is not enough simply to get and store a world model; it is also necessary
to keep it accurate and updated, or it becomes of little use or, worse yet,
becomes misleading to the robot. Keeping a model updated takes sensing,
computation, and memory.
Some models are very elaborate; they take a long time to construct and
are therefore kept around for the entire lifetime of the robot’s task. Detailed
metric maps are such models. Other models, on the other hand, may be
relatively quickly constructed, briefly used, and soon discarded. A snapshot
of the robot’s immediate environment showing the path out of the clutter
through the nearest door is an example of such a short-term model.
12.3 Costs of Representing

In addition to constructing and updating a representation, using a represen-
tation is also not free, in terms of computation and memory cost. Consider
12.3 Costs of Representing 149
maps again: to find a path from a particular location in the map to the goal
location, the robot must plan a path. As we will learn in more detail in Chap-
ter 19, this process involves finding all the free/navigable spaces in the map,
then searching through those to find any path, or the best path, to the goal.
Because of the processing requirements involved in the construction, main-
tenance, and use of representation, different architectures have very different
properties based on how representation is handled, as we will see in the next
few chapters. Some architectures do not facilitate the use of models (or do
not allow them at all, such as reactive ones), others utilize multiple types of
models (hybrid ones), still others impose constraints on the time and space
allowed for the models being used (behavior-based ones).
How much and what type of representation a robot should use depends
on its task, its sensors, and its environment. As we will see when we learn
more about particular robot control architectures, how representation is han-
dled and how much time is involved in handling it turns out to be a crucial
distinguishing feature for deciding what architecture to use for a given robot
and task. So, as you would imagine, what is in the robot’s head has a very
big influence on what that robot can do.
To Summarize
• Representation is the form in which information is stored in the robot.
• Representation of the world around the robot is called a world model.
• Representation can take a great variety of forms and can be used in a great
variety of ways by a robot.
• A robot can represent information about itself, other robots, objects, peo-
ple, the environment, tasks, and actions.
• Representations require constructing and updating, which have compu-

tational and memory costs for the robot.
• Different architectures treat representation very differently, from not hav-

ing any at all, to having centralized world models, to having distributed
ones.
Food for Thought
• Do you think animals use internal models? What about insects?
• Why might you not want to store and use internal models?
13 Think Hard, Act Later
Deliberative Control
Deliberation refers to thinking hard; it is defined as "thoughtfulness in deci-

sion and action." Deliberative control grew out of early artificial intelligence
(AI). As you remember from the brief history of robotics (Chapter 2), in those
days, AI was one of the main influences on how robotics was done.
In AI, deliberative systems were (and sometimes still are) used for solv-
ing problems such as playing chess, where thinking hard is exactly the right
thing to do. In games, and in some real-world situations, taking time to con-
sider possible outcomes of several actions is both affordable (there is time
to do it) and necessary (without strategy, things go bad). In the 1960s and
1970s, AI researchers were so fond of this type of reasoning, they theorized
that the human brain works this way, and so should robot control. As you
recall from Chapter 2, in the 1960s, early AI-based robots often used vision
sensors, which require a great deal of processing, and so it was worth it for a
robot to take the time (quite a lot of time back then, with those slow proces-
sors) to think hard about how to act given how difficult it was to make sense
out of the environment. The robot Shakey, a forerunner of many AI-inspired
robotics projects that followed, used the then state of the art in machine vi-
sion as input into a planner in order to decide what to do next, how and
where to move.
152 13 Think Hard, Act Later
Figure 13.1 A closeup of a maze the robot (marked with a black circle) needs to
navigate.
13.1 What Is Planning?
So what is planning?
P LANNING Planning is the process of looking ahead at the outcomes of the possible
actions, and searching for the sequence of actions that will reach the desired
goal.
S EARCH Search is an inherent part of planning. It involves looking through the

available representation “in search of” the goal state. Sometimes searching
the complete representation is necessary (which can be very slow, depending
on the size of the representation), while at other times only a partial search is
enough, to reach the first found solution.
For example, if a robot has a map of a maze, and knows where it is and
13.1 What Is Planning? 153
Figure 13.2 A graph of the maze, showing the two possible paths to the charger.
One of the paths is shorter/more optimal than the other.
where it wants to end up (say at a recharging station at the end of the maze),
it can then plan a path from its current position to the goal, such as those
shown in figure 13.2. The process of searching through the maze happens in
the robot’s head, in the representation of the maze (such as the one shown
in figure 13.1, not in the physical maze itself. The robot can search from
the goal backwards, or from where it is forward. Or it can even search in
both directions in parallel. That is the nice thing about using internal models
or representations: you can do things that you can’t do in the real world.
Consider the maze in figure 13.1, where the robot is marked with a black
circle, and the goal with the recharging battery. Note that at each junction in
the maze, the robot has to decide how to turn.
The process of planning involves the robot trying different turns at each
junction, until a path leads it to the battery. In this particular maze, there is
more than one path to the goal from where the robot is, and by searching
the entire maze (in its head, of course) the robot can find both of those paths,
and then choose the one it likes better. Usually the shortest path is considered
the best, since the robot uses the least time and battery power to reach it. But
in some cases other criteria may be used, such as which path is the safest
or the least crowded. The process of improving a solution to a problem by
O PTIMIZATION finding a better one is called optimization. As in the maze example above,
various values or properties of the given problem can be optimized, such as
O PTIMIZATION the distance of a path; those values are called optimization criteria. Usually
C RITERIA some optimization criteria conflict (for example, the shortest path may also
be the most crowded), so deciding what and how to optimize is not simple.
O PTIMIZING S EARCH Optimizing search looks for multiple solutions (paths, in the case of the maze),
in some cases all possible paths.
In general, in order to use search and plan a solution to a particular prob-
lem, it is necessary to represent the world as a set of states. In the maze
example, the states are the corridors, the junctions, the start (where the search
starts), the current state (wherever the robot happens to be at the time), and
the goal (where the robot wants to end up). Then search is performed to find
a path that can take the robot from the current state to the goal state. If the
robot wants to find the very best, optimal path, it has to search for all possi-
ble paths and pick the one that is optimal based on the selected optimization
criterion (or criteria, if more than one is used).
13.2 Costs of Planning
For small mazes like the one used here, planning is easy, because the state
space is small. But as the number of possible states becomes large (as in
chess, for example), planning becomes slow. The longer it takes to plan, the
longer it takes to solve the problem. In robotics, this is particularly important
since a robot must be able to avoid immediate danger, such as collisions with
objects. Therefore, if path planning takes too long, the robot either has to stop
and wait for planning to finish before moving on, or it may risk collisions or
running into blocked paths if it forges ahead without a finished plan. This is
not just a problem in robotics, of course; whenever a large state space is involved,
planning is difficult. To deal with this fundamental problem, AI researchers
have found various ways to speed things up. One popular approach is to
use hierarchies of states, where first only a small number of "large," "coarse,"
or "abstract" states is considered; after that, more refined and detailed states
are used in the parts of the state space where it really matters. The graph of
the maze shown in figure 13.2 is an example of doing exactly that: it lumps
all states within a corridor of the maze together and considers them as a
single corridor state in the graph. There are several other clever methods
Figure 13.3 A diagram of a classical deliberative architecture, showing the sequence

of sense-plan-act (SPA) components.
besides this one that can speed up search and planning. These are all opti-
mization methods for planning itself, and they always involve some type of
compromise.
Since the early days of AI, computing power has greatly improved and
continues to improve (as per Moore’s Law; see the end of the chapter for
more information). This means larger state spaces can be searched much
faster than ever before. However, there is still a limit to what can be done
R EAL T IME in real time, the time in which a physical robot moves around in a dynamic
environment.
In general, a real robot cannot afford to just sit and deliberate. Why?
Deliberative, planner-based architectures involve three steps that need to

be performed in sequence:
1. Sensing (S)
2. Planning (P)
3. Acting (A), executing the plan.

SPA For this reason, deliberative architectures, which are also called SPA (sense-
(S ENSE -P LAN -A CT ) plan-act) architectures, shown in figure 13.3, have serious drawbacks for robotics.
A RCHITECTURES
Here is what they are.
Drawback 1: Time-Scale
As we have said above, it can take a very long time to search in large
state spaces. Robots typically have collections of sensors: some simple digi-
tal sensors (e.g., switches, IRs), some more complex ones (e.g., sonars, lasers,
cameras), some analog sensors (e.g., encoders, gauges). The combined inputs
from these sensors in themselves constitute a large state space. When com-
bined with internal models or representations (maps, images of locations,
previous paths, etc.), the result is a state space that is large and slow to search.
If the planning process is slow compared with the robot’s moving speed,
it has to stop and wait for the plan to be finished, in order to be safe. So
to make progress, it is best to plan as rarely as possible and move as much
as possible in between. This encourages open loop control (see Chapter 10),
which we know is a bad idea in dynamic environments. If planning is fast,
then execution need not be open loop, since replanning can be done at each
step; unfortunately, typically this is impossible for real-world problems and
robots.
Generating a plan for a real environment can be very slow.
Drawback 2: Space
It may take a great deal of space (i.e., memory storage) to represent and
manipulate the robot’s state space representation. The representation must
contain all information needed for planning and optimization, such as dis-
tances, angles, snapshots of landmarks, views, and so on. Computer memory
is comparatively cheap, so space is not as much of a problem as time, but all
memory is finite, and some algorithms can run out of it.
Generating a plan for a real environment can be very memory-intensive.

Drawback 3: Information
The planner assumes that the representation of the state space is accurate
and up to date. This is a reasonable assumption, because if the representa-
tion is not accurate and updated, the resulting plan is useless. For example,
if a junction point in the maze is blocked, but the robot’s internal map of the
maze does not show it as blocked, the planned path might go through that
junction and would therefore be invalid, but the robot would not discover
that until it physically got to the junction. Then it would have to backtrack
to a different path. The representation used by the planner must be updated
and checked as often as necessary to keep it sufficiently accurate for the task.
Thus, the more information the better.
Generating a plan for a real environment requires updating the world

model, which takes time.
Drawback 4: Use of Plans
In addition to all of the above, any accurate plan is useful only if:
• The environment does not change during the execution of the plan in a
way that affects the plan
• The robot knows what state of the world and of the plan it is in at all times
• The robot’s effectors are accurate enough to execute each step of the plan
in order to make the next step possible.
Executing a plan, even when one is available, is not a trivial process.
All of the above challenges of deliberative, SPA control became obvious

to the early roboticists of the 1960s and 1970s, and they grew increasingly
dissatisfied. As we learned in Chapter 2, in the early 1980s they proposed
alternatives: reactive, hybrid, and behavior-based control, all of which are in
active use today.
What happened to purely deliberative systems?
As a result of robotics work since the 1980s, purely deliberative architec-

tures are no longer used for the majority of physical robots, because the com-
bination of real-world sensors, effectors, and time-scale challenges outlined
above renders them impractical. However, there are exceptions, because
some applications demand a great deal of advance planning and involve no

time pressure, while at the same time presenting a static environment and
low uncertainty in execution. Such domains of application are extremely
rare, but they do exist. Robot surgery is one such application; a perfect plan
is calculated for the robot to follow (for example, in drilling the patient’s
skull or hip bone), and the environment is kept perfectly static (literally by
attaching the patient’s body part being operated on to the operating table
with bolts), so that the plan remains accurate and its execution precise.
Pure deliberation is alive and well in other uses outside of robotics as well,
such as AI for game-playing (chess, Go, etc.). In general, whenever the world
is static, there is enough time to plan, and there is no state uncertainty, it is a
good thing to do.
What about robotics problems that require planning?
The SPA approach has not been abandoned in robotics; it has been ex-
panded. Given the fundamental problems with purely deliberative approach-
es, the following improvements have been made:
• Search/planning is slow, so save/cache important and/or urgent deci-

sions
• Open loop plan execution is bad, so use closed loop feedback, and be
ready to respond or replan when the plan fails.
In Chapter 15 we will see how the SPA model is currently incorporated

into modern robots in a useful way.
To Summarize
• Deliberative architectures are also called SPA architectures, for sense-plan-

act.
• They decompose control into functional modules which perform different

and independent functions (e.g., sense-world, generate-plan, translate-
plan-into-actions).
• They execute the functional modules sequentially, using the outputs of

one as inputs of the next.
• They use centralized representation and reasoning.
• They may require extensive, and therefore slow, reasoning computation.

• They encourage open loop execution of the generated plans.
Food for Thought
• Can you use deliberative control without having some internal represen-
tation?
• Can animals plan? Which ones, and what do they plan?
• If you had perfect memory, would you still need to plan?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Deliberative_Control
• In 1965, Gordon Moore, the co-founder of Intel, made the following im-
portant observation: the number of transistors per square inch on inte-
grated circuits had doubled every year since the integrated circuit was
invented. Moore predicted that this trend would continue for the foresee-
able future. In subsequent years, the pace slowed down a bit, to about
every eighteen months. So Moore’s Law states that the number of transis-
tors per square inch on integrated circuits will double about every eigh-
teen months until at least the year 2020. This law has an important impact
on the computer industry, robotics, and even global economics.
14 Don’t Think, React!
Reactive Control
Reactive control is one of the most commonly used methods for robot control.
It is based on a tight connection between the robot’s sensors and effectors.
R EACTIVE S YSTEMS Purely reactive systems do not use any internal representations of the environ-
ment, and do not look ahead at the possible outcomes of their actions: they
operate on a short time-scale and react to the current sensory information.
Reactive systems use a direct mapping between sensors and effectors, and
minimal, if any, state information. They consist of collections of rules that
couple specific situations to specific actions, as shown in figure 14.1. You can
think of reactive rules as being similar to reflexes, innate responses that do
not involve any thinking, such as jerking back your hand after touching a
hot stove. Reflexes are controlled by the neural fibers in the spinal cord, not
the brain. This is so that they can be very fast; the time it takes for a neural
signal to travel from the potentially burned finger that touched the hot stove
to the brain and back, and the computation in between to decide what to
do takes too long. To ensure a fast reaction, reflexes don’t go all the way to
the brain, but only to the spinal cord, which is much more centrally located
to most areas of the body. Reactive systems are based on exactly the same
principle: complex computation is removed entirely in favor of fast, stored
precomputed responses.
Reactive systems consist of a set of situations (stimuli, also called condi-
tions) and a set of actions (responses, also called actions or behaviors). The
situations may be based on sensory inputs or on internal state. For example,
a robot may turn to avoid an obstacle that is sensed, or because an inter-
nal clock indicated it was time to change direction and go to another area.
These examples are very simple; reactive rules can be much more complex,
involving arbitrary combinations of external inputs and internal state.
162 14 Don’t Think, React!
Figure 14.1 A diagram of a reactive architecture, showing the parallel, concurrent

task-achieving modules.
The best way to keep a reactive system simple and straightforward is to

have each unique situation (state) that can be detected by the robot’s sensors
trigger only one unique action of the robot. In such a design, the conditions
M UTUALLY E XCLUSIVE are said to be mutually exclusive, meaning they exclude one another; only one
C ONDITIONS can be true at a time.
However, it is often too difficult to split up all possible situations (world
states) in this way, and doing so may require unnecessary encoding. To en-
sure mutually exclusive conditions, the controller must encode rules for all
possible sensory input combinations. Recall from Chapter 3 that all those
combinations, when put together, define the robot’s sensor space. For a robot
with a 1-bit such as a switch, that is total of two possibilities (on and off), for
a robot with two switches, the total is four possibilities, and so on. As the
sensors grow in complexity and number, the combinatorial space of all pos-
sible sensory inputs, i.e., the sensor space, quickly becomes unwieldy. In
I NTRACTABLE computer science, AI, and in robotics, the formal term for “unwieldy” is in-
tractable. To encode and store such a large sensor space would take a giant
163
lookup table, and searching for entires in such a giant table would be slow,
unless some clever parallel lookup technique is used.
So to do a complete reactive system, the entire state space of the robot (all
possible external and internal states) should be uniquely coupled or mapped
to appropriate actions, resulting in the complete control space for the robot.
The design of the reactive system, then, is coming up with this complete set
of rules. This is done at “design-time,” not at “run-time,” when the robot is
active. This means it takes a lot of thinking by the designer (as any robot de-
sign process should), but no thinking by the robot (in contrast to deliberative
control, where there is a lot of thinking done by the system).
In general, complete mappings between the entire state space and all pos-
sible responses are not used in manually-designed reactive systems. Instead,
the designer/programmer identifies the important situations and writes the
rules for those; the rest are covered with default responses. Let’s see how
that is done.
Suppose that you are asked to write a reactive controller that will enable
a robot to move around and avoid obstacles. The robot has two simple
whiskers, one on the left and one on the right. Each whisker returns 1 bit,
“on” or “off”; “on” indicates contact with a surface (i.e., the whisker is bent).
A simple reactive controller for wall-following using those sensors would
look like this:
If left whisker bent, turn right.

If right whisker bent, turn left.
If both whiskers bent, back up and turn to the left.
Otherwise, keep going.
In the above example there are only four possible sensory inputs over all,
so the robots’ sensor space is four, and therefore there are four reactive rules.
The last rule is a default, although it covers only one possible remaining case.
A robot using the above controller could oscillate if it gets itself into a
corner where the two whiskers alternate in touching the walls. How
might you get around this rather typical problem?
There are two popular ways:
1. Use a little randomness: When turning, choose a random angle instead of

a fixed one. This introduces variety into the controller and prevents it
from getting permanently stuck in an oscillation. In general, adding a bit
Figure 14.2 Sonar ring configuration and the navigation zones for a mobile robot.
of randomness avoids any permanent stuck situation. However, it could

still take a long time to get out of a corner.
2. Keep a bit of history: Remember the direction the robot turned in the pre-
vious step (1 bit of memory) and turn in the same direction again, if the
situation occurs again soon. This keeps the robot turning in one direc-
tion and eventually gets it out of the corner, instead of getting it into an
oscillation. However, there are environments in which this may not work.
Now suppose that instead of just two whiskers, your robot has a ring of
sonars (twelve of them, to cover the 360-degree span, as you learned in Chap-
ter 9). The sonars are labeled from 1 to 12. Sonars 11, 12, 1 and 2 are at the
front of the robot, sonars 3 and 4 are on the right side of the robot, sonars 6
and 7 are in the back, and sonars 1 and 10 are on the left; sonars 5 and 8 also
get used, don’t worry. Figure 14.2) illustrates where the sonars are on the
body of the robot. That is a whole lot more sensory input than two whiskers,
which allows you to create a more intelligent robot by writing a whole lot
more reactive rules.
But just how many rules do you need to have, and which ones are the
right ones?
165
You could consider each sonar individually, but then you would have a lot
of combinations of sonars to worry about and a lot of possible sonar values
(from 0 to 32 feet each, recalling the range of Polaroid sonars from Chapter 9).
In reality, your robot does not really care what each of the sonars individu-
ally returns; it is more important to focus on any really short sonar values
(indicating that an obstacle is near) and specific areas around the robot (e.g.,
the front, the sides, etc.).
Let’s start by defining just two distance areas around the robot:
1. Danger-zone: short readings, things are too close
2. Safe-zone: reasonable readings, good for following edges of objects, but

not too far from things
Now let’s take those two zones and write a reactive controller for the robot
which considers groups of sonars instead of individual sonars:
(case
(if (minimum (sonars 11 12 1 2)) ⇐ danger-zone
and
(not stopped)
then
stop)
(if ((minimum (sonars 11 12 1 2)) ⇐ danger-zone
and
stopped)
then
move backward)
(otherwise
move forward))
The controller above consists of two reactive rules. As we have learned, it

typically takes more than a single rule to get something done.
The controller stops the robot if it is too near an object in front of it, detected
as the shortest reading of the front four sonars being in the danger-zone. If
the robot is already stopped, and the object is too near (in the danger-zone),
it backs up. The result is safe forward movement.
The controller does not check the sides and the back of the robot, however.
If we assume that the environment the robot is in is static (furniture, walls)
or that it contains rational people who keep their eyes open and do not run
into the robot from the side or back, this controller will do just fine. But if the
environment includes entities that are not very safe (children, other robots,
hurried people), then it is in fact necessary to check all the sonars, not just
the front four.
Moving forward and stopping for obstacles is not all we want from our
robot, so let’s add another controller, which we can consider another layer
or module in a reactive architecture, to make it better able to get around its
environment:
(case
(if ((sonar 11 or 12) <= safe-zone
and
(sonar 1 or 2) <= safe-zone)
then
turn left)
(if (sonar 3 or 4) <= safe-zone
then
turn right))
The above controller makes the robot turn away from detected obstacles.
Since safe-zone is larger than danger-zone, this allows the robot to turn away
gradually before getting too close to an obstacle and having to be forced to
stop, as in the previous controller. If obstacles are detected on both sides, the
robot consistently turns to the left, to avoid oscillations.
By combining the two controllers above we get a wandering behavior
which avoids obstacles at a safe distance while moving smoothly around
them, and also avoids collisions with unanticipated nearby obstacles by stop-
ping and backing up.
14.1 Action Selection
The controllers described above use specific, mutually exclusive conditions

(and defaults), so that their outputs are never in conflict, because only a sin-
gle unique situation/condition can be detected at a time. If the rules are
not triggered by mutually exclusive conditions, more than one rule can be
triggered by the same situation, resulting in two or more different action
commands being sent to the effector(s).
14.1 Action Selection 167
Figure 14.3 Two basic types of action selection: arbitration (top) and fusion (bot-
tom).
A CTION S ELECTION Action selection is the process of deciding among multiple possible actions
or behaviors. It may select only one output action or may combine the actions
to produce a result. These two approaches are called arbitration and fusion.
C OMMAND Command arbitration is the process of selecting one action or behavior from
A RBITRATION multiple candidates.
C OMMAND F USION Command fusion is the process of combining multiple candidate actions or
behaviors into a single output action/behavior for the robot.
The two alternatives are shown in figure 14.3 as part of the generic reactive
system.
It turns out that action selection is a major problem in robotics, beyond
reactive systems, and that there is a great deal of work (theory and practice)
on different methods for command arbitration and fusion. We will learn a
great deal more about this interesting problem in Chapter 17.
Although a reactive system may use arbitration to decide which action to
execute, and thus execute only one action at a time, it still needs to monitor
its rules in parallel, concurrently, in order to be ready to respond to any one

or more of them that might be triggered.
Using the example of the wall-following robot, both whiskers have to be
monitored all the time, or all sonars have to be monitored all the time, since
an obstacle could appear anywhere, at any time. In the wall-following ex-
ample, there are very few rules (at least in the whisker case), so given a fast
processor, those could be checked sequentially and no real response time
would be lost. But some rules may take time to execute. Consider the fol-
lowing controller:
If there is no obstacle in front, move forward.

If there is an obstacle in front, stop and turn away.
Start a counter. After 30 seconds, choose randomly
between left and right and turn by 30 degrees.
What kind of behavior does this controller produce? It produces random

wandering with obstacle avoidance (and not very good avoidance at that).
Note that the conditions of the three rules above require different amounts
of time to compute. The first two check sensory input, while the third uses a
timer. If the controller executes the rules sequentially, it would have to wait
for 30 seconds in the third rule, before being able to check the first rule again.
In that time, the robot could run into an obstacle.
Reactive systems must be able to support parallelism, the ability to moni-
tor and execute multiple rules at once. Practically, this means that the under-
M ULTITASKING lying programming language must have the ability to multitask, to execute
several processes/rules/commands in parallel. The ability to multitask is
critical in reactive systems: if a system cannot monitor its sensors in parallel,
and instead checks them in sequence, it may miss an event, or at least the
onset of an event, and thus fail to react in time.
You can begin to see that designing a reactive system for a robot can be
quite complicated, since multiple rules, potentially a large number of them,
have to be put together in a way that produces effective, reliable, and goal-
driven behavior.
How do we go about organizing a reactive controller in a principled
way? By using architectures that have been specifically designed for
that purpose.
The best known architecture for reactive control is Subsumption Architec-
ture, introduced by Prof. Rodney Brooks at MIT in 1985. It’s an oldie by now,
but it is still a goodie, finding its way into a vast number of reactive, hybrid,
and behavior-based systems.
14.2 Subsumption Architecture
The basic idea behind Subsumption Architecture is to build systems incre-

mentally, from the simple parts to the more complex, all the while using
the already existing components as much as possible in the new stuff being
added. Here is how it works.
Subsumption systems consist of a collection of modules or layers, each
of which achieves a task. For example, they might be move-around, avoid-
obstacles, find-doors, visit-rooms, pick-up-soda-cans, and so on. All of the
task-achieving layers work at the same time, instead of in sequence. This
means the rules for each of them are ready to be executed at any time, when-
ever the right situation presents itself. As you recall, this is the premise of
reactive systems.
The modules or layers are designed and added to the robot incrementally.
If we number the layers from 0 up, we first design, implement, and debug
layer 0. Let’s suppose that layer 0 is move-around, which keeps the robot
going. Next we add layer 1, avoid-obstacles, which stops and turns or backs
away whenever an obstacle is detected. This layer can already take advan-
tage of the existing layer 0, which moves the robot around, so that, together,
layers 0 and 1 result in the robot moving around without running into things
it can detect. Next we add layer 2, say find-doors, which looks for doors
while the robot roams around safely. And so on, until all of the desired tasks
can be achieved by the robot through the combination of its layers.
But there is a twist. Higher layers can also temporarily disable one or more
of those below them. For example, avoid-obstacles can stop the robot from
moving around. What would result is a robot that sits still, but can turn and
back away if somebody approaches it. In general, disabling obstacle avoid-
ance is a dangerous thing to do, and so it is almost never done in any real
robot system, but having higher layers selectively disable lower layers within
a reactive system is one of the principles of Subsumption Architecture. This
manipulation is done in one of only two ways, as shown in figure 14.4:
1. The inputs of a layer/module may be suppressed; this way the module

receives no sensory inputs, and so computes no reactions and sends no
outputs to effectors or other modules.
Figure 14.4 Subsumption methods for layer/module interaction: suppression of in-

puts (left) and inhibition of outputs (right).
2. The outputs of a layer/module may be inhibited; this way the module

receives sensory inputs, performs its computation, but cannot control any
effectors or other modules.
S UBSUMPTION The name "Subsumption Architecture" comes from the idea that higher lay-
A RCHITECTURE ers can assume the existence of the lower ones and the goals they are achiev-
ing, so that the higher layers can use the lower ones to help them in achieving
their own goals, either by using them while they are running or by inhibiting
them selectively. In this way, higher layers "subsume" lower ones.
There are several benefits to the subsumption style of organizing reactive
systems. First, by designing and debugging the system incrementally, we
avoid getting bogged down in the complexity of the overall task of the robot.
Second, if any higher-level layers or modules of a subsumption robot fail, the
lower-level ones will still continue to function unaffected.
B OTTOM -U P The design of subsumption controllers is called bottom-up, because it pro-
gresses from the simpler to the more complex, as layers are added incre-
mentally. This is good engineering practice, but its original inspiration came
to the inventor of Subsumption Architecture from biology. Brooks was in-
spired by the evolutionary process, which introduces new abilities based on
the existing ones. Genes operate using the process of mixing (crossover) and
Figure 14.5 A subsumption-based robot control system.
changing (mutation) of the existing genetic code, so complete creatures are

not thrown out and new ones created from scratch; instead, the good stuff
that works is saved and used as a foundation for adding more good stuff,
and so complexity grows over time. Figure 14.5 shows an example sub-
sumption control system; it would be constructed bottom-up, adding each
layer incrementally.
Building incrementally helps the design and debugging process, and us-
ing layers is useful for modularizing the robot controller. This is also good
engineering practice. Otherwise, if everything is lumped together, it is hard
to design, hard to debug, and hard to change and improve later.
The effectiveness of modularity also depends on not having all modules
connected to all others, as that defeats the purpose of dividing them up. So
in Subsumption Architecture, the goal is to have very few connections be-
tween different layers. The only intended connections are those used for
inhibition and suppression. Inside the layers, of course, there are plenty of
connections, as multiple rules are put together to produce a task-achieving
behavior. Clearly, it takes more than one rule to get the robot to avoid obsta-
cles, to find doors, and so on. But by keeping the rules for the separate tasks
apart, the system becomes more manageable to design and maintain.
Therefore, in Subsumption Architecture, we use strongly coupled connec-
tions within layers, and loosely coupled connections between layers.
How do we decide what makes a subsumption layer, and what should
go higher or lower?
Unfortunately, there are no fixed answers to those hard design questions.
It all depends on the specifics of the robot, the environment, and the task.
There is no strict recipe, but some solutions are better than others, and most
of the expertise that robot designers have, is gained through trial and error,
basically through sweating it.
14.3 Herbert, or How to Sequence Behaviors Through the World
How would you make a reactive robot execute a sequence of behaviors?

For example, consider the following task: Search for soda cans, find any
empty ones, pick them up, and bring them back to the starting point.
If you are thinking that you can have one layer or module activate another
in a sequence, you are right, but that is not the way to go. Coupling between
reactive rules or subsumption layers need not be through the system itself,
through explicit communication, but instead through the environment.
A well-known subsumption robot called Herbert, designed by Jonathan
Connell (a PhD student of Brooks’) in the late 1980s, used this idea to ac-
complish the task of collecting soda cans. Here is how. Herbert had a layer
that moved it around without running into obstacles. (As you will find out,
every mobile robot has such a layer/ability/behavior, and it is usually reac-
tive, for the obvious reason that collision-free navigation requires being able
to respond quickly.) Herbert’s navigation layer used infra red sensors. It also
had a layer that used a laser striper and a camera to detect soda cans (by
matching them to a particular width in a visual image, quite a clever trick).
Herbert also had an arm and a layer of control that could extend the arm,
sense if there was a can in its gripper, close the gripper, and pull the arm in.
Cleverly, all these arm actions were separate, and were activated not in a se-
quential, internal way but instead by sensing the environment and the robot
directly, like this:
14.3 Herbert, or How to Sequence Behaviors Through the World 173
If you see something that looks like a soda can

approach it
If it still looks like a soda can up close
then extend the arm
else turn and go away
Whenever the arm is extended

check between the gripper fingers
Whenever the gripper sensors (IR break beam) detects

something
close the gripper
Whenever the gripper is closed and the arm extended

pull back the arm
Whenever the arm is pulled back and the gripper closed

go to drop off the can
Herbert had a very clever controller; see figure 14.6 for its diagram. It
never used any "explicit" sequencing, but instead just checked the different
situations in the world (the environment and the robot together) and reacted
appropriately. Note, for example, that it was safe for Herbert to close its
gripper whenever there was something sensed in it, because the only way it
could sense something in the gripper was after it had already looked for cans
and extended the arm to reach one. Similarly, it was safe for Herbert to pull
the arm back when the gripper was closed, since it was closed only if it was
holding a can. And so on.
By the way, in case you are wondering how it knew if the cans were empty
or full, it used a simple strain gauge to measure the weight of the can it was
holding. It took only the light cans.
Herbert had no internal communication between the layers and rules that
achieved can-finding, grabbing, arm-retracting, and moving around, yet it
executed everything in the right sequence every time, and was able to per-
form its can-collecting task. The big idea here is that Herbert was using the
sensory inputs as a means of having its rules and layers interact. Brooks
called this "interaction through the world."
Another motto, also by Brooks, recommends that we "use the world as
its own best model." This is a key principle of Subsumption Architecture,
Figure 14.6 Herbert’s control architecture.
and of reactive systems in general: If the world can provide the information
directly (through sensing), it is best for the robot to get it that way rather than
to store it internally in a representation, which may be large, slow, expensive,
and outdated.
To summarize, the guiding principles of Subsumption Architecture are:
• Systems are built from the bottom up.
• Components are task-achieving actions/behaviors (not functional mod-

ules).
• Components can be executed in parallel (multitasking).

14.3 Herbert, or How to Sequence Behaviors Through the World 175
• Components are organized in layers.

• Lowest layers handle the most basic tasks.
• Newly added components and layers exploit the existing ones. Each com-
ponent provides and does not disrupt a tight coupling between sensing
and action.
• There is no use of internal models; "the world is its own best model."
Subsumption Architecture is not the only method for structuring reactive
systems, but it is a very popular one, due to its simplicity and robustness. It
has been widely used in various robots that successfully interact with uncer-
tain, dynamically changing environments.
How many rules does it take to put together a reactive system?
By now, you know the answer is "That depends on the task, the environ-
ment, and the sensors on the robot." But it is important to remember that for
any robot, task, and environment that can be specified in advance, a complete
reactive system can be defined that will achieve the robot’s goals. However,
that system may be prohibitively large, as it may have a huge number of
rules. For example, it is in theory possible to write a reactive system for play-
ing chess by encoding it as a giant lookup table that gives the optimal move
for each possible board position. Of course the total number of all possible
board positions, and all possible games from all those positions, is too huge
for a human to remember or even a machine to compute and store. But for
smaller problems, such as tic-tac-toe and backgammon, this approach works
very well. For chess, another approach is needed: using deliberative control,
as we learned in Chapter 13.
To Summarize
• Reactive control uses tight couplings between perception (sensing) and
action to produce timely robotic response in dynamic and unstructured
worlds (think of it as "stimulus-response").
• Subsumption Architecture is the best-known reactive architecture, but
certainly not the only one.
• Reactive control uses a task-oriented decomposition of the controller. The
control system consists of parallel (concurrently executed) modules that
achieve specific tasks (avoid-obstacle, follow-wall, etc.).
• Reactive control is part of other types of control, specifically hybrid con-

trol and behavior-based control.
• Reactive control is a powerful method; many animals are largely reactive.
• Reactive control has limitations:
– Minimal (if any) state

– No memory
– No learning
– No internal models / representations of the world.
Food for Thought
• Can you change the goal of a reactive system? If so, how? If not, why not?
You will soon learn how other methods for control deal with this problem
and what the trade-offs are.
• Can you always avoid using any representation/world model? If so,

how? If not, why not, and what could you do instead?
• Can a reactive robot learn a maze?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Reactive_Control
• Genghis, the six-legged robot we mentioned in Chapter 5 was also pro-

grammed with Subsumption Architecture. You can learn about how its
control system was put together, how many rules it took, and how it grad-
ually got better at walking over rough terrain as more subsumption layers
were added, from Cambrian Intelligence by Rodney Brooks.
Think and Act Separately, in
15 Parallel
Hybrid Control
As we have seen, reactive control is fast but inflexible, while deliberative con-
trol is smart but slow. The basic idea behind hybrid control is to get the best
of both worlds: the speed of reactive control and the brains of deliberative
control. Obvious, but not easy to do.
H YBRID C ONTROL Hybrid control involves the combination of reactive and deliberative control
within a single robot control system. This combination means that funda-
mentally different controllers, time-scales (short for reactive, long for delib-
erative), and representations (none for reactive, explicit and elaborate world
models for deliberative) must be made to work together effectively. And
that, as we will see, is a tall order.
In order to achieve the best of both worlds, a hybrid system typically con-
sists of three components, which we can call layers or modules (though they
are not the same as, and should not be confused with, layers/modules used
in reactive systems):
• A reactive layer
• A planner
• A layer that links the above two together.
As a result, hybrid architectures are often called three-layer architectures and
hybrid systems, three-layer systems. Figure 15.1 is a diagram of a hybrid ar-
chitecture and its layers.
We already know about deliberation with planners and about action with
reactive systems. What we now need to learn about is the real challenge and
"value added" of hybrid systems: the “magic middle.” The middle layer has
a hard job, because it has to:
178 15 Think and Act Separately, in Parallel
Figure 15.1 A diagram of a hybrid architecture, showing the three layers.
• Compensate for the limitations of both the planner and the reactive sys-
tem
• Reconcile their different time-scales
• Deal with their different representations
• Reconcile any contradictory commands they may send to the robot.
So the main challenge of hybrid control is achieving the right compromise

between the deliberative and reactive parts of the system.
Let’s work through an example to see what that means in practice. Sup-
pose we have a robot whose job is to deliver mail and paperwork through
the day to various offices within an office building. This was a very popular
mobile robot application that researchers liked to think about in the 1970s,
1980s, and 1990s, until email became the preferred means of communication.
However, office buildings still have a great many tasks that require "gopher-
ing" or running around with deliveries. Also, we can use the same controller
to deliver medicines to patients in a hospital, so let’s consider that example.
In order to get around a busy office/hospital environment, the robot needs
to be able to respond to unexpected obstacles, and fast-moving people and
15.1 Dealing with Changes in the World/Map/Task 179
objects (gurneys, stretchers, rolling lunch carts, etc.). For this it needs a ro-
bust reactive controller.
In order to efficiently find specific offices/rooms for making deliveries, the
robot needs to use a map and plan short paths to its destinations. For this it
needs an internal model and a planner of some kind.
So there you have it, a perfect candidate for a hybrid system. We already
know how to do both of the component systems (reactive collision-free nav-
igation and deliberative path planning), so how hard can it be to put the two
together?
Here is how hard:
• What happens if the robot needs to deliver a medication to a patient as

soon as possible, yet it does not have a plan for a short path to the patient’s
room? Should it wait for the plan to be computed, or should it move down
the corridor (in which direction?) while it is still planning in its head?
• What happens if the robot is headed down the shortest path but suddenly
a crew of doctors with a patient on a stretcher starts heading its way?
Should it just stop and get out of the way in any direction, and wait as
long as it may take, or should it start replanning an alternative path?
• What happens if the plan the robot computed from the map is blocked,
because the map is out of date?
• What happens if the patient was moved to another room without the robot
knowing?
• What happens if the robot keeps having to go to the same room, and so
has to replan that path or parts of it all the time?
• What if, what if, what if.
The methods for handling the above situations and various others are usu-
ally implemented in the “magic middle” layer of a hybrid system. Designing
the middle layer and its interactions with the other two layers (as shown in
figure 15.2) are the main challenges of hybrid control. Here are some com-
monly used methods and approaches.
15.1 Dealing with Changes in the World/Map/Task

When the reactive system discovers that it cannot do its job (for example, it
can’t proceed because of an obstacle, a closed door, or some other snag), it can
Figure 15.2 Some ways of managing layer interaction in hybrid controllers for
robots.
inform the deliberative layer about this new development. The deliberative
layer can use this information to update its representation of the world, so
that it can now, and in the future, generate more accurate and useful plans.
This is a good idea not only because it is necessary to update the internal
model when things change, but also because we already know that updat-
ing internal models and generating plans takes time and computation, so it
cannot be afforded continually. The input from the reactive layer provides
an indication of a very good time for such an update.
15.2 Planning and Replanning
Whenever the reactive layer discovers that it cannot proceed, this can be used
as a signal to the deliberative layer to do some thinking, in order to generate
a new plan. This is called dynamic replanning.
But not all information flows from the bottom up, from the reactive to the
deliberative layer. In fact, the deliberative layer, which provides the path
15.3 Avoiding Replanning 181
to the goal (i.e., the room to go to), gives the robot the directions to follow,
turns to take, distances to go. If the planner is computing while the robot
is moving, it may send a message to the reactive navigation layer to stop,
turn around, and head in a different direction because a better way has been
discovered in the map.
In general, a complete plan, when finished, has the best answer the deliber-
ator can generate. But sometimes there is not enough time to wait for that full
and optimal answer (such as when the patient really needs immediate help).
In those cases, it is often best to get the robot going in the generally right
direction (for example, toward the correct wing of the hospital) and in the
meantime keep generating a more accurate and detailed plan, and updating
the navigation layer as needed.
Of course the timing of these two processes, the reactive navigation (run-
ning around) and deliberative planning (thinking hard), is obviously not
synched up, so the robot may get to the right wing of the hospital and not
know where to go next, and thus may have to stop and wait for the planner
to get done with its next step. Alternatively, the planner may need to wait for
the robot to move around or get out of a crowded area in order to figure out
exactly where it is, so that it can generate a useful plan. We will learn more
about the various challenges of navigation in Chapter 19.
15.3 Avoiding Replanning
A useful idea researchers in planning had quite a while ago was to remem-
ber/save/store plans so that they would not have to be generated again in
the future. Of course, every plan is specific to the particular initial state and
goal state, but if those are at all likely to recur, it is worth stashing away the
plan for future use.
This idea is really popular for situations that happen often and need a
fast decision. In our example above, such situations include dealing with
crowded junctions of corridors or dynamic obstacles that interfere with a par-
ticular direction of movement. Instead of "thinking out" how to deal with this
every time, controller designers often preprogram many of these responses
ahead of time. But because these are not quite reactive rules, because they
involve multiple steps, and yet they are not quite plans, because they do not
involve very many steps, they find a perfect home in the middle layer of the
three-layer system.
This basic idea of storing and reusing mini plans for repeated situations
has been used in "contingency tables," literally lookup tables that tell the
robot what to do (as in a reactive system) and pull out a little plan as a re-
sponse (as in a deliberative system). This same idea has appeared in the form
of fancy terms such as "intermediate-level actions" and "macro operators." In
all cases, it is basically the same thing: plans that are computed (either off-
line or during the robot’s life) and stored for fast lookup in the future.
So who has ultimate priority, the deliberative layer or the reactive layer?
The answer is, as usual, "It depends." What it depends on includes the type
of environment, task, sensing, timing, and reaction requirements, and so on.
Some hybrid systems use a hierarchical structure, so that one of the systems
is always in charge. In some cases, the deliberator’s plan is "the law" for
the system, while in others the reactive system merely considers it as advice
that can be ignored. In the most effective systems, the interaction between
thinking and acting is coupled, so that each can inform and interrupt the
other. But in order to know who should be in charge when, it is necessary
to consider the different modes of the system and specify who gets its say.
For example, the planner can interrupt the reactive layer if it (the planner)
has a better path than the current being executed, but it should do so only if
the new plan is sufficiently better to make it worth the bother. On the other
hand, the reactive layer can interrupt the planner if it (the reactive layer)
finds a blocked path and cannot proceed, but should only do so after it has
tried to get around the barrier for a while.
15.4 On-Line and Off-Line Planning
In everything we’ve talked about so far, we have assumed that the deliber-
ation and the reaction are happening as the robot is moving around. But as
we saw from the role of the middle layer, it is useful to store plans once they
are generated. So the next good idea is to preplan for all the situations that
O FF -L INE P LANNING might come up, and store those plans ahead of time. This off-line planning
takes place while the robot is being developed and does not have much to
O N -L INE P LANNING worry about, as compared with on-line planning of the kind that a busy robot
has to worry about while it is trying to get its job done and its goals achieved.
If we can preplan, why not generate all possible plans ahead of time,
store them all, and just look them up, without ever having to do search
and deliberation at run-time, while the robot is trying to be fast and
efficient?
This is the idea behind universal plans.
U NIVERSAL P LAN A universal plan is a set of all possible plans for all initial states and all goals
within the state space of a particular system.
If for each situation a robot has a preexisting optimal plan it only has to
look up, then it can always react optimally, and so have both reactive and
deliberative capabilities without deliberating at all. Such a robot is reactive,
since the planning is all done off-line and not at run-time.
Another good feature of such precompiled plans is that information can
be put into the system in a clean, principled way. Such information about
D OMAIN K NOWLEDGE the robot, the task, and the environment is called domain knowledge. It is
compiled into a reactive controller, so the information does not have to be
reasoned about (or planned with) on line, in real time but instead becomes a
set of real-time reactive rules that can be looked up.
This idea was so popular that researchers even developed a way of gener-
ating such precompiled plans automatically, by using a special programming
language and a compiler. Robot programs written in that language produced
a kind of a universal plan. To do so, the program took as input a mathemati-
cal specification of the world and of the robot’s goals, and produced a control
"circuit" (a diagram of what is connected to what) for a reactive "machine."
These machines were called situated automata; they were not real, physical
machines but formal (idealized) ones whose inputs were connected to ab-
stract (again idealized, not real) sensors, and whose outputs were connected
S ITUATED A UTOMATA to abstract effectors. To be situated means to exist in a complex world and to
interact with it; automata are computing machines with particular mathemat-
ical properties.
Unfortunately, this is too good to be true for real-world robots. Here is
why:
• The state space is too large for most realistic problems, so either generat-
ing or storing a universal plan is simply not possible.
• The world must not change; if it does, new plans need to be generated for
the changed environment.
• The goals must not change; this is the same as with reactive systems,
where if the goals change, at least some of the rules need to change as
well.
Situated automata could not be situated in the real, physical world after
all. Robots exist in the real world but we can’t easily write mathematical
specifications that fully describe this world, which is what that the situated
automata approach required.
And so we are back to having to do deliberation and reaction in real time,
and hybrid systems are a good way to do that. They do have their own
drawbacks, of course, including:
• The middle layer is hard to design and implement, and it tends to be very
special-purpose, crafted for the specific robot and task, so it has to be rein-
vented for almost every new robot and task.
• The best of both worlds can end up being the worst of both worlds; if mis-
managed, a hybrid system can degenerate into having the planner slow
down the reactive system, and the reactive system ignore the planner en-
tirely, minimizing the effectiveness of both.
• An effective hybrid system is not easy to design or debug, but that is true
for any robot system.
In spite of the drawbacks of hybrid systems, they are the most popular
choice for a great many problems in robotics, especially those involving a
single robot that has to perform one or more tasks that involve thinking of
some type (such as delivery, mapping, and many others) as well as reacting
to a dynamic environment.
To Summarize
• Hybrid control aims to bring together the best aspects of reactive control
and deliberative control, allowing the robot to both plan and react.
• Hybrid control involves real-time reactive control in one part of the sys-
tem (usually the low level) and more time-expensive deliberative control
in another part of the system (usually the high level), with an intermediate
level in between.
• Hybrid architectures are also called three-layer architectures because of

the three distinct layers or components of the control system.
• The main challenge of hybrid systems lies in bringing together the re-
active and deliberative components in a way that results in consistent,
timely, and robust behavior over all.
• Hybrid systems, unlike reactive systems, are capable of storing represen-

tation, planning, and learning.
Food for Thought
Is there an alternative to hybrid systems, or is that all there is in terms of

intelligent, real-time robot behavior? Can you think of another way that a
robot can be able to both think and react? After you come up with an an-
swer, check out the next chapter.
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Hybrid_Control
• Hybrid control is a major area of research in control theory, a field of

study usually pursued in electrical engineering, which we described in
Chapter 2. (Recall that control theory covers feedback control, which we
studied in Chapter 10.) In control theory, the problem of hybrid control
deals with the general interplay between continuous and discrete signals
and representations, and the control of systems that contain them. Here
are a couple of popular textbooks on control theory:
– Signals and Systems by Alan V. Oppenheim, Alan S. Willsky, and with

Nawab S. Hamid.
– Modern Control Systems, by Richard C. Dorf and Robert H. Bishop.
16 Think the Way You Act
Behavior-Based Control
As we have learned, reactive control and deliberative control each have their
limitations, and hybrid control, an attempt to combine their best compo-
nents, has its own challenges. In this chapter we will learn about behavior-
based control, another popular way of controlling robots, which also incorpo-
rates the best of reactive systems but does not involve a hybrid solution.
Behavior-based control (BBC) grew out of reactive control, and was sim-
ilarly inspired by biological systems. Actually, if you think about it, all ap-
proaches to control (reactive, deliberative, hybrid, and behavior-based) were
inspired by biology, at one level or another. This just goes to show that bio-
logical systems are so complex that they can serve as inspiration for a variety
of disparate methods for control, yet still remain more complicated and ef-
fective than anything artificial that has been made so far. Just when you start
to feel really good about artificial systems, all you need to do is to go outside
and look at some bugs to see how far we have yet to go.
Back to behavior-based systems. The primary inspiration came from sev-
eral main challenges:
• Reactive systems are too inflexible, incapable of representation, adapta-
tion, or learning.
• Deliberative systems are too slow and cumbersome.
• Hybrid systems require complex means of interaction among the compo-
nents.
• Biology seems to have evolved complexity from simple and consistent
components.
If you think about the different control methodologies along a line, BBC is
closer to reactive control than to hybrid control, and farthest from delibera-
188 16 Think the Way You Act
Figure 16.1 A diagram of a behavior-based architecture, showing behavior organi-

zation.
tive control. In fact, as you will see, behavior-based systems have reactive
components, just as hybrid systems do, but they do not have traditional
deliberative components at all. Figure 16.1 shows a diagram of a generic
behavior-based control architecture.
So what is behavior-based control?
B EHAVIOR -B ASED Behavior-based control (BBC) involves the use of "behaviors" as modules for
C ONTROL control. Thus, BBC controllers are implemented as collections of behaviors.
The first property of behavior-based control to remember is that it’s all about
behaviors. That brings up the obvious question:
What is a behavior?
There is a definite answer to that question, and in fact one of the strengths
of BBC comes from different ways in which people have encoded and imple-
mented behaviors, which are also sometimes called behavior-achieving mod-
ules. But don’t assume that anything goes, and any piece of code can be
a behavior. Fortunately, there are some rules of thumb about behaviors and
constraints on how to design them and what to avoid in implementing them:
189
• Behaviors achieve and/or maintain particular goals. A homing behavior

achieves the goal of getting the robot to the home location. A wall-following
behavior maintains the goal of following a wall.
• Behaviors are time-extended, not instantaneous. That means they take

some time to achieve and/or maintain their goals. After all, it takes a
while to go home or follow a wall.
• Behaviors can take inputs from sensors and also from other behaviors,
and can send outputs to effectors and to other behaviors. This means we
can create networks of behaviors that "talk to" each other.
• Behaviors are more complex than actions. While a reactive system may
use simple actions like stop and turn-right, a BBC uses time-extended be-
haviors like the ones we saw above, as well as others like find-object, follow-
target, get-recharged, hide-from-the-light, aggregate-with-your-team, find-mate,
etc.
As you can see from the short list above, behaviors can be designed at a
L EVEL OF variety of levels of detail or description. This is called their level of abstraction,
A BSTRACTION because to abstract is to take details away and make things less specific. Be-
haviors can take different amounts of time and computation. In short, they
are quite flexible, and that is one of the key advantages of BBC.
The power and flexibility of BBC comes not only from behaviors but also
from the organization of those behaviors, from the way they are put together
into a control system. Here are some principles for good BBC design:
• Behaviors are typically executed in parallel/concurrently, much as in re-

active systems, in order to enable the controller to respond immediately
when needed.
• Networks of behaviors are used to store state and to construct world mod-
els/representations. When assembled into distributed representations,
behaviors can be used to store history and to look ahead into the future.
• Behaviors are designed so that they operate on compatible time-scales.

This means it is not good BBC design to have some very fast behaviors and
some very slow ones. Why not? Because that makes the system hybrid
in terms of the time-scale, and we’ve already seen (in Chapter 15) that
interfacing different time-scales is a challenging problem.
Because of the above properties of behaviors and their combinations, BBC

is not limited in the ways that reactive systems are. At the same time, it
does not employ a hybrid structure, as hybrid systems do. Instead, behavior-
based systems have the following key properties:
1. The ability to react in real-time
2. The ability to use representations to generate efficient (not only reactive)
behavior
3. The ability to use a uniform structure and representation throughout the
system (with no intermediate layer(s)).
Before we move on to the other advantages of BBC, let’s clear something
up first. You might be wondering about the difference between behaviors
inside the robot and inside a behavior-based controller, which cannot be ob-
served from the outside, and behaviors that the robot performs and that can
be observed from the outside. If you are not wondering about this, you
should be. Here is why.
In some behavior-based systems, the internal behavior structure exactly
matches the externally manifested behaviors. This means that if the robot
performs wall-following, then there is a distinct piece of the control program
we can call “wall-following” that achieves that observable behavior. How-
ever, this is not always the case, especially for more complex behaviors, so
most controllers are not designed this way.
Why not? After all, having an intuitive match between the internal control
program and the externally observable behavior sounds really clean and easy
to understand. Unfortunately, it turns out to be impractical.
Most interesting observable robot behaviors are a result not just of the con-
trol program, but also of the interaction of the internal behaviors among
themselves and with the environment the robot is in. This is basically the
same idea that reactive systems used (recall Chapter 14): simple reactive
rules can interact to produce interesting observable behavior of the robot.
The same is true for internal control behaviors.
Consider flocking, the behavior in which a group of robots moves together
in a group. A robot that flocks with others does not necessarily need to have
an internal flocking behavior. In fact, flocking can be implemented very ele-
gantly in a completely distributed way. To learn more about this, see Chap-
ter 18.
This is one of the clever philosophies of behavior-based systems. Such
systems are typically designed so the effects of the behaviors interact in the
191
environment rather than internally through the system, in order to take ad-
I NTERACTION vantage of interaction dynamics. In this context, those dynamics refer to pat-
D YNAMICS terns and history of interaction and change. The idea that rules or behaviors
can interact to produce more complex outputs is called emergent behavior; we
will devote entire Chapter 18 to it because it is quite confusing, yet interesting
and important.
In general, in designing a behavior-based system, the designer begins by
listing the desirable observable, externally manifested behaviors. Next, the
designer figures out how best to program those behaviors with internal be-
haviors. Those internal behaviors, as we saw, may or may not directly achieve
the observable goals. To make this process easier, various compilers and pro-
gramming languages have been developed.
Now we know that behavior-based systems are structured as collections
of internal behaviors, which, when moving in an environment, produce a
collection of externally manifested behaviors, and the two are not necessarily
the same. And that’s all right, as long as the robot’s observable behaviors
achieve its overall goals.
Let’s consider the following example problem: We need the robot to
move around a building and water plants whenever they get dry.
You can imagine that the robot needs the following behaviors: avoid-
collisions, find-plant, check-if-dry, water, refill-water-reservoir, recharge-bat-
teries. To program such a robot, would you program those exact six behav-
iors and be done? Possibly, but chances are that some of them, especially the
more complicated ones (like find-plant, for example), may themselves con-
sist of internal behaviors themselves, such as wander-around, detect-green,
approach-green, and so on. But don’t assume that this means behavior-
based systems are simply hierarchical, with each behavior having its own
component behaviors. Some systems are hierarchical, but usually multiple
behaviors use and share the same underlying component behaviors. In our
example, the refill-water-reservoir behavior may also use the wander-around
behavior in order to get back to the water supply and refill its reservoir.
In general, behavior-based controllers are networks of internal behaviors
which interact (send messages to each other) in order to produce the
desired external, observable, manifested robot behavior.
As we outlined earlier, the key differences among the various control meth-
ods are based on the way each treats modularity, time, and representation.
In the case of BBC, the approach to modularity is that of using a collection of

behaviors, and having those behaviors be relatively similar in terms of execu-
tion time. This means that having one behavior which contains a centralized
world model and performs reasoning on it, as in deliberative or hybrid sys-
tems, would not fit the behavior-based philosophy, and would thus not make
a good controller.
Since BBC is based on the reactive philosophy (but not limited by it), it also
mandates that behaviors be incrementally added to the system, and that they
be executed concurrently, in parallel, rather than sequentially, one at a time.
Behaviors are activated in response to external and/or internal conditions,
sensory inputs, and internal state or messages from other behaviors. Dy-
namics of interaction arise both within the system itself (from the interaction
among the behaviors) and within the environment (from the interaction of
the behaviors with the external world). This is similar to the principles and
effects of reactive systems, but can be exploited in richer and more interesting
ways because:
Behaviors are more expressive (more can be done with them) than simple
reactive rules are.
Remember that behavior-based systems can have reactive components. In
fact, many of these systems do not use complex representations at all. Inter-
nal representation is not necessary for all problems, and so it can be avoided
for some. This does not mean that the resulting controller is reactive; as long
as the controller is structured using behaviors, in the ways described above,
it creates a BBC system, with all the benefits that come with the approach,
even without the use of representation.
Because behaviors are more complex and more flexible than reactive rules,
they can be used in clever ways to program robots. By having behaviors
interact with each other within the robot, they can be used to store represen-
tation. Therefore they can serve as a basis for learning and prediction, and
this means that behavior-based systems can achieve the same things that hy-
brid systems can, but in a different way. One of the major differences is in
how representation is used.
16.1 Distributed Representation

The philosophy of behavior-based systems mandates that information used
as internal representation not be centralized or centrally manipulated. This
is in direct contrast to deliberative and hybrid systems, which typically use a
centralized representation (such as a global map) and a centralized reasoning

mechanism (such as a planner that uses the global map to find a path). What
is an alternative to that approach, and why is it a good thing? That’s what
we are going to learn about next.
The key challenge in using representation in BBC is in how that repre-
sentation (any form of world model) can be effectively distributed over the
behavior structure. In order to avoid the pitfalls of deliberative control, the
representation must be able to act on a time-scale that is close to (if not the
same as) the real-time components of the system. Similarly, to avoid the chal-
lenges of the middle layer in hybrid control, the representation needs to use
the same underlying behavior structure as the rest of the system.
16.2 An Example: Distributed Mapping
Let’s return to our plant-watering robot. Suppose that your task is to im-
prove that robot so it does not just move around randomly, look for plants,
and water them, but also to make a map of its environment as it goes along,
and then to use that map to find shortest paths to particular plants, or other
locations in the environment.
Navigation and mapping are the most common mobile robot tasks, as we
will learn in Chapter 19. So let’s see how we can build (or learn) a map in a
BBC. We already know that it shouldn’t be a traditional centralized map that
a deliberative or hybrid system would use.
Instead, we have to somehow distribute information about the environ-
ment over behaviors in a useful way. How can we do that?
Here is how: we can distribute parts of a map over different behaviors.
Then we can connect parts of the map that are adjacent in the environment
(such as a wall that is next to a corridor) by connecting the behaviors that
represent them in the robot’s head/controller. This way we will build a net-
work of behaviors which represents a network of locations in the environ-
ment. What is connected in the map is connected in the environment, and
vice versa. A fancier formal way to say this is that the topology of the map is
isomorphic (has the same form) with the topology of the specific landmarks
in the environment.
Figure 16.2 Toto the robot. (Photo courtesy of the author)
16.2.1 Toto the Robot

There once was a robot named Toto (see Toto’s photo in figure 16.2) that used
just such a representation to learn the map of its office environment at MIT, in
order to hang out with students and help someone1 get her PhD in robotics.
Let’s see how Toto learned its environment and got around to do its task. To
be honest, Toto did not in fact water any plants, as water and mobile robots
rarely mix well.
16.2.2 Toto’s Navigation

Toto was a behavior-based system; you can see its control diagram in fig-
ure 16.3. It was equipped with a simple ring of twelve Polaroid sonars (if you
can’t remember why twelve, go back to Chapter 14), and a low-resolution
compass (2 bits). Toto lived around 1990, in the "old days," well before laser
scanners became commonplace. As in any good BBC, Toto’s controller con-
sisted of a collection of behaviors. Its lowest levels of the system took care
of moving Toto around safely, without collisions. Next, Toto had a behav-
ior that kept it near walls and other objects, based on the distance perceived
1. Ok, I’ll tell you who: it was this book’s author.

Figure 16.3 Toto’s behavior-based controller.
Figure 16.4 Toto’s sonar-based navigation regions. (Image courtesy of the author)
with its sonar sensors. Just as we learned in Chapter 14, Toto used a set of
sonar sensing regions for navigation (shown in figure 16.4); its low-level safe
navigation was in fact a reactive system. In fact, Toto’s navigation controller
used some of the programs you saw in Chapter 14.
Following boundaries was a good idea for Toto. Remember why? If not,
look back at Chapter 9 and recall that sonars are most likely to be accurate at
short and intermediate distances from objects. It pays to get accurate sensor
data when building maps and going places, so Toto did it by staying near
objects whenever possible. It turns out that rats and mice also like to stick
near objects, perhaps for other reasons (safety from people?), but nobody
would have confused Toto with a rat, at least based on its looks. We’ll see
later that Toto’s brain, in particular its use of distributed representation, was
also rat-like, but more on that in a bit.
To summarize, Toto’s navigation layer manifested the following behaviors:
1. The lowest-level behavior kept it moving safely around its environment

without colliding with static obstacles or moving people.
2. The next level behavior kept it near walls and other objects in the envi-
ronment.
16.2.3 Toto’s Landmark Detection

The next thing Toto did was to pay attention to what it was sensing and how
it was moving. This is generally a good thing to do. It allowed Toto to notice
when it was moving straight in the same direction for a while, or when it was
meandering around a lot. Moving straight indicated that it might be next to
a wall or in a corridor, and so Toto was able to detect walls and corridors.
Meandering around indicated that it was in a cluttered area, and so Toto
would detect that as well. So while Toto moved about, one of its behaviors
was paying attention to what its sonars “saw,” and in that way detecting and
recognizing landmarks: If Toto kept moving along in a consistent compass
direction (in a near-straight line) and kept close to a wall on one side (say the
left), it would recognize and detect a left-wall (LW). If the wall was on the
other side, it detected a right-walls (RW), and if it saw walls on both sides, it
recognized a corridor (C). Finally, if it detected nothing straight for a while
(no walls on either side), causing it to meander about, it declared a messy-
area (MA). In Toto’s mind, left walls, right walls, corridors, and messy areas
were special, because it could notice them easily from its sensors and its own
behavior. Therefore, they formed nice landmarks for his map.
Toto’s landmark detection layer also noted the compass direction for each
of the landmarks it detected. Its 2-bit compass (that’s not meant as an insult,
it is an accurate description) indicated north as 0, east as 4, south as 8, west
as 12, and so on, as shown in figure 16.5. In addition to the compass direction
of the landmark, Toto also noted the landmark’s approximate length.
To summarize, Toto’s landmark detection layer manifested the following
behavior:
1. It kept track of the consistency in the sonar and compass readings, and,
noticing those, recognized one of the landmark types (left-wall, right-
wall, corridor, messy-area).
2. For each landmark type, it also noted the compass direction and length of
the landmark.
16.2.4 Toto’s Mapping Behaviors

In order to store those landmarks, Toto constructed a map, as shown in fig-
ure 16.5. Here we come upon the really interesting fact about Toto: its dis-
tributed map representation. Each of the landmarks that Toto discovered
was stored in its own behavior. When Toto’s landmark detection layer dis-
covered a corridor (C) going north (0) for 6.5 feet, it created a behavior that
looked something like this:
my-behavior-type: C
my-compass-direction: 0
my-approximate-location: (x,y)
my-approximate-length: 6.5
whenever received (input)
if input(behavior-type) = my-behavior-type
and
input(compass-direction) = my-compass-direction
then
active <- true
What you see above is pseudo-code, not really Toto’s code but something
close to it and easer to read. The code above simply says that the behavior
represents a corridor (C) going north (0), so that whenever the input into that
behavior matches that description (i.e., when the input is C and 0, and the
Figure 16.5 An example of an indoor environment Toto navigated in, and a map it
constructed of that environment.
approximate length and location are right), the landmark matches and the
robot is in fact in that particular corridor.
Toto’s landmarks were clever: they worked regardless of which way Toto
was going down a corridor or along a wall. A C0 shown above would also
match C8 (because that corridor goes north-south, so Toto might be heading
down the same corridor in the opposite direction), and C4 would also match
C12 (because that corridor goes east-west). The same rule applied to walls as
well: a left wall going north (LW0) matched a right wall going South (RW8),
and so on.
Whenever the landmark-detecting layer detected a landmark, its descrip-
tion (type and compass direction) was sent to all map behaviors at the same
time (in parallel). If any behavior in the map matched the input, as described
above, then that behavior would become active, meaning Toto knew where it
was in its map; this is called localization and is a very important part of the
navigation problem. We will talk about it in detail in Chapter 19.
Ok, so far we have Toto moving around safely, detecting landmarks, and
reporting those to the behavior map. Whenever one matches, Toto recognizes
that it is in a place where it has been before.
What happens if no behavior in the map matches the landmark Toto is
detecting?
That means Toto has discovered a new place/landmark, which needs to be
added to the map. Suppose Toto is seeing a corridor (C) going east (4) that is
approximately 5 feet long. Here is how it adds that new landmarks/behavior
to its map: it takes a brand new, empty behavior “shell” and assigns it to the
newly found landmark information, so it becomes:
my-behavior-type: C
my-compass-direction: 4 or 12
my-approximate-location: (x,y)
my-approximate-length: 5.0
whenever received (input)
if input(behavior-type) = my-behavior-type
and
input(compass-direction) = my-compass-direction
then
active <- true
It then connects the previous behavior (suppose it is the C0 we just saw

earlier) to the new one by putting a communication link between them. This
means that when Toto is going north on C0, the next landmark it will see is
C4.
To take advantage of knowing where it is (being localized), Toto’s map
did another clever thing: the behavior that was active sent messages to its
neighbor in the direction in which Toto was traveling (for example, in the
map shown in figure 16.5, C0 sent messages to C4 if Toto was going north,
but to C12 if Toto was going south), to warn it that it should be next to be
active. If the neighbor was next to be recognized, then Toto was even more
confident about the accuracy of its map and its localization. On the other
hand, if the expecting neighbor was not recognized next, that usually meant
that Toto had branched off on a new path it had not tried before, and would
soon discover new landmarks to be added to the map.
Besides sending messages to its neighbors to tell them who is coming next,
Toto’s map behaviors sent inhibitory messages to one another. Specifically,
when a behavior matches the incoming landmark, it sends messages to all
of its neighbors to inhibit them, so that only one map behavior can be active
at one time, since Toto can really be only in one place at one time. That’s
another part of Toto’s localization. (If this seems like a lot of machinery to
assure Toto where it is, you’ll see in Chapter 19 how hard it is for any robot
to know where it is.)
To summarize, Toto’s mapping layer performed the following behaviors:
1. Matched the received landmark against all landmark behaviors in the

map
2. Activated the map behavior that matched the current landmark
3. Inhibited the previous landmark
4. Reported if no landmark matched, and created a new landmark behavior

in response, storing the new landmark and connecting it to the previous
landmark.
16.2.5 Path Planning in Toto’s Behavior Map

As we have seen, while Toto roamed around its environment, it built up a
landmark map of its world. This is an illustration of how you can build a
distributed map with a behavior-based system. Now let’s see how such a
map is used to find/plan paths.
Let’s suppose that Toto’s watering can was empty and it needed to go to
the corridor with the water supply, so this landmark became Toto’s goal. If
the goal landmark and Toto’s current position were the same behavior in the
map, then Toto was already at the goal and did not need to go anywhere.
That’s too lucky to have been the case most of the time. Instead, Toto had to
plan a path between its current location and the goal. Here is how it did that.
The behavior that corresponded to the goal sent messages saying “Come
this way!” to all of its neighbors. The neighbors, in turn, sent the messages on
to their own neighbors (but not back the way the message came; that would
be a waste and create some problems). Each map behavior, when passing a
message along, also added to it its landmark length, so that as the path went
Figure 16.6 Toto’s way of path planning.
through the network, its length grew based on the number and lengths of
map behaviors it passed through.
Rather quickly, these messages all reached the currently active behavior
which represented wherever Toto was at the time. When they arrived, the
messages contained the total length of the path taken to get there, in terms of
physical landmark length. The current landmark paid attention only to the
message with the minimum summed landmark length, since that indicated
the direction toward the shortest path to the goal.
Besides going to a particular landmark, such as a specific corridor, Toto
could also find the nearest landmark with a particular property. For example,
suppose that Toto needed to find the nearest right wall. To make this happen,
all right wall landmarks in the map would start sending messages. Toto
followed the shortest path, and reached the nearest right wall in the map.
16.2.6 Toto’s Map-Following

As we learned in Chapter 13, the environment can change at any point, the
robot can get lost or confused about where it is, and so on, and therefore
planning a path once and sticking to in an open loop fashion is a dangerous
prospect. So Toto did not plan once, but kept planning continually. Here is
how.
Toto’s map behaviors sent messages continuously, so wherever in the map
Toto happened to be (whatever behavior happened to be active) it could re-
spond to the shortest path it received. This is a very useful property if the
environment is likely to change while the robot is moving around. Another
reason for such continual updating of the plan is the possibility of the so-
K IDNAPPED R OBOT called kidnapped robot problem, in which somebody picks up the robot and
P ROBLEM moves it to another location. Continuously recomputing the path by send-
ing the messages around means that a blocked path or changed location will
immediately be replaced by a better (shorter) path in the network. In fact, it
is very important to notice that Toto did not really have any notion of a "path"
at all! Instead, it just went from one landmark to the next, which eventually
led it to the goal, as efficiently as possible within its map. It did not compute
the path explicitly or store it anywhere, but instead used the active behavior
map to find its way, no matter where it was.
Toto’s way of building a map and planning within its active structure is
not easy to describe, but it is easy (efficient) to compute. It is also not easy
to illustrate, but figure 16.6 tries to do so, by showing how messages are
passed between neighboring behaviors in Toto’s map in order to plan a path
between its current behavior and goal behavior.
To summarize, Toto’s path finding layer performed the following behav-
iors:
1. Repeatedly sent (one or more) messages from the goal throughout the
map network, until the goal was reached
2. Kept a running total of the length of the path as the messages were passed
through the network
3. At each landmark Toto reached (and thus recognized and activated), the
current total of incoming messages indicated where Toto should go next
(to what behavior in the map), telling it how to turn and where to head.
As a result of all of its behaviors operating in parallel and interacting with

each other and with the environment, Toto was able to:
1. Move around safely and quickly

2. Detect new and known landmarks
3. Build a map and find its position in the map
4. Find shortest paths toward various goals.
All of the above was achieved in real time, so no detectable time-scale is-
sues arose. Additionally, no intermediate layer existed in Toto because there
was no need to explicitly interface any layers at different time-scales (e.g.,
navigation vs. path finding). Toto is a good example of a complete behavior-
based system capable of achieving both real-time behavior and optimization
without the use of a hybrid controller.
This example, albeit somewhat complicated, shows us that BBC and hy-
brid control are equally powerful and expressive. By using distributed rep-
resentations, we can do arbitrary tasks with BBC, just as with hybrid systems.
How does one decide when to use a hybrid controller and when to use a
behavior-based one?
There is no fixed answer to that question; the decision is often based on
the designer’s personal preferences and past experience and expertise. In
some domains, there are advantages to one approach or the other, however.
For example, for single-robot control, especially for complex tasks involving
representation and planning, hybrid systems are typically most popular. In
contrast, for controlling groups and teams of robots, behavior-based control
is often preferred, as we will learn in Chapter 20.
Some fans of behavior-based control claim that BBC is the most realistic
model of biological cognition (thinking), because it includes the spectrum
of computational capabilities, from simple reflexes to distributed represen-
tations, all within a consistent time-frame and representation. Other fans
prefer BBC for pragmatic reasons, finding it easy to program and use for de-
veloping robust robots. (Of course, those who prefer hybrid systems often
list the very same reasons for their opposite choice!)
To Summarize
• Behavior-based systems (BBS) use behaviors as the underlying modular-

ity and representation.
• BBS enable fast real-time responses as well as the use of representation
and learning.
• BBS use distributed representation and computation over concurrent be-

haviors.
• BBS are an alternative to hybrid systems, with equal expressive power.
• BBS require some expertise to program cleverly, as any control approach

does.
Food for Thought
• Some people say that behaviors should be more precisely defined so that
behavior-based robot programming would be easier to figure out. Others
believe that having behaviors be loosely defined allows robot program-
mers to be creative and come up with interesting and novel ideas for ways
to design robot controllers. What do you think?
• The field of psychology went through a phase called “behaviorism”, which

defined all organisms based on their observable behaviors. This general
idea is similar to behavior-based robotics. However, behaviorism believed
that there was no difference between externally observable behaviors (ac-
tions) and internally observable behaviors (thinking and feeling). Some
behaviorists did not even believe in internal state (recall that from Chap-
ter 3). As we saw in this chapter, BBS do have internal state and represen-
tation, so they are not consistent with behaviorism. People often confuse
behaviorism and behavior-based systems. Is the difference clear to you?
If not, you might want to read more about both.
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Behavior-Based_Cont-
rol
• You can learn more about Toto from the following papers:
Matarić, Maja J. (1990), “A Distributed Model for Mobile Robot Environm-
ent-Learning and Navigation”, MIT EECS Master’s Thesis, Jan 1990, MIT
AI Lab Tech Report AITR-1228, May.
Matarić, Maja J. (1992), “Integration of Representation Into Goal-Driven
Behavior-Based Robots”, IEEE Transactions on Robotics and Automation, 8(3),
June, 304-312.
• Cambrian Intelligence by Rodney A. Brooks is a great collection of papers

about behavior-based robots built at MIT during the 1980s and 1990s, in-
cluding Genghis, Attila, Toto, and several others.
• After finishing this book, Ronald Arkin’s Behavior-Based Robotics is an ex-

cellent next step in learning about robotics.
17 Making Your Robot Behave
Behavior Coordination
Any robot that has at its disposal more than one behavior or action has to
B EHAVIOR solve the surprisingly difficult problem of action selection or behavior coordina-
COORDINATION tion.
The problem is simple on the surface: what action/behavior should be
executed next?
Solving this problem in a way that makes the robot do the right thing over
time, so as to achieve its goals, is anything but simple. Both hybrid and
behavior-based systems have multiple modules, and therefore have to solve
this coordination problem. In hybrid systems, the middle layer is often stuck
with the job, while in behavior-based systems the job may be distributed
throughout the entire control system.
In general, there are two basic ways of selecting the next behavior or the
next action: picking one behavior/action or combining multiple behaviors/a-
ctions. As we mentioned in Chapter 14, these two approaches are called ar-
bitration and fusion. Figure 17.1 shows some of the many methods available
for implementing each and combinations of the two.
17.1 Behavior Arbitration: Make a Choice
A RBITRATION Arbitration is the process of selecting one action or behavior from multiple
possible candidates. We have already discussed arbitration in the context
of reactive control, in Chapter 14, but it has a much broader role to play in
robotics in general, since it applies any time there is a choice to be made
among different actions or behaviors.
208 17 Making Your Robot Behave
Figure 17.1 Various methods for behavior coordination.
Arbitration-based behavior coordination is also called competitive behavior

coordination, because multiple candidate behaviors compete but only one
can win.
Arbitration can be done with:
• A fixed priority hierarchy (behaviors have preassigned priorities)
• A dynamic hierarchy (behavior priorities change at run-time).
Subsumption Architecture (from Chapter 14) used a fixed priority hierar-

chy of behaviors, which was implemented through inhibition of the behav-
iors’ outputs and suppression of their inputs.
Many hybrid systems employ fixed priority hierarchies of control. As we
saw in Chapter 15, in some systems the planner always has control over the
reactive system, and in others the reverse is always the case.
More sophisticated hybrid systems use dynamic hierarchies in order to

have the control of the system transition between the reactive and delibera-
tive parts of the controller, for best performance. Similarly, behavior-based
systems often employ dynamic arbitration to decide what behavior wins and
takes control of the robot next.
Dynamic arbitration usually involves computing some function to decide
who wins. The function can be anything, including voting (so the winner is
"elected"), fuzzy logic, probability, or spreading of activation, among many
others.
17.2 Behavior Fusion: Sum It Up
In contrast to behavior arbitration, a competitive method of control, there are

various architectures that employ the execution of multiple behaviors at the
same time. This is the basis of behavior fusion.
B EHAVIOR F USION Behavior fusion is the process of combining multiple possible candidate ac-
tions or behaviors into a single output action/behavior for the robot. Behav-
ior fusion is also called a "cooperative" method because it combines outputs
of multiple behaviors to produce the final result, which may be one or more
of the existing behaviors, or even a wholly new result. (How can that be?
Read on, you’ll see.)
Combining behaviors, however, is a very tricky business. Even walking
and chewing gum at the same time is sometimes a challenge, and things
only get more difficult from there. (Robots are no good with gum, and they
are not great at walking yet, either.)
In order to combine two or more behaviors, the behaviors must be repre-
sented in a way that makes that combination possible and, hopefully, easy.
Here is an example of how that could be done.
Consider a robot that has the following behaviors:
Stop:
command velocity 0 to both wheels
GoForward:
command velocity V1 to both wheels
TurnRightABit:
command velocity V2 to the right wheel,
and 2*V2 to the left wheel
TurnLeftABit:
command velocity V3 to the left wheel,
and 2*V3 to the right wheel
TurnRightALot:
and 3*V4 to the left wheel
TurnLeftALot:
and 3*V5 to the right wheel
SpinInPlaceToTheRight:
and -V6 to the right wheel
SpinInPlaceToTheLeft:
and -V7 to the left wheel
The control fusion mechanism takes the outputs of all of the behaviors,
sums the velocity commands for each wheel, and sends the result to each
wheel.
Do you see any potential problems with this approach?
It is certainly nice and simple, and since it uses only velocities, all values
can be directly added and subtracted. But what happens if two behaviors
cancel each other out? What if V2 and V3 are equal, and TurnRightABit and
TurnLeftABit are commanded at the same time? What will the robot do? It
will not turn at all! Worse yet, if we combine the Stop behavior with any
other, the robot will not stop, and that could be a very bad thing if there is an
obstacle in front of it.
The most obvious basic case in which this problem arises is around obsta-
cles. Consider the following (popular) basic controller:
GoToGoal:
compute the direction toward the goal
go in that direction
AvoidObstacle:
compute the direction away from the obstacle
go in that direction
Together, these two behaviors can make the robot oscillate in front of ob-
stacles, doing a little dance and never making progress.
How might we get around some of these problems?
It turns out that roboticists have thought about this long and hard, and
there is a great deal of theory and even more practice showing what works
and what doesn’t. For example, some behaviors could be weighted so that
their outputs have a stronger influence on the final result. Of course, that
will not always work either. Another possibility is to have some logic in
the system which prevents certain combinations and certain outcomes. This
makes the otherwise nice and simple controller not so nice and simple any
more, but it is often necessary. There is no free lunch.
It is in general easy to do command fusion when the commands are some
numerical value, such as velocity, turning angle, and so on. Here is an exam-
ple from "the street":
Consider the DAMN (Dynamic Autonomous Mobile Navigation) Archi-
tecture, which was used to control a van that autonomously drove on roads
over very large distances. (DAMN was developed by Julio Rosenblatt at
CMU; see the end of the chapter for details.) The demo was called "No Hands
Across America" and got a lot of well-deserved publicity. In the DAMN
controller, various low-level actions, such as turning angles, were all voting
members. Each action got to vote (for itself or others), all votes were tallied,
and the result, a weighted sum of actions, was executed. Figure 17.2 shows
a diagram of the control architecture. This is a clear example of command
fusion and it turned out to be a very effective way of controlling the driving
robot. Later on, it was improved further through the use of fuzzy logic, an-
other way to fuse commands. You can find more information about fuzzy
logic at the end of the chapter.
Formal methods for command fusion include potential fields, motor sche-
mas, and dynamical systems, to name but a few. These are all well-developed
Figure 17.2 A diagram of the DAMN architecture for dynamic autonomous mobile
navigation (thus the name).
and deserve careful study. We’ll skip them here but you can find pointers to
additional reading at the end of this chapter.
In general, fusion methods are applied to lower-level representations of
actions or behaviors (direction, velocity), while arbitration is used for higher-
level ones (turn, stop, grasp). Many control systems use a combination of
methods, with fusion at one level and arbitration at another.
To Summarize
• Deciding what action, behavior, or set of behaviors to execute is a complex

problem in robotics.
• This problem is called behavior coordination.
• Behavior coordination has been studied for a long time and there are a
great many possible solutions. However, no solution is perfect and there
are no strict rules about what is the right solution for a given robot sys-
tem, environment, and task.
Food for Thought
Does it ever happen to you that you can’t decide between two options?
What do you do? Theories from cognitive science and neurosciences say
Does it ever happen to you that you can’t decide between two options? What
do you do? Theories from cognitive science and neurosciences say that hu-
man emotions are useful for helping us decide between similar options; there
exist people with certain neural deficits who are both unemotional and un-
able to choose between things that appear similarly good (or bad). Does this
mean robots need emotions? Some researchers in AI have argued that both
robots and AI software in general could benefit from such emotions. What
do you think?
Looking for More?
• Ronald Arkin’s Behavior-Based Control gives an easy-to-understand expla-

nations for some of the more popular methods for behavior fusion, in-
cluding schemas and potential fields.
• Julio Rosenblatt’s PhD dissertation was titled DAMN: A Distributed Ar-

chitecture for Mobile Navigation; you can get it as a technical report from
CMU. In general, you can get any PhD dissertation from the institution
that issued it.
• Paolo Pirjanian’s PhD dissertation Multiple Objective Action Selection & Be-
havior Fusion using Voting from Aalborg University in Denmark provides
an excellent overview of behavior coordination approaches.
• Fuzzy logic is a rather popular topic with a great many books published
on the subject. You might want to start with Fuzzy Logic: Intelligence, Con-
trol and Information by John Yen and Reza Langari.
18 When the Unexpected Happens
Emergent Behavior
The whole idea that a robot can produce unexpected behavior is somewhat
scary, at least to a non-roboticist.
But is any robot behavior really unexpected? And is all unexpected
behavior bad?
In this chapter we will learn about emergent behavior, an important but
poorly understood phenomenon.
All robot behavior results from the interaction of the robot’s controller and
the environment. If either or both of these components are simple, the result-
ing behavior may be quite predictable. On the other hand, if the environment
is dynamic, or if the controller has various interacting components, or both,
the resulting robot behavior may be surprising, in a good or bad sense. If
such unexpected behavior has some structure, pattern, or meaning to an ob-
server, it is often called emergent. But hang on, we won’t define emergent
behavior yet; we are not quite ready for that.
18.1 An Example: Emergent Wall-Following
Although it is easy to imagine how unexpected behavior results from com-

plicated systems, you should not assume that it takes a complex environment
or a complex controller to generate emergent behavior. For a counterexam-
ple, consider the following simple controller (which should by now be quite
familiar to you):
If left whisker bent, turn right.

If right whisker bent, turn left.
If both whiskers bent, back up and turn to the left.
216 18 When the Unexpected Happens
Figure 18.1 Emerging wall-following behavior by a simple mobile robot.
Otherwise, keep going.
What happens if you put a robot with this controller next to a wall? See
figure 18.1 for the answer. It follows the wall! And yet, the robot itself knows
nothing about walls, since the controller has nothing explicit about walls in
its rules. No walls are detected or recognized, yet the robot follows walls
reliably and repeatably. In fact, the controller above might have been written
simply for obstacle avoidance, but it follows walls, too.
So is wall-following an emergent behavior of this controller?
Well, that depends.
What is emergent behavior, and is it magic?
18.2 The Whole Is Greater Than the Sum of Its Parts
E MERGENT B EHAVIOR The intuitive if rather imprecise notion of emergent behavior is that it involves
some type of “holistic” capability, where the behavior of the robot is greater
than the sum of its “parts,” which are the rules in the controller. Emergent be-
18.2 The Whole Is Greater Than the Sum of Its Parts 217
havior appears to produce more than what the robot was programmed to do.
Somehow we get more than we built in, something for nothing, something
magic.
As you saw in Chapters 14 and 16, interactions between rules, behaviors,
and the environment provide a source of richness and expressive power for
system design, and good roboticists often exploit this potential to design
clever and elegant controllers. Specifically, reactive and behavior-based sys-
tems are often designed to take advantage of just such interactions; they are
designed to manifest what some might call emergent behaviors.
Let’s consider another example. Suppose you have a group of robots and
you need to make them flock together. Suppose further that you cannot
have the robots communicate with each other and must have each robot
just follow its own rules and its own local sensory data (just as birds do
when they flock). How would you program each robot so as to have the
group flock?
It turns out that such distributed flocking can be achieved with the follow-
ing set of simple rules for each of the robots:
Don’t get too close to other robots (or other obstacles)

Don’t get too far from other robots
Keep moving if you can
What we set “close” and “far” depends on various parameters, including

the velocities of the robots, their sensor range, and on how tight we want the
flock to be.
When run in parallel, the rules above will result in the group of robots
flocking. And furthermore, if you want to get the flock to go to a partic-
ular place, such as South (useful for birds) or Home (useful for everyone),
you need to add only one more rule, which “nudges” each robot to the de-
sired destination, assuming the rest of the rules above are satisfied. To learn
more about this type of flocking, wait for Chapter 20 and also look into the
resources at the end of this chapter.
What does it take to create an emergent behavior?

18.3 Components of Emergence

The notion of emergence depends on two components:
1. The existence of an external observer, to see the emergent behavior and

describe it (because if nobody sees the emergent behavior, how do we
know it has happened?)
2. Access to the innards of the controller, to verify that the behavior is not
explicitly specified anywhere in the system (because if it is, then it is not
emergent, merely expected).
Many roboticists use the combination of the two components above as the
definition of emergent behavior. The wall-following and flocking behaviors
described above work quite nicely for this definition.
Of course, not all researchers agree with that definition (such ready agree-
ment in a research field would be just too easy and too boring). Some argue
that there is nothing emergent about the wall following and flocking behav-
iors, and that they are merely examples of a particular style of indirect robot
programming that uses the environment and side-effects.
These critics propose that in order for a behavior to be emergent, it must be
truly unexpected; it must be discovered “surreptitiously” by observing the
system, and must come as a true surprise to the external observer.
18.4 Expect the Unexpected

The problem with the unexpected surprise as a property of a behavior is
that it depends entirely on the expectations of the observer, and those ex-
pectations are completely subjective. A naive observer is surprised by many
observations. A prehistoric man would be very surprised indeed by a bicycle
staying upright or a plane flying. (In fact, the only reason most of us are not
surprised by these phenomena is from familiarity; it is certainly not because
most of us actually understand exactly why bicycles and airplanes do not
fall.) In contrast, an informed observer may not be surprised by much, if he
or she understands the system extremely well or has seen it in action before.
This last point above brings us to an even trickier problem with unex-
pected emergent behavior. Once such a behavior is observed, it is no longer
unexpected. Just as with the bicycle and airplane, even a caveman would
eventually get used to seeing them upright or in the sky, and would expect
them to stay so, even without understanding how the system works. And
18.5 Predictability of Surprise 219
that lack of surprise – the fulfilled expectation, the predictability – invali-

dates the second definition of emergence. But this is no good, because a
behavior either is or isn’t emergent; it should not be emergent once (the first
time it is seen), and then never again.
18.5 Predictability of Surprise
So, it seems pretty clear that the knowledge of the observer cannot be a valid
measure of emergence, and that something more objective is needed. A more
objective and concrete way to try to define and understand emergence is to
look a bit more closely at the property of the controller “not knowing” about
the behavior and yet generating it.
As we saw in the wall-following example, the controller (and thus the
robot) had no notion of “walls” or “following.” Wall-following emerged
from the interaction of the simple avoidance rules and the structure of the en-
vironment. If there were no walls in the environment, wall-following would
not emerge. This brings us to another definition of emergent behavior:
E MERGENT B EHAVIOR Emergent behavior is structured (patterned, meaningful) behavior that is ap-
parent at one level of the system (from the observer’s viewpoint), but not at
another (from the controller’s/robot’s viewpoint).
This means that if the robot generates interesting and useful behavior with-
out being explicitly programmed to so, the behavior can be called emergent,
whether or not it comes as a surprise to any observer.
Some such behaviors, such as wall-following and flocking examples, can
be preprogrammed through clever rules and are not particularly surprising
to the designer, but are still clever ways to do robot control. However, a
great deal of emergent behavior cannot be designed in advance, and is in-
deed unexpected, produced by the system as it runs and emerging from the
dynamics of interaction that arise only at execution time. In such cases, the
system has to be run and observed in order for the behavior to manifest itself;
it cannot be predicted.
Why is that? Why can’t we, in theory, predict all possible behaviors of
the system (assuming we are very smart)?
Well, we could in theory predict all possible behaviors of a system, but
it would take forever, because we would have to consider all possible se-
quences and combinations of the robot’s actions and behaviors in all possible
environments. And to make it even more complicated, we would also have
to take into account the unavoidable uncertainty of the real world, which
affects all of the robot’s sensing and actions.
It is easy to misunderstand and misinterpret these ideas. The fact that
we cannot predict everything in advance does not mean we cannot predict
anything. We can make certain guarantees and predictions for all robots and
environments and tasks. We can, for example, predict, and thus guarantee
the performance of a surgical robot. If we couldn’t, who would use it? Still,
we cannot be sure that there may not be an earthquake while it operates,
which may well interfere with its performance. On the other hand, for some
controllers and robots we may not, for example, be able to predict what robot
will put what brick where but can guarantee that the robots will, together,
build a wall (unless of course somebody steals all the bricks or there is an
earthquake, again).
And so, we are left with the realization that the potential for emergent
behavior exists in any complex system, and since we know that robots that
exist in the real world are, by definition, complex systems, emergence be-
comes a fact of life in robotics. Different roboticists approach it in different
ways: some work hard to avoid it, others aim to use it, and still others try to
ignore it whenever possible.
18.6 Good vs. Bad Emergent Behavior

Notice that so far we have assumed that all emergent behavior is, in some
way, desirable and good. Of course all sorts of structured behaviors can
emerge that are not at all desirable, but instead quite problematic for the
robot. Oscillations are one obvious case in point; a robot with simple turn-
away-from-the-obstacle rules can easily get stuck in a corner for a while. Can
you think of other examples?
Can you think of how to make an emergent behavior come about without
programming it in directly?
Furthermore, for any kind of emergent behavior:
What is the difference between an emergent bug and an emergent fea-
ture?
That is entirely dependent on the observer and his or her goals. One ob-
server’s bug is another’s feature. If you want your robot to get stuck in cor-
ners, then oscillations are a feature; otherwise they are a bug.
Now let’s get back to purposefully creating emergent behavior in our robots.
18.7 Architectures and Emergence 221
Recalling our examples of wall-following and flocking, we can see that emer-
gent behavior can arise from the interactions of the robot and the environ-
ment over time and/or over space. In wall following, the rules (go left, go
right, keep going), in the presence of a wall, result over time in following the
wall. So what is necessary, besides the controller itself, is the structure in the
environment and enough time for the behavior to emerge.
In contrast, in the flocking behavior, what is necessary is the many robots
executing the behaviors together, in parallel. This is again a type of envi-
ronment structure, albeit in the form of other robots. The rules themselves
(don’t go too close, don’t go too far, keep going) mean nothing if the robot is
alone, and in fact it takes three or more robots for such flocking to be stable;
otherwise it falls apart. But given enough robots, and enough space and time
for them to move, a flock emerges.
Given the necessary structure in the environment, and enough space
and/or time, there are numerous ways in which emergent behaviors can
be programmed and exploited.
This is often particularly useful in programming teams of robots; see Chap-
ter 20 to learn more.
18.7 Architectures and Emergence

As you would expect, the different control architectures affect the likelihood
of generating and using emergent behavior in different ways, because system
modularity directly affects emergence.
Reactive and behavior-based systems employ parallel rules and behaviors,
respectively, which interact with each other and the environment, thus pro-
viding the perfect foundation for exploiting emergent behavior by design.
In contrast, deliberative systems are sequential, and thus typically have no
parallel interactions between the components, and thus would require en-
vironment structure in order to have any behavior emerge over time. Simi-
larly, hybrid systems follow the deliberative model in attempting to produce
a coherent, uniform output of the system, minimizing interactions and thus
minimizing emergence.
To Summarize
• Emergent behavior can be expected or unexpected, desirable or undesir-

able.
• Roboticists try to avoid the undesirable but exploit the desirable emergent
behavior.
• Emergent behavior requires an observer of the system’s behavior and in-

formation about the insides of the system
• Different control architectures have different methods for exploiting or

avoiding emergent behavior.
• You can spend a lot of years studying emergent behavior and philoso-
phizing about it.
• People, the ultimate complex systems situated in complex environments,

produce emergent behavior, too.
Food for Thought
• Besides wall-following, or following (servoing to some target), which is a

natural behavior to produce via emergence, can you think of others you
might want to implement this way?
• Can you think of ways of using emergent behavior to do something quite

complicated, such as build a bridge or a wall or a building? Ants do it all
the time, see Chapter 20.
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Emergent_Behavior
• Craig Reynolds invented “boids”, simulated bird-like creatures that were

programmed with simple rules and produced elegant flocking behaviors.
You can read about boids in Artificial Life by Steven Levy. You can also find
more information about boids, and watch videos of the flocks in action,
on the Internet.
19 Going Places
Navigation
N AVIGATION Navigation refers to the way a robot finds its way in the environment.
Getting from one place to another is remarkably challenging for a robot.
In general, you will find that any robot controller spends a great deal of its
code getting where it needs to be at any given time, compared with its other
“high-level” goals. Getting any body part where it needs to be is hard, and
the more complicated the robot’s body, the harder the problem.
The term “navigation” applies to the problem of moving the robot’s whole
body to various destinations. Although the problem of navigation has been
studied in the domain of mobile robots (as well as flying and swimming
robots), it applies to any robot that can move around. The body of the robot
may be of any form; the locomotion mechanism takes care of moving the
body appropriately, as we saw in Chapter 5, and the navigation mechanism
tells it where to go.
What’s so hard about knowing where to go?
As usual, the problem is rooted in uncertainty. Since a robot typically does

not know exactly where it is, this makes it rather hard to know how to get
to its next destination, especially since that destination may not be within its
immediate sensory range. To better understand the problem, let’s break it
down into a few possible scenarios, in all of which the robot has to find some
object, let’s say a puck.
• Suppose the robot has a map of its world which shows where the puck
is. Suppose also that the robot knows where it is in its world, the map. What
remains to be done to get to the puck is to plan a path between the robot’s
current location and the goal (the puck), then follow that plan. This is the
path planning problem. Of course if anything goes wrong - if the map is in-
correct or the world changes - the robot may have to update the map, search
224 19 Going Places
around, replan, and so on. We already talked a bit about how to deal with
such challenges of deliberative and hybrid control in Chapters 13 and 15.
• Now suppose the robot has a map of its world which shows where the
puck is, but the robot does not know where it is in the map. The first thing
the robot must do is find itself in the map. This is the localization problem.
Once the robot localizes within its map, i.e., knows where it is, it can plan the
path, just as above.
• Now suppose the robot has a map of its world and knows where it is in
its map, but does not know where the puck is in the map (or world, same
thing). What good is the map if the location of the puck is not marked?
Actually, it is a good thing indeed. Here is why. Since the robot does not
know where the puck is, it will have to go around searching. Given that it
has a map, it can use the map to plan out a good searching strategy, one that
covers every bit of its map and is sure to find the puck, if one exists. This is
C OVERAGE P ROBLEM the coverage problem.
• Now suppose the robot does not have a map of its world. In that case,
it may want to build a map as it goes along; this is the mapping problem.
Notice that not having a map does not mean the robot does not know where
it is; you may not have a map of New York City, but if you are standing
next to the Statue of Liberty, you know where you are: next to the Statue of
Liberty. Basically, a robot’s local sensors can tell it where it is if the location
can be uniquely recognized, such as with a major landmark like the Statue
of Liberty, or with a global positioning system (GPS). However, without a
map you won’t know how to get to the Empire State Building. To make
the problem really fun, and also quite realistic for many real-world robotics
domains, suppose the robot does not know where it is. Now the robot has
two things it has to do: figure out where it is (localization) and find the puck
(search and coverage).
• Now suppose that the robot, which has no map of its world and does not
know where it is, chooses to build a map of its world as it goes along trying
to localize and search for the puck. This is the simultaneous localization and
mapping (SLAM) problem, also called concurrent mapping and localization
(CML), but that term is not as catchy or popular. This is a “chicken or egg”
problem: to make a map, you have to know where you are, but to know
where you are, you have to have a map. With SLAM, you have to do both at
the same time.
All of the various problems mentioned above are components of the navi-
gation problem. To summarize, they are:
• Localization: finding out where you are
• Search: looking for the goal location
• Path planning: planning a path to the goal location
• Coverage: covering all of an area
• SLAM: localization and mapping at the same time.
As it turns out, every one of the above problems has been studied ex-
tensively in robotics, and numerous research papers, and even some books,
have been written about them. Various algorithms exist for every one of the
above problems, and more are developed every year by many eager robotics
researchers. In short, all of the above are hard problems, still considered
“open,” meaning research is ongoing in trying to solve them in increasingly
better ways. Given that, it is hard to do justice to the large body of work
that exists in these areas of navigation. Next, we will briefly mention some
of the approaches to the problems above, but to learn more, check out the
additional reading sources given at the end of the chapter.
19.1 Localization
O DOMETRY One way to stay localized is through the use of odometry. Odometry comes
from the Greek hodos meaning “journey” and metros meaning “measure,” so
PATH I NTEGRATION it literally means measuring the journey. A more formal term for it is path
integration. Cars have odometers, as do robots. Robot odometry is usually
based on some sensors of the movement of the robot’s wheels, typically shaft
encoders, as we learned in Chapter 8. However, the farther the robot goes
and the more turns it makes, the more inaccurate odometry becomes. This
is unavoidable, since any measurement of a physical system is noisy and
inaccurate, and any small error in the odometry will grow and accumulate
over time. The robot will need to recognize a landmark or find some other
way to reset its odometer if it is to have any chance of localizing. Odometry
in robots is the same basic process that people use in dead reckoning.
Odometry allows the robot to keep track of its location relative to the start-
ing or reference point. Therefore, it is localized relative to that point. In
general, localization is always relative to some reference frame, whether it is
an arbitrary starting point or a GPS map.
To deal with localization in a more formal way, it is usually treated as a
S TATE E STIMATION state estimation problem. State estimation is the process of estimating the state
226 19 Going Places
Figure 19.1 Problems associated with localization.
of a system from measurements. This is a hard problem for the following

reasons:
1. The estimation process is indirect. One does not usually have the luxury of
directly measuring the quantity. Instead, one measures what one can, and
estimates the state based on those data (“data” is the plural form of the
word “datum”). As we discussed above, a robot may keep track of its
wheel rotations, and thus use odometry in order to estimate its location.
It is not measuring the location directly; it is estimating it from odometry
measurements.
2. The measurements are noisy. Of course they are; they involve real-world
sensors and real-world properties. By now this should come as no sur-
prise to you, right?
3. The measurements may not be available all the time. If they are, the state es-
timation may be continuous; otherwise, it may be done in batches, when
enough data have been collected. For example, odometry is continuous,
but other measurements may not be, such as counting mile markers or
other types of intermittent landmarks.
State estimation usually requires a model of the relevant quantities in the

environment. For odometry, the only model needed is the reference frame in
which the distance is measured. However, more complex types of localiza-
tion require more complex models. For example, consider a robot that has a
range sensor (sonar or laser) and a map of the environment. The robot can
take a set of range measurements and compare those with the map in order
to see where it is most likely to be. This is a tricky process, because several
places in the environment may look the same to the particular sensors. Sup-
pose that the robot is in a room that has four empty corners, all of which
look the same as the robot is facing them. In order to tell them apart, the
robot has to move around and perceive other features (walls, furniture, etc.)
so it can see where it is, relative to the map. This type of confusion is very
common in robot navigation, whether from corners, corridors, doorways, or
other features that often look similar, especially to robot sensors, which are
limited and noisy.
As discussed in Chapter 12, maps are just one way to represent the envi-
ronment, and they come in all sorts of varieties. Instead of having a detailed
map, the robot may know about only a few landmarks or targets, in which
case it has to compare its sensor readings (i.e., measurements) to the model
it has of those landmarks. It may only know their distances from each other,
and will have to calculate where it is based on that information. The local-
ization problem varies according to the type of representation used.
If the robot has a topological map (remember that concept from Chap-
ter 12?), localization means it has to be able to uniquely identify which land-
mark of that map it is in. This is not so hard if the nodes are unique (such as
the Statue of Liberty; there is only one in New York City), but is much harder
if there are multiple landmarks that look similar (such as a busy street corner
with a coffee shop and a newspaper stand; there are plenty of those in New
York City). If the robot has a metric map, localization means it has to be able
to identify its global Cartesian position (latitude and longitude). Fortunately,
that position is sure to be unique.
L OCALIZATION In summary, localization is the process of figuring out where the robot is
relative to some model of the environment, using whatever sensor measure-
ments are available. As the robot keeps moving, the estimate of its position
drifts and changes, and has to be kept updated through active computation.
So knowing where you are (when you are a robot) is not a trivial matter, and
being told to “Get lost!” is probably the easiest command to follow.
228 19 Going Places
19.2 Search and Path Planning
As we saw above, path planning involves finding a path from the robot’s
current location to the destination. This naturally involves the robot know-
ing its current location (i.e., being localized) and knowing the destination or
goal location, both in a common frame of reference. It turns out that path
planning is a pretty well understood problem, at least when all of the above
information is available to the robot. Given a map of the environment and
two special points on it (current and goal locations), the robot just has to
“connect the dots.”
Typically, there are many possible paths between the start and the goal,
and finding them involves searching the map. To make this searching com-
putationally efficient, the map is usually turned into a graph, a set of nodes
(points) and lines that connect them. Why bother? Because graphs are easy
to search using algorithms that have been developed in computer science
and artificial intelligence.
A path planner may look for the optimal (the very best) path based on some
criterion. A path may be optimal based on distance, meaning it is the short-
est, it may be optimal based on danger, meaning it is the safest; or it may
be optimal based on scenery, meaning it is the prettiest. Admittedly, this
last criterion is not usually used in robotics; the point is that there are vari-
ous criteria for path optimality, depending on the particular robot’s task and
environment.
Finding the optimal path requires searching all possible paths, because if
the planner does not check all possibilities, it may miss the very best one.
Therefore, this is a computationally complex (read: potentially slow) pro-
cess. In fact, path planning is one of those tasks that require robots to per-
form higher-level thinking or reasoning. This ends up being an important
capability that has major impact on how we program robots, as you saw in
Chapter 11.
Not all planners look for the optimal path, or even a complete path. Some
perform only local path planning, looking at the nearby part of the map, not
the whole thing, and therefore speeding things up (but potentially running
into problems like closed doors and walls later on). Other planners look for
the first path that gets to the goal, rather than the optimal one, again in order
to save time. As we talked about at length in Chapter 11 and thereafter, time
is very valuable to a robot, and path planning is one important task in a
mobile robot’s life that can be rather time-consuming. On the other hand,
getting to where the robot needs to go is important.
19.3 SLAM 229
Figure 19.2 Problems associated with path planning.
There is a great deal of work in robotics on various ways to represent the

environment, plan a path, and convert that path into a set of movement com-
mands to the robot. Check out some of the relevant books listed at the end
of the chapter.
19.3 SLAM
As you can imagine, SLAM is a difficult problem, since it involves having

the robot perform two ongoing and related parallel processes: figuring out
where it is and constructing a map of the environment. What makes this
difficult is the apparent confusion among multiple places that look similar
D ATA A SSOCIATION and are therefore ambiguous. This is the data association problem, the problem
P ROBLEM of uniquely associating the sensed data with absolute ground truth.
If the robot is constructing a topological map, it has to contend with the
challenge of uniquely identifying landmarks we described above. If it sees
a busy street corner with a coffee shop and a newspaper stand, how does
it know exactly which landmark it is seeing? For that matter, even if it is
standing next to a large statue holding a torch, how does it know that there
aren’t two such statues in New York City? It doesn’t, until it has the complete
230 19 Going Places
map, and by the definition of SLAM, it does not have a map – it is building
it as it goes along.
If the robot is constructing a metric map, it has to contend with the errors
in odometry and other sensor measurements that make it hard to be sure
what global (x,y) position it is actually at. For example, when it goes around
a few blocks and corners, can it be sure it has come back where it started? Or
is it a block off?
In short, it is easier if the robot can localize when building a map, or if it
can have a map in order to localize, but fortunately robotics researchers are
working out some nice methods for solving the interesting SLAM problem
that forces the robot to do both.
19.4 Coverage
The coverage problem has two basic versions: with a map and without a
map.
If the robot has a map of the environment, then the coverage problem be-
comes one of checking all the navigable spaces of the map until either the
object being sought is found or the whole map is covered. (This of course
assumes the object does not move; if it does, it is a different and harder prob-
lem having to do with pursuit that we’ll mention in Chapter 20.) Fortunately,
this is clearly a search problem, and therefore there are nice algorithms devel-
oped in computer science and AI to solve it, however slowly. For example,
there is the British Museum algorithm, an exhaustive brute-force search in-
spired by the problem of visiting all of the exhibits in that vast museum in
London.
H EURISTICS There are also various search algorithms that use clever heuristics, rules of
thumb that help guide and (hopefully) speed up the search. For example, if
the robot knows that the object is most likely to be in a corner, it will check
corners first, and so on.
So if the robot has a map, it uses that map to plan how to cover the space
carefully. But what if it does not have a map? In that case it has to move
in some systematic fashion and hope to find what it is looking for. This is
a hard problem since the environment may have a really complicated shape
and make it quite difficult to cover it all without knowing what it looks like.
In that case, mapping it first may be a better approach. But that takes time,
as we saw.
19.4 Coverage 231
There are various heuristics for coverage of unknown environments as

well. For example, one approach follows continuous boundaries, and an-
other spirals out from a starting point. Still another approach moves the
robot randomly, and given enough time and a closed environment, the robot
will cover it all. As you can see, which approach and what heuristics to use
depend on the robot’s sensors, task requirements (how quickly and/or thor-
oughly it has to get the job done), and whatever it can figure out about the
environment.
To Summarize
• Navigation is perhaps the very oldest and most studied problem in mobile
robotics.
• Navigation involves a great many subproblems, which include odometry,
localization, search and path planning, path optimization, and mapping.
Each of these is a separate field of study.
• Today’s robots are capable of navigating safely in the environment, but
there are still many elements of navigation that can be greatly improved,
so the areas above continue to be subjects of research and development.
Food for Thought
• Ants are tremendously effective at dead reckoning, finding their way in

the vast open spaces of the desert through odometry. They seem to be
much better at it than people, based on some studies that have been done.
Of course, no studies have actually pitted a human against an ant in the
desert, but comparable experiments find people lost and confused, and
ants on a nearly straight path back home.
• A great deal of research has gone into understanding how rats navigate,
since they are very good at it (so are ants, as we noted above, but their
brains are much harder to study, being so small). Researchers have put
rats into regular mazes and water mazes, and even mazes filled with milk.
One thing that is very interesting about rats that learn a maze well is that
they may run into a wall (if a new wall is stuck in by the experimenter)
even if they can see it, seeming to run based on their stored map rather
than their sensed data. What type of robot control system would produce
such behavior? Rats do it only once, though.
232 19 Going Places
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Navigation
• To learn about the search problem, check out the following excellent text-
books on computer algorithms: Introduction to Algorithms by Corman,
Leiserson, and Rivest, and AI: Artificial Intelligence, A Modern Approach
by Stuart Russell and Peter Norvig.
• To discover how mathematical statistics and probability are used to help

robots cope with uncertainty in navigation, brace yourself and ready Prob-
abilistic Robotics by Sebastian Thrun, Wolfram Burgard, and Dieter Fox
Building. Warning: this is not simple reading and requires a fair amount
of mathematics background.
• Autonomous Mobile Robots by Roland Siegwart and Illah Nourbakhsh goes

into more details about mobile robot kinematics and localization than we
did here and is a good resource.
• Here are a couple of textbooks on robot navigation you might want to

check out: Principles of Robot Motion: Theory, Algorithms, and Implementa-
tions by H. Choset, K. M. Lynch, S. Hutchinson, G. Kantor, W. Burgard,
L. E. Kavraki and S. Thrun and Planning Algorithms by Steven M. LaValle.
Note that these are the same as the texts we recommended for manipula-
tion, since the underlying planning problem is the same, it just gets harder
in higher dimensions.
20 Go, Team!
Group Robotics
What makes a group of robots a team? If you can control one robot,
what’s different about getting a group of them to do something together?
Why is it a hard thing to do well?
Controlling a group of robots is an interesting problem that presents a
whole new set of challenges compared with controlling one robot. These
include:
1. Inherently dynamic environment
2. Complex local and global interaction
3. Increased uncertainty
4. The need for coordination
5. The need for communication.
Having an environment full of robots creates a very fast-changing, dy-

namic world. As we already know, the more and the faster the environment
changes around the robot, the harder control can be. Multiple robots inher-
ently create a complex, dynamic environment because each moves around
and affects the environment and other robots around it. All this naturally
leads to increased complexity of the environment, and also more uncertainty
for everyone involved. A single robot has to contend with uncertainty in its
sensors, effectors, and any knowledge it has. A robot that is part of a group
or team also has to deal with uncertainty about the other robot’s state (Who is
that guy? Where is that guy?), actions (What is he doing?), intentions (What
is he going to do?), communication (What did he say? Did he actually say
234 20 Go, Team!
that? Who said that?), and plans (What does he plan to do? Does he have a
plan? Will that interfere with my plan?).
In a group situation, a robot has a whole new type of interaction to deal
with, one that involves other robots. Just like people, groups of robots must
coordinate their behavior effectively in order to get a job done as a team.
There are different kinds of "jobs" that teams can do, and different kinds of
C OORDINATION teams. Coordination has to do with arranging things in some kind of order
(ord and ordin mean “order” in Latin). Notice that the definition does not
C OOPERATION specify how that order is achieved. Cooperation, on the other hand, refers to
joint action with a mutual benefit. As you will see, some robot teams are
coordinated without really being cooperative, while others must cooperate
in order to coordinate team performance.
Quite a large area of active robotics research is concerned with the chal-
lenges of coordinating teams of robots. It is variously called group robotics,
social robotics, team robotics, swarm robotics, and multi-robot systems. Let’s
see what kinds of things it studies.
20.1 Benefits of Teamwork
To begin with, let’s consider why we would want to use more than one robot
at a time. Here are some potential benefits of using multiple robots:
• It takes two (or more): Some tasks, by their very nature, simply cannot
be done by a single robot. Perhaps the most popular example is pushing a
box (as in figure 20.1), a prototypical task that has to do with transporting
a large, heavy, awkward, or fragile object that cannot be effectively handled
by a single robot. Cooperative transportation is useful in construction and
habitat building, clearing barriers after a disaster, moving wounded people,
and many other real-world applications that require teamwork.
• Better, faster, cheaper: Some tasks do not require multiple robots, but
can be performed better by a team than by an individual robot. Foraging
F ORAGING is the most popular example. Foraging is the process of finding and collect-
ing objects (or information) from some specified area. Foraging has been
studied a great deal, because it is a prototype for a variety of real-world
applications of group robotics, such as locating and disabling land mines,
collectively distributing seeds or harvesting crops in agriculture, laying ca-
bles in construction, covering an area with sensors in surveillance, and so
on.
20.1 Benefits of Teamwork 235
Figure 20.1 A robot team collectively pushing a box; one robot (behind the box)
is the watcher, which steers the team toward the goal, while the two others are the
pushers. In this case, it takes three to tango, or at least to push this particular box.
(Photo courtesy of Dr. Brian Gerkey)
Foraging can be performed by a single robot, but if the area to be foraged

is large compared with the size of the robot, which is usually the case, the
job can be performed faster by a robot team. However, determining the right
number of robots for such a task is a tricky proposition, since too many will
get in each other’s way, and too few will not be as efficient as a larger team
might be. Determining the right team size is just one of the challenges of
group robotics. We’ll talk about others later in this chapter.
• Being everywhere at once: Some tasks require that a large area be mon-
itored so that robots can respond to an emergency event if/wherever it oc-
S ENSOR -A CTUATOR curs. Sensor networks are a popular example. Sensor-actuator networks are
N ETWORKS groups of mobile sensors in the environment that can communicate with
each other, usually through wireless radio, and can move about. By hav-
ing a large number of sensors in a building, an intruder or an emergency can
be detected in any area, and nearby robots can be sent to fix the problem.
Such networks are also used for habitat monitoring – for example, for tracking
animals and fish in the wild in order to conserve their environment, for de-
236 20 Go, Team!
Figure 20.2 Robot dogs playing soccer. (Photo courtesy of Messe Bremen)
tecting and measuring oil spills in the ocean, and for tracking the growth of
algae or some contaminant in the water supply. The more such robots there
are, the larger the area that can be monitored.
• Having nine lives: Having a team of robots can result in increased ro-
R OBUSTNESS bustness in performing a task. Robustness in robotics refers to the ability to
resist failure. In a team of robots, if one or a small number of robots fails, the
others can make up for it, so that the overall team is more robust than any
R EDUNDANCY individual alone would be. This type of robustness results from redundancy,
the repetition of capabilities on the team. Of course, not all teams are redun-
dant: if every member on a team is different, and thereby none can make up
for the failures of any others, then the team is not redundant, and is accord-
ingly is not very robust, since the whole team depends on each and every
member, and the failure of any one team member can break down the entire
team. This type of team organization is usually avoided everywhere, from
sports to armies.
20.2 Challenges of Teamwork

The above-listed advantages of robot teams have their price. Here are some
disadvantages that create the challenges of multi-robot control:
I NTERFERENCE • Get out of my way! Interference among robots on a team is the main chal-
lenge of group robotics. The more robots there are on a team, the more
chance there is that they will interfere with each other. Since the laws of
physics just won’t allow two or more robots to be in the same place at
the same time, there is always the potential for physical interference on a
team. Besides this basic type of interference which stems from the robot’s
20.2 Challenges of Teamwork 237
physical embodiment, there is also another kind: goal interference, which

has to do with robots’ goals conflicting. One robot can undo the work of
another, intentionally or otherwise, if their goals are conflicting. It turns
out that understanding, predicting, and managing interference is one of
the great challenges of working with more than one robot at a time.
• It’s my turn to talk! Many multi-robot tasks require the members of the
team to communicate in some way. The most popular means of communi-
cation is through wireless radio. A whole field of engineering deals with
communications and all the challenges that come with it. As in human
communication, there are interference issues here as well. If the robots
share a communication channel, they may need to take turns sending
messages and receiving them, to avoid "talking over each other", which
is the same as collisions in communication-space (rather than in physical
space, as in physical interference). Simply put, the more robots there are,
the more communication may be useful but also challenging.
• What’s going on? We have already seen that uncertainty is an inescapable

part of a robot’s life. The more robots are involved in a task, the more
uncertainty there can be. Similarly, the less each individual robot team
member knows, the harder it is to get the whole team to agree, in case
such agreement is needed (such as in moving to the next part of the task,
or recognizing when the task is over; not all tasks require this, as you will
see shortly). Some clever multi-robot controllers use multiple members
of a robot team to reduce uncertainty, but this does not come for free: it
requires clever algorithms and communication among robots.
• Two for the price of one? Even with discounts for bulk, more robots are
always costlier than fewer, and more robots have more hardware failures
and maintenance requirements to worry about as well. Cost is a practical
issue that influences real-life decisions in robotics as it does everywhere
else.
The challenges of group robotics are what makes programming a team of

robots a new and different problem from programming one robot. Before we
look into how to program them, let’s see what kinds of robot teams there are.
238 20 Go, Team!
Figure 20.3 A wheeled robot and a six-legged robot cooperatively pushing a box.
(Photo courtesy of Dr. Lynne Parker)
20.3 Types of Groups and Teams
How would you program a bunch of robots to play soccer (such as those
in figure 20.2)?
Would you program each robot to act is if it were alone – to chase the ball
and try to take it, run for the goal and shoot, and run back and protect its
own goal or goalie? What would happen if you did? (What would happen
if that were how people played soccer?) It would be a complete mess! All
robots would run around and into each other, trying to play both offense
and defense, and getting into each other’s way. Soccer, like most other group
D IVISION OF L ABOR activities, requires teamwork and some type of division of labor or role assign-
R OLE A SSIGNMENT ment, giving each member of the team a job so each knows what to do to help
the team as a whole without getting in the way of the others.
Do all group tasks require division of labor? What about foraging?
Could you just have a group of identical robots all doing foraging and
staying out of each other’s way?
Definitely. This shows us that there are different types of tasks and differ-
ent types of collective multi-robot systems.
First, teams can be most easily divided into two types, based on the indi-
viduals that constitute them:
H OMOGENEOUS • Homogeneous teams are those with identical, and therefore interchangeable,
T EAMS members. Members may be identical in their form (such as all having four
wheels and a ring of sonars) and/or in their function (such as all being
20.3 Types of Groups and Teams 239
Figure 20.4 A highly effective foraging group of merely coexisting robots. (Photo
courtesy of Dr. Chris Melhuish)
able to find and pick up objects and send messages to other team mem-
bers). Homogeneous teams can be coordinated with simple mechanisms,
and in some cases require no active, intentional cooperation to achieve ef-
fective group behavior (remember emergent flocking from Chapter 18?)
.
H ETEROGENEOUS • Heterogeneous teams are those with different, non-interchangeable mem-

T EAMS bers. Members may differ in form (some have four wheels, while others
have two wheels and a caster) and/or in function (some play offense by
chasing the ball while some play defense by getting in the way of oppo-
nents). Heterogeneous teams typically require active cooperation in order
to produce coordinated behavior.
Figures 20.1 and 20.3 show examples of real-world robot teams pushing
boxes. Both teams are heterogeneous, but one is heterogeneous only in
form (different bodies, same function), while the other is heterogeneous
in both form and function. Can you tell which is which? The first, a three-
member team, features members with different roles: two pushers and a
watcher, so it is heterogeneous in both form and function. The second, a
two-member team, features two members with different from (wheels vs.
legs) but identical function (pushing the box).
The next way of classifying robot teams is based on the type of coordina-
tion strategy they use:
Merely coexisting: In this approach, multiple robots work on the same

task in a shared environment, but do not communicate or, in some cases,
240 20 Go, Team!
even recognize each other. They merely treat each other as obstacles. Such
systems require no algorithms for coordination or communication, but as the
number of robots grows and the environment shrinks, interference increases,
quickly removing any benefit of having a group. Foraging, construction, and
other tasks can be achieved with this approach as long as the team size is
carefully designed to suit the task and environment so that interference is
minimized.
Let’s take a closer look, using the foraging example: We have a group of
foraging robots whose job is to look over a field for scattered objects, pick
them up, and bring them back to a particular deposit location. While they do
that task, they must avoid collisions with obstacles, including other robots.
You can see how the more robots are doing the task, the more potential in-
terference there is between them, as long as they are in a bounded, limited
amount of space (which is reasonable to expect, since foraging in the vastness
of outer space is not currently a major robotics application). In this approach,
the robots do not help each other or even recognize each other. Therefore, to
make this work, there is a carefully designed balance among the task pa-
rameters (including the size of the space and the number of objects being
collected), the number of robots, their physical size relative to the environ-
ment, their sensor range, and their uncertainty properties. As you can see,
this takes some work. Figure 20.4 shows a group of robots that very effec-
tively collected objects into a single pile through merely coexisting. We’ll see
later how they did it.
Loosely coupled: In this approach, the robots recognize each other as

members of a group and may even use simple coordination, such as moving
away from each other to make space and minimize interference. However,
they do not depend on each other for completing the task, so members of the
group can be removed without influencing the behavior of the others. Such
teams are robust, but difficult to coordinate to do precise tasks. Foraging,
herding, distributed mapping, and related group tasks are well suited for
this approach.
Back to our foraging example. Now, instead of treating each other as ob-
stacles, the robots can actually react to each other in more interesting ways.
For example, a robot that has found no objects can follow another robot that
is carrying an object, in hopes that it will lead it toward more objects. Alter-
natively, a robot can avoid other robots under the assumption that it should
go where others haven’t been yet, and so find as yet undiscovered objects.
Figure 20.5 The Nerd Herd, one of the first multi-robot teams, flocking around the
lab. (Photo courtesy of the author)
The robots can also flock (as shown in figure 20.5) or form lanes as they are
heading to deposit the objects, in order to minimize interference.
Tightly Coupled: In this approach, the robots cooperate on a precise task,

usually by using communication, turn-taking, and other means of tight co-
ordination. They depend on each other, which gives the system improved
group performance, but also less robustness through redundancy. Remov-
ing team members makes the team performance suffer. Playing soccer, mov-
ing in formation, and transporting objects are done well through the use of
tightly coupled teams.
Returning to foraging yet again: A tightly coupled foraging team would
likely use a global plan in which each member of the team goes to a specific
area of the environment to explore. This is convenient when a map of the
area is available. If no map is given, then the robots may use collaborative
SLAM to build one (remember SLAM from Chapter 19?) as they coordinate
their exploration and object collection.
20.4 Communication
Let’s talk about talking. Members of a team usually (but not always) need to
communicate with each other in order to achieve the collective goal(s).
242 20 Go, Team!
Why would robots communicate?
Here are some good reasons:
1. Improving perception: Robots can sense only so much; by using commu-

nication they can know more about the world, without having to sense it
directly, by exchanging information with other robots.
2. Synchronizing action: Because robots on a team usually can’t perceive all

others on the team instantaneously, if they all want to do (or not do, or
stop doing) something together at the same time, they need to communi-
cate, or signal, to each other.
3. Enabling coordination and negotiation: Communication is not required

for coordinated behavior, as you will see below. However, in some cases
and for some tasks, it helps a great deal for robots to be able to cooperate
and negotiate in order to get things done right.
The reasons to communicate give us some ideas as to what robots might

communicate. Consider the options through the example of foraging:
• Nothing: That’s what happens in merely coexisting, and it can work very
well, as you will soon see.
• Task-related state: The location of the objects, the number of recently seen
robots, etc.
• Individual state: ID number, energy level, number of objects collected,

etc.
• Environment state: Blocked exits and paths, dangerous conditions, new-

found shortcuts, etc.
• Goal(s): Direction to the nearest object, etc.
• Intentions: I’m going this way because I’ve found objects there before; I’m
not going that way because there are too many others there; etc.
And how can robots communicate?
We have already mentioned wireless radio communication, but that’s not

nearly the whole story. Consider the myriad ways in which people communi-
cate information: we gesticulate, we shout to a crowd, we whisper to a friend,
we post signs, we email, we leave phone messages, we write letters, cards,

papers, and books, and so on. Interestingly, most of those forms of commu-
nication can be used by robots. Let’s see what options there are for robots
who want to communicate:
E XPLICIT Explicit communication, sometimes also called intentional communication, re-
C OMMUNICATION quires an individual to purposefully behave in a way that will convey a mes-
I NTENTIONAL
sage. In robotics that typically involves sending a message over the commu-
C OMMUNICATION
nication channel, these days wireless radio.
Because wireless communication is ubiquitous (found everywhere), multi-
robot communication has gotten a lot easier. Robotics can now rather easily
B ROADCAST use broadcast communication, sending a message to everyone on the communi-
C OMMUNICATION cation channel, or peer-to-peer communication, sending a message to a selected
P EER - TO -P EER
recipient.
C OMMUNICATION
P UBLISH -S UBSCRIBE Publish-subscribe communication is much like using an email list or a news
C OMMUNICATION group: a select group of recipients interested in a particular topic signs up
for the list, and only those on the list receive messages.
Explicit communication involves a cost because it requires both hardware
and software. For any given team task, the designer must think hard about
whether communication is needed at all, and if it is, what its range and type
(among those listed above) should be, what the information content should
be, and what performance level is needed and can be expected from the com-
munication channel. (This last is another reality of life: all communication
channels are imperfect, so some messages and parts of messages are lost or
corrupted by noise.)
Thus, if the task requires the robots to be able to negotiate one-on-one and
in a short amount of time, that presents a very strict set of communication
requirements which are fundamentally different from those that may be in-
volved in a task that can be achieved with occasional broadcast messages to
all team members.
The forms of communication described so far have all been explicit, used
with the purpose of communicating. However, another very powerful and
effective type of communication exists:
I MPLICIT Implicit communication involves an individual leaving information in the
C OMMUNICATION environment, thereby communicating with others without using any explicit
communication channels.
S TIGMERGY Stigmergy is the form of communication in which information is conveyed
through changing the environment. Ant trails are a perfect example: as ants
P HEROMONE move about, they leave behind small amounts of pheromone, a fragrant hor-
mone that can be detected by others of the same ant species. Ants tend to
244 20 Go, Team!
follow other ants’ pheromone trails, and so the more ants go down a cer-
tain path, the stronger the pheromone trail becomes and the more ants are
recruited. This is a great example of positive feedback, which results from
P OSITIVE F EEDBACK the feedback control we learned about in Chapter 10. In positive feedback, the
more something happens, the more it "feeds" on itself, and so the more it hap-
pens, and so on. Examples of positive feedback behavior include stampedes,
lynch mobs, grazing patterns of cattle, termite nests, and of course ant trails.
A MPLIFYING This type of feedback is said to be amplifying because it makes some signal
N EGATIVE F EEDBACK or behavior stronger. Positive feedback is the opposite of negative feedback,
R EGULATORY which is regulatory, making sure that a system does not go out of control,
but rather that it stays close to the set point, just as we discussed in detail in
Chapter 10.
So, positive feedback can result not only from explicit but also from im-
plicit communication such as stigmergy, communication through sensing the
effects of others in the environment. Robotics researchers have shown how a
team of merely coexisting robots that has no way of sensing each other as any
different from obstacles can reliably build a barrier using pucks, or collect all
pucks within a large area into a single group, purely through stigmergy. Can
you figure out how that could be done?
How can you get a team of robots that can’t detect each other but can
only tell if they have bumped into something, and that can push a puck,
go forward, turn, and back up, to reliably collect all pucks in a single
"pile" (the quotes are there because the pile is in 2D, on the flat floor)?
Assume that the environment looks just like that shown in figure 20.4, be-
cause those really are snapshots of that very system.
As you have learned already, the form of the robot (its body) must be well
matched to its function (its task). The roboticists who designed the puck-collecti-
ng robots, Ralph Beckers and Owen Holland, really embraced that principle
and used it very ingeniously. They designed a very simple robot, shown in
figure 20.6, which had a scoop that served a dual purpose: (1) it collected
pucks, because as the robot moved forward, the scoop scooped pucks up
and pushed them along; and (2) it detected collisions, because when it was
pushed in by the weight of the puck(s), it triggered a 1-bit contact sensor (a
simple switch). Using that clever physical robot design, Holland and Beckers
then developed a suitably clever controller:
Figure 20.6 A close-up of the cleverly designed puck-collecting robot, part of the
swarm that piled up pucks using a single 1-bit contact sensor and no communication.
(Photo courtesy of Dr. Chris Melhuish)
When hard contact detected

stop and back up, then turn and go
When soft contact detected

turn and keep going
Amazingly, that was it! The robot’s scoop was calibrated (remember cali-
bration from Chapter 8) very carefully not to be too sensitive, so that it could
collect quite a few pucks (six to eight) and push them along. Only when
it encountered a stronger barrier or heavier weight (more than about six to
eight pucks) was its contact switch triggered. Then it stopped, backed up,
turned, and moved on, just as its simple minimalist controller told it to do.
Consider what that looks like when you put the robot in a room full of
pucks. The robot will move around and, by chance, scoop up pucks. It won’t
even be able to distinguish pucks from other objects in the environment. For-
tunately, there is nothing else in the environment besides pucks, walls, and
other robots. We’ll get to those things in a bit.
246 20 Go, Team!
When the robot collects more than eight or so pucks in its scoop, the pucks
will be so heavy that its contact sensor will be triggered, and it will stop and
back up. And what will that do? It will cause it to leave the pile of pucks
there and go on without it! In this way, it will keep making little piles of
pucks, and whenever it happens upon a pile of pucks and is pushing another
pile, it will drop its new pile next to the old, which will make a larger pile.
If you let the robot do this for a while, it will end up with quite a few small
piles. If you let it run a while longer, it will produce fewer, larger piles.
Eventually, after quite a long long time, it will produce a single large pile.
And of course if you put multiple robots in the same area with the same
controller, the overall task will be achieved faster.
What happens when the robot runs into the wall?
Well, if the wall is rigid, the robot will stop and back up, but also leave any
pucks it has next to the wall, which would not be desirable. The designers
cleverly made walls out of flexible fabric, so that the scoop would contact
them but the robot would not stop; it would only turn away, keeping the
pucks (at least most of the time).
What happens when the robot runs into another robot?
If the contact is soft, it goes on with its pucks, and if the contact is hard,
which is the case when robots are heading right into each other, both will
back up and move away. And notice this clever side-effect: any pucks they
are pushing will end up in a single joined pile.
There are some great videos of this system, which was implemented and
validated several times. See the end of this Chapter for pointers to more
information.
We can learn a lot from this clever system. First, it is a superb example
of effectively using both the robot’s physical design and its simple controller
to exploit the dynamics of the robots’ interaction with each other and the
environment in order to accomplish their task. In this system, robots cannot
detect each other or communicate with each other. This means they cannot
coordinate where the pile of pucks will end up, since that is the result of the
system dynamics and is different each time the system runs. On the other
hand, it’s hard to imagine a simpler and more elegant controller and design
for this rather complex task.
The point of the above example is not to say that robots should always use
only a single contact sensor and no communication to do collective foraging.
Instead, the point is to show that there are a great many truly innovative
and powerful solutions to problems when one thinks broadly and gives up
assumptions that may be incorrect, such as "The robots must communicate
to get the job done well" or "The robots must recognized each other in order
to cooperate." Let’s talk about that last one a bit more.
20.4.1 Kin Recognition
Do you think it would be easier for robots in the example above if they
could distinguish other robots from pucks and walls?
The answer is yes, if you wanted to write a more complex controller that
took advantage of that information. However, that requires more complex
sensing and sensory data processing; as you have already seen in Chap-
ters 8 and 9, object recognition is a hard problem. Still, being able to rec-
ognize ”others like me” is a very useful thing, and is something that most
animals do, in multiple ways, through the use of multiple sensors ranging
from pheromones (yes, those again), to sound, to vision.
K IN R ECOGNITION In nature, kin recognition refers to being able to recognize members of the
immediate family, those that share genetic material with the individual. This
ability is directly useful in deciding with whom to share food, to whom to
signal about predators, and other such potentially "expensive" activities.
In group robotics, kin recognition can refer to something as simple as dis-
tinguishing another robot from other objects in the environment, to some-
thing as complicated as recognizing one’s team members, as in a game of
robot soccer. In all cases, kin recognition is a very useful ability, typically
worth the sensory and computational overhead it may involve.
Without kin recognition, the types of cooperation that can be achieved are
greatly diminished. Kin recognition does not necessarily, or even typically,
involve recognizing the identities of all others. Even without identities, so-
phisticated coordination and cooperation is possible. For example, a team
D OMINANCE of robots, just like a group of hermit crabs, can establish a dominance hierar-
H IERARCHY chy, more informally known as a pecking order. A pecking order helps to give
structure and order to a group, so there is less interference.
248 20 Go, Team!
20.5 Getting a Team to Play Together
So how can we control a group of robots?
There are basically two options: centralized control and distributed con-
trol, with some compromises thrown in. Let’s take a look.
20.5.1 I’m the Boss: Centralized Control

C ENTRALIZED In centralized control, a single, centralized controller takes the information
C ONTROL from and/or about all of the robots on the team, thinks about it as long as it
takes (and it could take a while if the team is any size above three or so and is
doing anything not too trivial in the real world), and then sends commands
to all of the robots as to what to do.
No doubt you can readily see the many problems with that idea. It requires
a lot of information to come together in a single place; it requires global com-
munication; it is slow and gets slower as the team gets larger; and it is not
robust because the centralized controller is a bottleneck in the system: if it
fails, the system fails.
Why would we even consider this type of robot team organization?
Because it has one advantage: centralized control allows the team to com-
pute the optimal solution to a given problem it is facing. Putting it more
simply: too many cooks can spoil the broth. In an ideal world, if the cen-
tralized planner has all the needed correct and complete information and is
given enough time to think and make decisions, then the team as a whole
can be given perfect instructions.
Does that sound familiar? It should: it’s the basic idea of centralized plan-
ning which you saw in Chapter 13, so you already know the pros and cons
of this approach.
20.5.2 Work It Out as a Team: Distributed Control

D ISTRIBUTED In distributed control, there is no single, centralized focus of control; instead,
C ONTROL it is spread over multiple or even all members of the team. Typically, each
robot uses its own controller to decide what to do.
There are many advantages to this approach: no information needs to be
gathered centrally, so there are no bottlenecks and communication can be
minimized or sometimes avoided. As a result, distributed control works well
20.5 Getting a Team to Play Together 249
with large teams, and does not slow down in performance as the team grows
or changes in size.
But as usual, there is no such thing as a free lunch: distributed control
brings its own set of challenges. The largest is the issue of coordination:
distributed control requires that the desired group-level collective behavior
be produced in a decentralized, non-planned fashion from the interactions
of the individuals. This means that the job of designing the individual, local
behaviors for each robot is more challenging, because it needs to work well
with the others in order to produce the desired group behavior.
Given a set of individuals, whether they be bugs or people or robots, it is
difficult to predict what they will do together as a group, even if we know
what rules or behaviors they execute. This is because what happens lo-
cally, between two or three individuals, can have an impact on a large group
through positive feedback and other mechanisms of propagation, passing on
the effects of actions.
Various sciences, including physics, chemistry, social science, meteorology,
and astronomy, have studied such collective behavior of multiple individu-
als for a long time, and are still at it. The components being studied range
from atoms to amino acids to stars and planets to people on the crowded
subway. If the individual components in the system being studied are rela-
tively simple, analysis can be somewhat easier, but there are many examples
of very simple entities creating very complex collective behaviors (see the
end of the chapter for more information). The problem often gets easier if
there are very many components, because there exist good statistical tools
in mathematics that can be used. In robotics, we are stuck with the hardest
of situations, because multi-robot systems have pretty complicated compo-
nents (even simple robots are not very simple to predict) and not too many
of them (until nano robots and smart dust are invented; see Chapter 22 for
more on that).
Going from local rules to global behavior is a hard problem; it is hard to
predict what a lot of robots with particular controllers will do when you let
I NVERSE P ROBLEM them run. Harder still is the so-called inverse problem, which involves going
from the global behavior to the local rules. This means it is harder still to
figure out what controllers to put on each robot in order to get the group
as a whole to do a particular desired behavior (such as puck collection or
competitive soccer).
Distributed robot control requires the designer to solve the inverse prob-
lem: to figure out what each robot’s controller should be, whether they are
250 20 Go, Team!
all the same, or different, i.e., whether the team is homogeneous or heteroge-
neous, as we saw above, in order to produce the desired group behavior.
20.6 Architectures for Multi-Robot Control
Whether you are controlling a single robot as a centralized controller for a

team or a group of robots individually as parts of a team, you are still left
with the usual set of control approaches we have already covered: delibera-
tive, reactive, hybrid, and behavior-based control.
Based on all you know, can you figure out which of the control ap-
proaches is good for what type of team control?
Your intuition is probably right:
• Deliberative control is well suited for the centralized control approach.

The single controller performs the standard SPA loop: gathers the sensory
data, uses it all to make a plan for all of the robots, sends the plan to each
robot, and each robot executes its part.
• Reactive control is well suited for implementing the distributed control

approach. Each robot executes its own controller, and can communicate
and cooperate with others as needed. The group-level behavior emerges
from the interaction of the individuals; in fact, most of the emergent be-
haviors seen in robotics result from this type of system: a team of robots
controlled using distributed reactive control.
• Hybrid control is also well suited for the centralized control approach,
but can be used in a distributed fashion, too. The centralized controller
performs the SPA loop, individual robots monitor their sensors, and up-
date the planner with any changes, so that a new plan can be generated
when needed. Each robot can run its own hybrid controller, but it needs
information from all the others to plan, and synchronizing the plans is
hard.
• Behavior-based control is well suited for implementing the distributed

control approach. Each robot behaves according to its own, local behavior-
based controller; it can learn over time and display adaptive behavior as
a result, so the group-level behavior can also be improved and optimized.
20.6.1 Pecking Orders: Hierarchies

It is easiest to imagine how homogeneous robot teams can be controlled with
the above control approaches. However, heterogeneous teams with mem-
bers featuring different abilities and different amounts of control can also be
implemented with any of the above.
H IERARCHIES Hierarchies are groups or organizations that are ordered by power; the term
comes from the Greek word hierarche meaning “high priest.’ In robot hierar-
chies, different robots have different amounts of control over others in the
team.
As you might guess, two basic kinds of hierarchies exist: fixed (static) and
F IXED H IERARCHIES dynamic. Fixed hierarchies are determined once and do not change; they are
like royal families, where the order of power is determined by heredity. In
D YNAMIC contrast, dynamic hierarchies are based on some quality that can keep chang-
H IERARCHIES ing; they are like dominance hierarchies in some animals that are based on
whichever is the largest, strongest, and wins the most battles.
Going back to our different approaches to robot control, here are some
options for implementing hierarchies in robot teams:
• Fixed hierarchies can be generated by a planner within a deliberative or
hybrid system.
• Dynamic, changing, adaptive hierarchies can be formed by behavior-based
systems.
• Reactive distributed multi-robot systems can form hierarchies either by
preprogramming or dynamically (e.g., based on size, color, ID number).
Coordinating a group of robots, whether it be a swarm of simple ones or
a small group of complex ones, or anything in between, is a major challenge
and an active area of robotics research. It is also a highly promising one, since
its results are being aimed directly at real-world applications that include
autonomous driving and multi-car convoying for safer highway transporta-
tion, at sensor network deployment in the ocean for detecting and cleaning
up toxic spills, on the ground for fire fighting, for terraforming and habitat
construction in space applications, for land-mine detection, farming, recon-
naissance and surveillance, and many more. So jump in and join the team!
To Summarize
• Controlling a group of robots is a problem different from controlling a sin-
gle robot; it involves considerable interaction dynamics and interference.
252 20 Go, Team!
• Robot groups and teams can be homogeneous or heterogeneous.
• Robot teams can be controlled centrally or in a distributed fashion.
• Communication is not a necessary, but is often a useful, component of

group robotics.
• Robot teams can be used in a wide variety of real-world applications.
Food for Thought
• The ways in which robots in a group interfere is not very different from
the way people in a group interfere with each other. Some roboticists
are interested in studying how people resolve conflict and collaborate,
to see if some of those methods can be used to make robots better team
players as well. Unlike people, robots don’t tend to be self-interested or
to show a vast variety of personal differences, unless they are specifically
programmed that way.
• It is theoretically impossible to produce totally predictable group behav-

ior in multi-robot systems. In fact, it is a lost cause to attempt to prove or
guarantee precisely where each robot will be and what it will do after the
system is running. Fortunately, that does not mean that multi-robot sys-
tem behavior is random. Far from it; we can program our robots so that
it is possible to characterize, even prove, what the behavior of the group
will be. The important fact of life for multi-robot systems is that we can
know a great deal about what the group as a whole will do, but we cannot
know exactly and precisely what each individual in such a group will do.
• Imagine that the robots in the group can change their behavior over time,
by adapting and learning. In Chapter 21 we will study what and how
robots can learn. Learning in a team of robots makes coordination even
more complex, but it also makes the system more interesting and poten-
tially more robust and useful.
• The ideas of positive and negative feedback are used in common English
terminology to refer to giving praise/reward and punishment. Actually,
that is an example of loose use of terms; we’ll talk about how positive and
negative feedback relate to reward and punishment in Chapter 21, which
deals with robot learning.
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Group_Robotics
• A team of twenty very simple wheeled robots called the Nerd Herd was
one of the very first to ever perform group behaviors, such as following,
flocking, herding, aggregating, dispersing, and even parking neatly in a
row. That was back in the early 1990s at the MIT Mobot Lab, which you
already read about in Chapter 5. You can see videos of the Nerd Herd at
http://robotics.usc.edu/robotmovies. Here are a couple of papers about
how the Nerd Herd was controlled:
– Matarić, Maja J. (1995), “Designing and Understanding Adaptive Group

Behavior”, Adaptive Behavior 4(1), December, 51-90.
– Matarić, Maja J. (1995), “From Local Interactions to Collective Intelli-
gence”, The Biology and Technology of Intelligent Autonomous Agents, L.
Steels, ed., NATO ASI Series F, 144, Springer-Verlag, 275-295.
• Here are some of the papers published by the researchers who did the
very elegant work on collective sorting we described in this chapter:
– Beckers, R., Holland, O. E. and Deneubourg J_L (1994). “From Lo-

cal Actions to Global Tasks: Stigmergy in Collective Robotics,” in R.
Brooks and P. Maes eds. Artificial Life IV, Cambridge, Mass.: MIT
Press, 181-9.
– Deneubourg, J. L., Goss, S., Franks, N. R., Sendova-Franks, A., De-
train, C., and Chretien, L. (1990). “The Dynamics of Collective Sorting:
Robot-like Ants and Ant-like Robots.” In Meyer, J-A, and Wilson, S.,
eds, Simulation of Adaptive Behaviour: from animals to animats, Cam-
bridge, Mass.: MIT Press, 356-65.
• For a look at how complex collective behavior in insects can be, read the
impressive book Ants by E. O. Wilson.
• Ethology is the science of studying animals in their natural habitat; the

term comes from the Greek ethos meaning “custom.” Textbooks on ethol-
ogy provide fascinating background about animal social and collective
behavior, which has served as inspiration for many roboticists. Etholo-
gists who studied insect behavior and other topics relevant to what we
254 20 Go, Team!
discussed in this chapter include Niko Tinbergen, Konrad Lorenz, David

Macfarland, Frans de Waal, and E. O. Wilson. You can find textbooks by
each of them.
21 Things Keep Getting Better
Learning
L EARNING Learning, the ability to acquire new knowledge or skills and improve one’s
performance, is one of the distinguishing features of intelligence, human or
robotic.
How can a robot change its programming and get better at what it does?
There are, it turns out, quite a lot of ways. But first, let’s consider what
kinds of things a robot can learn.
A robot can learn about itself. Remember that just because a robot can do
something, that does not mean it knows it can do it or how well it can do it.
So, it is quite useful for a robot to learn how its sensors tend to work and fail
(e.g., "If I sense something at location (x,y), where is it really likely to be?"),
how accurate its actuators are (e.g., "If I want to go forward by 10 cm, how
far am I really likely to go?"), what behavior it tends to perform over time
(e.g., "I seem to run into obstacles a lot."), how effective it is at achieving its
goal ("I can’t seem to achieve my goal for a very long time."), and so on. Such
knowledge is not easily pre-programmed, and may vary over time, so it is
best learned.
Notice that this gives us a hint about what is worth learning. But we’ll talk
about that a bit later in this chapter.
Besides learning about itself, a robot can learn about its environment. The
most popular example is learning maps, which we talked about in Chap-
ter 19. Even without learning a map, a robot can learn paths to a goal, such
as learning a way through a maze, as a sequence of moves. The robot can
also learn where unsafe areas of the environment are, and much more, based
on what about its environment matters for performing its task. By the way,
that is another hint about what is worth learning.
256 21 Things Keep Getting Better
As we saw in Chapter 20, for some robots the environment includes other
robots. In that case, a robot can learn about other robots, things such as how
many of them there are, what kinds, how they tend to behave, how well the
team works together, and what works well for the team and what does not.
Let’s return to the basics: why should a robot bother to learn? Here are
some good reasons to consider:
• Learning can enable the robot to perform its task better. This is important
because no robot controller is perfect, nor is any robot programmer.
• Learning can help the robot adapt to changes in its environment and/or
its task. This is hard to pre-program if changes cannot be anticipated by
the programmer.
• Learning can simplify the programming work for the designer of the robo-
t’s controller. Some things are just too tedious or too difficult to program
by hand, but can be learned by the robot itself, as you will see shortly.
Now that we have the what and why of robot learning covered, we can
turn to the real question: how do we get a robot to learn?
21.1 Reinforcement Learning

R EINFORCEMENT Reinforcement learning is a popular approach for getting robots to learn based
L EARNING on feedback received from the environment. The basic idea is inspired by
learning in nature, the way animals (including people) learn. Reinforcement
learning involves trying different things and seeing what happens; if good
things happen, we tend to do the behavior again, and if bad things happen,
we tend to avoid it.
This basic process turns out to be a remarkably versatile tool for learn-
ing. It allows robots to learn what to do and not to do in various situations.
Consider a typical reactive controller that tells the robot how to react under
different sensory inputs. By using reinforcement learning, you can have the
robot learn such a controller instead of programming it in. Well, close to it,
anyway.
Consider the robot’s input space (remember what that is from Chapter 3)
as the set of all possible situations. Now consider the set of all of its possible
actions. If the robot knows what action to perform in each state, it will have
a complete controller. But suppose we would prefer not to pre-program the
controller, but to have the robot learn it on its own.
How could a robot learn to match the right states with the right actions?
By trial and error! In reinforcement learning, the robot tries different ac-
tions (in fact, all of the actions it has at its disposal) in all states it gets into. It
keeps track of what happens each time in some type of representation (usu-
ally some type of a table or matrix). In order to learn what action is best in
each state, it has to try everything; in machine learning (for robots, software
agents, and other kinds of programs), the process of trying all possible state-
E XPLORATION action combinations is called exploration. Exploration is important, because
until the robot has tried all possible actions in all states, it can’t be sure that
it has found the best action for each state.
Once the robot has tried all combinations of states and actions, it knows
what is good and what is bad to do relative to its task and goals. Wouldn’t it
be great if it were done now? Unfortunately, it’s not. Here is why:
• Things are not what they seem: If there is any error in the robot’s sensing
of the current state or in its execution of its actions, the result of trying any
state-action combination may be incorrectly learned. Uncertainty strikes
again.
• Things may change: If the environment or task changes while the robot is
learning, what has it learned may be invalid.
Both of those reasons make it important for the robot to keep checking
what it has learned, not only in reinforcement learning but in any learning
approach. In reinforcement learning in particular, the robot has to keep ex-
ploring instead of always exploiting what it has learned. In machine learning,
E XPLOITATION the process of using what has been learned is called exploitation. We tend
to think of exploitation as some kind of unfair use, but it really means us-
ing something to its fullest advantage. If a robot purely exploits what it has
learned without exploring further, it will not be able to adapt to changes
or continue to learn. The trade-off between constantly learning (at the cost
of doing things less than perfectly) and using what is known to work well
E XPLORATION VS . (at the cost of missing out on further improvements) is called exploration vs.
E XPLOITATION exploitation.
Ok, we know that the robot tries different things and remembers how they
turn out. But how does it do that, and exactly what does it remember?
Imagine that the robot has a table, like the one shown in figure 21.1, that
has all possible states as the rows, and all possible actions as the columns. Ini-
tially this table is empty or all of its entries are the same, say 0. After the robot
Figure 21.1 A policy table for a simple learning robot.
tries a particular state-action combination and sees what happens, it updates

that entry in the table. An intuitive way to think about this is to imagine
that the value of a state-action entry in the table grows if good things hap-
pen and shrinks if bad things happen. To keep this mathematically correct,
the values in the table need to be normalized and updated in a consistent
way. Entries can be updated in a variety of different ways, depending on
what reinforcement learning algorithm the robot is using; two popular ones
are temporal differencing and Q-learning. There are numerous papers and
textbooks written about robot learning that describe these algorithms, so we
won’t spend much time on them here; but you can easily look them up. The
most important thing to realize is that the robot uses reinforcement learning
to learn state-action combinations. The complete state-action table is called
C ONTROL P OLICY a control policy; these are reactive rules that program a robot for a particular
goal.
Every policy is specific to a particular goal, just as any controller is specific
to a particular task. If the goal/task changes, so must the policy/controller.
VALUE F UNCTION As an alternative to a policy, the robot could learn the value function, the
value of being in each state relative to the goal. In practice, this is the same
as learning a control policy, so we won’t discuss it further.
For reinforcement learning to work on a robot, the robot has to be able to

evaluate how well a state-action pairing went, what the outcome was of a
trial. For example, if the robot is learning how to walk, it has to be able to
measure if it has moved forward after trying some state-action combination.
P OSITIVE F EEDBACK If it has, that state-action combination receives positive feedback, more simply
R EWARD called reward. Conversely, if the robot moved backward, it will receive neg-
N EGATIVE F EEDBACK ative feedback, more simply called punishment. (It is important not to confuse
P UNISHMENT
these notions of feedback in reinforcement learning with those we discussed
in Chapter 10; as they are not the same.)
Notice that feedback can come in different ways, and usually simple good
or bad is not enough. In the example above, the robot will learn about going
forward and backward, but it will not learn what to do if the outcome of a
state-action pair leaves it stopped.
This is just a simple example that illustrates that setting up the right kind
of feedback mechanism is one of the major challenges of reinforcement learn-
ing.
How does the robot know exactly how much to reward or punish a par-
ticular outcome?
This turns out to be really tricky.
Reinforcement learning works quite well under some specific conditions:
• The problem is not too huge: If there are so many states that it would prac-
tically take the robot days to try them all and learn, then the approach may
not be desirable or practical; it all depends on how hard it is to program
the system by hand in comparison.
• The outcome is clear: The result of tested state-action combinations are

known soon after the trial and can be sensed by the robot.
Let’s take another example. Consider a robot learning a maze. Taking

the right turn early on can make all the difference in eventually finding the
way out, but at the time the robot takes the turn, it will not know whether
that is a good thing to do or not. In that case, the robot has to remember
the sequence of state-action pairs it took, and then send back the reward or
punishment to each of them in order to learn. With uncertainty, this can get
T EMPORAL C REDIT quite complicated and messy, but not impossible. The general problem of
A SSIGNMENT assigning the credit or blame to actions taken over time is called temporal
credit assignment.
The reinforcement learning problem can also be made more difficult in a

multi-robot environment, where it is not always easy to tell which robot’s
action brought about a good (or bad) outcome. A robot may do an action in
a state that does nothing, but if another robot happens to do just the right
next thing next to it, either may assume the desirable outcome is due to itself
S PATIAL C REDIT and not the other. This problem is called spatial credit assignment.
A SSIGNMENT Reinforcement learning has been used with great success in non-robotics
applications; for example, the world backgammon champion was a com-
puter program that learned the game by playing against itself and using re-
inforcement learning. Of course in robotics, learning is much messier due to
uncertainty, dynamically changing environments, and hidden/partially ob-
servable state (remember Chapter 3). Nevertheless, robots have successfully
used variations of reinforcement learning to learn to forage, navigate mazes,
juggle, pass and score in soccer, and many other activities.
21.2 Supervised Learning
U NSUPERVISED Reinforcement learning is an example of unsupervised learning, meaning that

L EARNING there is no external supervisor or teacher who tells the robot what to do.
Instead, the robot learns from the outcomes of its own actions. The natural
S UPERVISED L EARNING opposite of unsupervised learning is supervised learning, which involves an
external teacher who provides the answer or at least tells the robot what it
did wrong.
N EURAL N ETWORK Neural network learning is a very popular approach for supervised machine
L EARNING learning for robots and other non-embodied machines. A wide variety of
C ONNECTIONIST neural network learning algorithms exists, all collectively called connectionist
L EARNING learning, because typically what is being learned is the weights on the vari-
ous connections in the networks (how strongly nodes are connected), or in
some cases the structure of the network itself (what is connected to what;
S TATISTICAL N EURAL how many nodes there are). There exists a growing field of statistical neu-
N ETWORKS ral networks in which the learning is done with very mathematically formal
techniques from statistics and probability. In fact, the use of statistics and
probability in robot control and learning is widespread; see the end of the
chapter for pointers to more literature on the topic.
Reinforcement learning and neural networks learning are both only dis-
tantly related to the way learning works in nature. They are not meant to
be careful models of natural systems, but practical algorithms for helping
machines improve behavior. So don’t get confused by any claims to the con-
21.2 Supervised Learning 261
trary; neural networks are not like the brain (brains are much, much more
complex) and reinforcement learning is not like training circus animals (cir-
cus animals, at least for now, are much smarter).
In unsupervised learning, the robot gets a bit of reward or punishment
some of the time. In contrast, in supervised learning, the robot gets the full
answer in the form of the magnitude and direction of the error. Consider
the case of ALVIN, the autonomous driving robot van which learned to steer
on the roads around Pittsburgh, Pennsylvania, through the use of a neural
network. Each time it turned the steering wheel, the supervised learning
algorithm told it just how it should have done it precisely right. It then com-
E RROR puted the error, the difference between what it did and what it should have
done (just as we defined it in Chapter 10), and used that difference to update
its network (the weights on the links). By doing this type of training some
hundreds of thousands of times, mostly in simulation (using real images of
the road but not actually driving on the physical road), the robot learned to
steer quite well.
You can see how supervised learning provides much more feedback to
the robot than unsupervised learning does. Both require a lot of training
trials, and each is suitable for different types of learning problems. Learning
robot inverse kinematics and dynamics (remember Chapter 6?) lends itself to
supervised learning, where the right answer is readily available and the error
can easily be computed. As an example, consider a robot that tries to reach
and point at a particular object; the distance from that object and where the
robot actually reached and pointed provides the error, and allows the robot
to learn by reaching over and over again. Not surprisingly, a great many
robots have learned their inverse kinematics in just this way.
As another example, consider maze learning: the robot has no map of the
maze (if it did, then it would not be maze learning, just map searching and
path planning) and has to learn the best path to the cheese. It does this by
trying different paths; if it runs into a dead end in the maze, it learns that
was a bad place to go, but it cannot compute the error from the right thing to
do, because it does not know the right thing to do (that would require a map
of the maze, which it does not have). Therefore, it must use reinforcement
learning and discover the path to the cheese that way.
21.3 Learning by Imitation/From Demonstration

Some types of learning are best done by being shown by a teacher. Consider
how babies learn: they do a great deal of both supervised and unsupervised
learning as described above (reaching, throwing, putting in the mouth and
chewing, babbling, crawling, and so on), but they also learn a huge amount
from friendly humans.
Learning from demonstration and by imitation is incredibly powerful, be-
cause it frees the learner from having to do things by trial and error, and
from making large mistakes. The learner, human or robot, gets the perfect
example of what to do, and merely needs to repeat what it sees.
If only it were that simple. As it turns out, it is not. In fact, very few species
on Earth are capable of learning this way; only humans, dolphins, and chim-
panzees are believed to be able to learn arbitrary new skills by seeing them
performed by others. The rest of the animals either can’t do it at all, or can
learn just a few things this way (monkeys and parrots fall into this category).
And although many pet owners would disagree, and claim that their partic-
ular pet can learn anything by adoringly watching the owner, science so far
tells us otherwise.
Imitation is very powerful yet very hard. Fortunately for robots, we can
try to understand why it is hard, yet still make it possible and useful. So why
is it hard?
In order for a robot to learn by demonstration, it must be able to:
• Pay attention to a demonstration. (Something buzzed, but I can’t look

away! Do I watch the legs or the arms? Oops, I moved and missed some-
thing.)
• Separate what is relevant to the task being taught from all irrelevant in-
formation. (Does the color of the shoes matter? What about the specific
path being traced?)
• Match the observed behavior to its own behaviors and effectors, taking
care about reference frames. (The teacher’s right is my left and vice versa.
The teacher has two arms but I have two wheels, so what do I do? The
teacher is more flexible than I am, my joint limits do not allow me to do
what I saw.)
• Adjust the parameters so the imitation makes sense and looks good. (The
teacher did this slowly but I can do it faster. Will it look as good and get
the job done? Do I reach lower since I’m shorter?)
Figure 21.2 A robot learning by following a teacher and being “put through” the
task. (Photo courtesy of Dr. Monica Nicolescu)
• Recognize and achieve the goals of the performed behavior. (Was that a
wave or a reach to a particular position in space? Was the object moved
on purpose or as a side-effect?)
The above list is not easy for people (just imagine or recall learning a new
complicated dance), much less for robots, for all the reasons of uncertainty,
limited perception, and constrained actuation we have discussed already in
several chapters.
While it may not be easy, learning by demonstration/from imitation is a
growing area of robot learning, because it not only promises to make robot
programming easier, but it creates the opportunity for people and robots to
interact in a playful way. This is important for the rapidly growing field of
H UMAN -R OBOT human-robot interaction, which we’ll talk more about in the next chapter.
I NTERACTION Learning by imitation involves careful decisions about internal representa-
tion. In one implemented system, shown in figure 21.2, a mobile robot learns
a new task by following the teacher around and doing whatever she does. In
the process, the robot experiences a set of very similar inputs to those of the
teacher, and can observe itself performing a set of behavior, for example: “Go
along the corridor, then grab the red box, and carry it to the next room, and
drop it in the corner.” By “mapping" its observations to its own behaviors,
Figure 21.3 A robot teaching another robot the same task by "putting through."
(Photo courtesy of Dr. Monica Nicolescu)
the robot learns new tasks. For example, the robot was able to learn tasks
such as fetching and delivering objects and navigating around an obstacle
course, and could then teach those new tasks to another robot, through the
same process of demonstration and imitation, as shown in figure 21.3. This
type of approach finds inspiration in theories from neuroscience, and works
well with modular behavior-based systems.
P UTTING T HROUGH The idea of having the learner experience the task directly is called putting
through. This is a very powerful way to learn, and is often used for teaching
people skills, such as how to swing in golf or tennis. It helps to try something
out and know how it “feels” when done correctly. The same idea can be used
in teaching robots. For this to work, the robot needs to have some internal
representation of its own behaviors or actions, so it can remember what it
experienced and how it can generate that type of behavior again. It can then
teach others, too, as shown in figure 21.3.
In another example of robot learning by demonstration, a humanoid robot,
shown in figure 21.4, learned to balance a pole on its “finger” by watching
a human do it. Unlike humans, the robot could learn the task after only a
few trials, over a few minutes, and then perform it perfectly thereafter. How
is that possible, and why can’t people do that? A part of the reason has to
Figure 21.4 A humanoid robot learning to balance a pole by watching a human

teacher. (Photo courtesy of Dr. Sethu Vijayakumar)
do with the robot’s ability to have a model of its own motor control. In this
example, the robot knew exactly how its arms worked, and thus could per-
fect its pole-balancing by tuning a few parameters after a small number of
trials with the pole. Sometimes robots do have an unfair advantage, in this
case because the robot had direct computational access to its inverse kine-
matics and dynamics, while people do not. We have to learn by trial and
error, while robots can learn by quickly building a model of what works for
their particular explicitly modeled system parameters.
If you remember what we learned in Chapter 9 you may be wondering
how the robot could track the pole accurately by using vision while it was
moving around, since that is not a simple problem. In fact, the robot watched
only the two colored balls on the pole, and the way they moved relative to
each other. It had a model of pole-balancing which involved the top ball
being above the bottom one, and so it learned how to keep it in that position.
What it really learned was to balance a pole (which was attached to those
balls), which happens to be a well-studied inverted pendulum problem from
control theory (remember that from Chapter 10?).
21.4 Learning and Forgetting

Because robot learning and learning in general are such fascinating and chal-
lenging problems, various methods for learning have been developed, and
have been tried on robots with different levels of success. Besides the ap-
proaches described above, there are memory-based learning, evolutionary
learning, case-based learning, statistical learning, and many more. In fact,
there are too many to put in one chapter, but hopefully you are curious and
will look them up and learn more about them.
Something you must not forget when considering learning is forgetting!
Forgetting seems like a bad thing, at least when it comes to human memory,
assuming that all information we have ever acquired is worth keeping. How-
ever, that is not actually the case, and it is impossible with limited memory.
F ORGETTING In machine learning, therefore, purposeful and careful forgetting is useful,
because it allows for discarding outdated previously learned information in
favor of newer, more current information. Forgetting is necessary for two
reasons:
1. Making room for new information in finite memory space
2. Replacing old information that is no longer correct.
Computer memory these days is cheap, but it’s not infinite, so eventu-
ally some stuff has to be thrown out to make room for new stuff. Even if
memory were to become, in effect, infinite, and robots could store all they
ever learned, we would still have some challenges. First, there would have
to be methods for very quickly searching all that stored information in or-
der to get to something specific. This is a well-studied problem in computer
science; there are ways to organize information so things can be looked up
efficiently, but it takes work. More important, given a lot of information on
a given topic or task, some of which may be contradictory (because what the
robot learns over time naturally changes), how does the robot decide what
information/knowledge to use? A simple answer is to use the latest, newest
information, but that may not always work, since just because something
was learned or experienced recently, that does not mean it is the best ap-
proach or solution. For example, the most recent way you saw from home to
the store may have been circuitous and slow; you should not reuse that way
but instead reuse the path you took last week, which was much shorter.
Knowing something or having some previously learned skill can interfere
with learning something new. Rats that learn to run a particular maze have a
21.4 Learning and Forgetting 267
hard time adapting if a path to the cheese is unexpectedly blocked; they run
into the new wall even though they see it, because they believe their learned
internal map of the maze. Eventually they relearn, but it takes a while. Sim-
ilarly, people who have learned to play tennis take a bit to adjust to playing
squash, because in the former you have to hold your wrist rigid while in the
latter you have to bend it a lot to swing the (smaller and lighter) racquet.
Finally, sometimes people absolutely refuse to accept new information or
knowledge because it conflicts with something they already know and pre-
fer to believe; we call this denial. These are just a few illustrative examples,
ranging from rodents to people, of how learning, unlearning, and forgetting
need to work together and can cause problems when they do not.
L IFELONG L EARNING Lifelong learning, the idea of a robot learning continually and constantly
improving, is something roboticists dream about but do not yet know how
to do. This is a great challenge for many reasons. Aside from learning itself
being challenging, for all the reasons outlined so far in this chapter, lifelong
learning is also hard because so far we have different approaches to learning
different things (from maps to juggling to delivery), as you saw above. Since
a robot needs to learn (and forget) a variety of things in its lifetime, such
multiple learning methods have to be combined into a single system in some
principled and efficient way.
To summarize
• Robot learning is harder than other types of learning because it involves

uncertainty in sensing and action, and large and often continuous sensory
input spaces and effector action output spaces.
• The amount and type of information (feedback, reward, punishment, er-

ror) available to the learning robot determine what type of learning meth-
ods are possible for a particular learning problem.
• Various robot learning methods have been studied and continue to be pur-
sued in research, including reinforcement learning, supervised vs. unsu-
pervised learning, learning by imitation/demonstration, and evolution-
ary learning, to name a few.
• A robot can use multiple learning methods and learn a variety of things
at the same time.
• Robot learning is in its infancy, with great new discoveries yet to be made.
Food for Thought
Some people find the idea of robots learning frightening, because they are
uncomfortable with the notion of robot behavior being anything but com-
pletely predictable. Does adding learning to the robot’s abilities make it more
unpredictable? If so, how can this be helped? If not, how can people’s wor-
ries be put to rest?
Looking for More?
http://roboticsprimer.sourceforge.net/workbook/Learning
• Tom Mitchell’s Machine Learning is a popular and comprehensive textbook

spanning the wide variety of learning approaches.
• Reinforcement Learning: An Introduction by Rich Sutton and Andy Barto

is an excellent introduction to this popular learning approach.
• Robot Learning edited by Jonathan Connell and Sridhar Mahadevan is a

nice collection of real robot learning approaches and results.
• Probabilistic Robotics by Sebastian Thrun, Wolfram Burgard, and Dieter Fox

Building, the same text we recommended in Chapter 19, covers the uses of
statistical and probabilistic approaches in robotics for applications beyond
navigation. The same warning applies: this is not simple reading and
requires a fair amount of mathematics background.
22 Where To Next?
The Future of Robotics
Right now, as you are reading this book, we are at a particularly interesting
time in the history of robotics.
The great thing about the above statement is that it seems true no matter
when you read the book. Roboticists have been continually excited about
the future of their field. Most will admit that they have predicted major
breakthroughs at various times, and many such breakthroughs have taken
place.
However, the start of the twenty first century is an especially pivotal time
for the field. Here is why:
• Sensors, effectors, and bodies are becoming very sophisticated. We are creating
the most complex robots yet, with bodies that mimic biological form and
attempt to model biological function.
• Computers are faster and cheaper than ever before. This means robot brains can
be more sophisticated than ever, allowing them to think and act efficiently
in the real world.
• Wireless communication is everywhere. This means robots can communicate

with other computers in the environment, so they can be better informed,
and therefore smarter.
Other advances have played a role as well. Importantly, as you recall from
Chapter 2, robotics got its name from menial labor and has historically done
its share of factory work. But now robots are starting to enter our daily lives,
and that is going to change everything.
The most pervasive use of robotics until recently was in car factories, as
shown in figure 22.1. Robots were perfected to be so effective that they are
270 22 Where To Next?
Figure 22.1 Robot arms assembling a car. (Photo courtesy of KUKA Schweissanla-
gen GmbH in Augsburg, Germany)
Figure 22.2 An assembly robot involved in gene sequencing. (Photo courtesy Andre
Nantel)
271
Figure 22.3 Sony’s Aibo robot dogs. (Photo courtesy of David Feil-Seifer)
not considered proper robots any more, but fit into the more general cate-
FACTORY gory of factory automation. At the end of the twentieth century, robots did
A UTOMATION some finer work of the assembly type: they were instrumental in helping
people sequence the human genome. As in car assembly, the brain work was
done by people, but the robots did the hard, precise, and repetitive work,
as shown in figure 22.2. As a result, the field of genetics took a tremendous
leap, and the process is being applied to genome sequencing of numerous
other species, toward doing better science and developing better methods
for improving human health.
You probably had no idea that robots were involved in sequencing the hu-
man genome, since they were far from the spotlight. Hidden in factories
and labs, the robots are not getting much exposure. A much more appeal-
ing robot that got a great deal more exposure was the Sony Aibo robot dog,
shown in figure 22.3. The Aibo went through several generations, in which
its overall shape was refined, but in all cases it was universally deemed
“cute” and was used in a variety of research and educational projects. Unfor-
tunately, while Aibo was cute, it was neither useful nor affordable, and thus
not very appealing to a broad audience.
In the meantime, perhaps the most popular question for roboticists over
the decades has been “Can it clean my house?” The answer, finally, is “Yes, it
Figure 22.4 An autonomous vacuum cleaner by iRobot Inc. (Photo courtesy of

iRobot Inc.)
can!” Several robotic vacuum cleaners were developed in the last decade of
the twentieth century, but one in particular, the small, simple, and affordable
Roomba by iRobot, shown in figure 22.4, filled the special niche of an af-
fordable and sufficiently useful novelty high-tech product. Over two million
such robots have been sold to households in the United States and beyond,
and a wet version that mops the floor, called Scooba, has also been released.
The Roomba is a very important robot, albeit a simple and modest one.
Because it is so simple and affordable, the Roomba has the chance to be-
come a part of people’s everyday lives (or at least as often as people clean
their floors). In this way, numerous people can finally begin to interact with
robotics technology in an approachable setting. And as a result, we can hope
that many young people will be inspired to take the Roomba and aim to
develop the next highly successful robot that will go into millions of homes.
As robotics is poised on the thresholds of our homes, let us consider some
of the exciting directions the field of robotics is taking that will shape the
future not only of the field but also of our world.
22.1 Space Robotics 273
Figure 22.5 NASA’s Sojourner, the first rover to roam on Mars. (Photo courtesy of
NASA)
22.1 Space Robotics
Working in harsh environments has always been one of the most natural
uses for robots. Space is the ultimate frontier, and one where robots have al-
ready made significant contributions. NASA’s well-known mission to Mars
dropped off a very large and well-bundled bouncing “ball” on the surface
of the planet. The ball, when it stopped bouncing and rolling, and opened
up, revealed the Pathfinder lander. Shortly thereafter, a hatch on the lan-
der opened and a rover drove out. That amazing little rover was Sojourner,
shown in figure 22.5. On July 4, 1997, it was the first to drive autonomously
around the surface of Mars, collect images, and send them back to Earth. So-
journer’s name means “traveler” and was chosen from 3500 entries submit-
ted as part of a yearlong worldwide student competition. Valerie Ambrose,
of Bridgeport, Connecticut, who was twelve at the time, submitted the win-
ning essay about Sojourner Truth, an African-American reformist from the
Civil War era whose mission was to “travel up and down the land” advocat-
ing the rights of all people to be free and the rights of women to participate
fully in society. It’s good to name good robots after good people.
Figure 22.6 NASA’s Robonaut, the first space robot with a humanoid torso. (Photo
courtesy of NASA)
Since Sojourner, NASA and other space agencies around the world are
considering terraforming first the moon, and eventually Mars, with a view
to potential human colonization. What used to be science fiction may be-
come the future, and robots will be necessary to make it possible. Currently
robots such as NASA’s Robonaut, shown in figure 22.6, are being developed
to be capable of working side by side with astronauts, as well as completely
autonomously, in order to do tasks that range from fixing problems with
spacecraft to doing exploration and sample return from planets, moons, and
even asteroids, to constructing habitats for human use.
22.2 Surgical Robotics
Let’s jump from very large and distant spaces to very small and up-close-
and-personal ones, our bodies. Surgical robotics is one of the better-establish-
ed uses of robots in challenging real-world environments. As briefly men-
tioned in Chapter 6, robots have been involved in hip surgery and even brain
surgery for some time. They are excellent replacements for human workers
in those domains because precision is of critical importance (Can you walk
without part of a hip or think without part of your brain? Would you want
22.2 Surgical Robotics 275
Figure 22.7 A surgical robot at work. (Photo ©[2007] Intuitive Surgical, Inc.)
to?), but intelligence is not. In fact, these machines are not intelligent and
autonomous at all, but are remotely controlled by highly skilled surgeons,
who not only are well practiced in surgery but also are well trained in us-
R OBOT-A SSISTED ing robots to do surgery, which is typically called robot-assisted surgery. The
S URGERY surgery is merely “assisted,” in that the human surgeon does all the perceiv-
ing (looking through the special glasses into what the camera on the robot
sees), thinking (deciding what actions to take), and doing (taking the actions
with the remote controls). The robot is a teleoperated manipulator, with spe-
cial endeffectors ideally suited for the particular surgical task. This process
is so well worked out that in some cases the surgeon need not be in the same
room or even state as the patient; surgery has been performed across con-
tinents, with patients in the United States and surgeons in France, and vice
versa. This sounds like a great idea for providing surgical aid to people in
underdeveloped countries or on the battlefield, until one considers that such
robots currently cost about a million dollars. (Perhaps there will be a good
sale soon.)
It is easy to imagine the many new directions that surgical robotics re-
search will pursue. Perception is the first challenge; surgeons would like the
ability to “feel” the area in which the robot is operating, the way they would
if they were doing the work by hand (so to speak). This involves complex
H APTICS research challenges in haptics, the study of the sense of touch. Interestingly,
the term comes from the Greek haptikos meaning “being able to grasp or per-
ceive.” Take a second to envision the environment of a surgical robot: there
are no simple rigid or flat surfaces, and the lighting is less than ideal. Finally,
today there are no detailed models of the environment available in advance,
though they could be made before surgery with sophisticated imaging equip-
ment, if it is available.
Surgeons are far from ready to give up control to a robot (or anyone?)
yet, but if we are to contemplate robots that can perform any of this work
on their own, research will have to address real-time decision-making and
control within the human body. If that is not scary enough, consider that
N ANOROBOTICS there is also ongoing research into nanorobotics, truly tiny machines that can
float freely in our bloodstreams and hopefully do some useful work such
as scraping plaque from blood vessels, detecting danger areas, and so on.
Needless to say, this research is still in its infancy, since manufacturing a tiny
thing requires another tiny thing, and we do not yet have such tiny things
S ELF -A SSEMBLY available. To counter this problem, some researchers are looking into self-
assembly, which is exactly what you would guess: the ability for robots to
assemble themselves and each other, especially at small scales. That is still
mostly in the zone of science fiction, so let’s turn to some large and more
current directions.
22.3 Self-Reconfigurable Robotics
In theory at least, a swarm of tiny robots can get together to create any shape.
In reality, robots are constructed with rigid parts and do not shape-shift eas-
R ECONFIGURABLE ily. But an area of robotics research called reconfigurable robotics has been de-
R OBOTICS signing robots that have modules or components that can come together in
various ways, creating not only different shapes but differently shaped robot
bodies that can move around. Such robots have been demonstrated to shape-
shift from being like six-legged insects to being like slithering snakes. We
say “like” because six-legged insect robots and snakelike robots have been
built before, and were much more agile and elegant than these reconfigurable
robots. However, they could not shape-shift.
The mechanics of physical reconfiguration are complex, involving wires
and connectors and physics, among other issues to consider and overcome.
Individual modules require a certain amount of intelligence to know their
state and role at various times and in various configurations. In some cases,
22.4 Humanoid Robotics 277
Figure 22.8 A reconfigurable robot striking (just one) pose. (Photo courtesy of Dr.
Wei-Min Shen)
the modules are all the same (homogeneous), while in others, they are not
(heterogeneous), which makes the coordination problem in some ways re-
lated to multi-robot control we learned about in Chapter 20. Reconfigurable
robots that can autonomously decide when and how to change their shapes
are called self-reconfigurable. Such systems are under active research for
navigation in hard-to-reach places, and beyond.
22.4 Humanoid Robotics
Some bodies change shape, and others are so complicated that one shape
is plenty hard to manage. Humanoid robots fit into that category. As we
learned in Chapter 4, parts of the human body have numerous degrees of
freedom (DOF), and control of high-DOF systems is very difficult, especially
in the light of unavoidable uncertainty found in all robotics. Humanoid con-
trol brings together nearly all challenges we have discussed so far: all aspects
of navigation and all aspects of manipulation, along with balance and, as you
will see shortly, the intricacies of human-robot interaction.
There is a reason why it takes people so long to learn to walk and talk and
be productive, in spite of our large brains; it is very, very hard. Human be-
Figure 22.9 Sarcos humanoid robot. (Photo courtest of Sarcos Inc.)
ings take numerous days, months, and years to acquire skills and knowledge,
while robots typically do not have that luxury. Through the development of
sophisticated humanoid robots with complex biomimetic sensors and actua-
tors, the field of robotics is finally having the opportunity to study humanoid
control, and is gaining even more respect for the biological counterpart in the
process.
22.5 Social Robotics and Human-Robot Interaction
Robotics is moving away from the factory and laboratory, and ever closer
to human everyday environments and uses, demanding the development of
robots capable of naturally interacting socially with people. The rather new
HUMAN - ROBOT field of human-robot interaction (HRI) is faced with a slew of challenges which
INTERACTION (HRI) include perceiving and understanding human behavior in real time (Who is
that talking to me? What is she saying? Is she happy or sad or giddy or
22.5 Social Robotics and Human-Robot Interaction 279
mad? Is she getting closer or moving away?), responding in real-time (What

should I say? What should I do?), and doing so in a socially appropriate and
natural way that engages the human participant.
Humans are naturally social creatures, and robots that interact with us will
need to be appropriately social as well if we are to accept them into our lives.
When you think about social interaction, you probably immediately think
about two types of signals: facial expressions and language. Based on what
you know about perception in robotics, you may be wondering how robots
will ever become social, given how hard it is for a moving robot to find, much
less read, a moving face in real time (recall Chapter 9), or how hard it is for
a moving target to hear and understand speech from a moving source in an
environment with ambient noise (recall Chapter 8).
Fortunately, faces and language are not the only ways of interacting so-
cially. Perception for human-robot interaction must be broad enough to in-
H UMAN I NFORMATION clude human information processing, which consists of a variety of complex sig-
P ROCESSING nals. Consider speech, for example: it contains important nuances in its rate,
volume level, pitch, and other indicators of personality and feeling. All that
is extremely important and useful even before the robot (or person) starts to
understand what words are being said, namely before language processing.
P ROXEMICS Another important social indicator is the social use of space, called proxemics:
how close one stands, along with body posture and movement (amount, size,
direction, etc.), are rich with information about the social interaction. In some
cases, the robot may have sensors that provide physiological responses, such
as heart rate, blood pressure, body temperature, and galvanic skin response
(the conductivity of the skin), which also give very useful cues about the
interaction.
But while measuring the changes in heart rate and body temperature is
very useful, the robot still needs to be aware of the person’s face in order to
S OCIAL G AZE “look” at it. Social gaze, the ability to make eye contact, is critical for nor-
mal social interpersonal interaction, so even if the robot can’t understand
the facial expression, it needs at least to point its head or camera to the
face. Therefore, vision research is doing a great deal of work toward creating
robots (and other machines) capable of finding and understanding faces as
efficiently and correctly as possible.
N ATURAL L ANGUAGE And then there is language. The field of natural language processing (NLP)
P ROCESSING (NLP) is one of the early components of artificial intelligence (AI) , which split off
early and has been making great progress. Unfortunately for robotics, the
vast majority of success in NLP has been in written language, which is why
search engines and data mining are very powerful already, but robots and
computers still can’t understand what you are saying unless you say very
few words they already know, very slowly and without an accent. Speech
processing is a challenging field independent from NLP; the two sometimes
work well together, and sometimes not, much the same as has been the case
for robotics and AI. Great strides have been made in human speech process-
ing, but largely for speech that can be recorded next to the mouth of the
speaker. This works well for telephone systems, which can now detect when
you are getting frustrated by endless menus (“If you’d like to learn about our
special useless discount, please press 9” and the all-time favorite “Please lis-
ten to all the options, as our menu has changed”), by analyzing the qualities
of emotion in speech we mentioned above. However, it would be better if
robot users did not have to wear a microphone to be understood by a robot.
To make this possible, research is taking place at the levels you learned about
in Chapter 7: hardware, signal processing, and software.
The above just scrapes the surface of the many interesting problems in
HRI. This brand-new field that brings together experts with backgrounds not
only in robotics but also in psychology, cognitive science, communications,
social science, and neuroscience, among others, will be great to watch as it
develops.
22.6 Service, Assistive and Rehabilitation Robotics
S ERVICE R OBOTS The twenty first century will witness a great spectrum of service robots, ma-
chines used in a variety of domains of human life. Some of the more mun-
dane ones, which include window washers and fetch-and-carry delivery sys-
tems, are already deployed in controlled environments such as factories and
warehouses. The more interesting systems, aiming at daily use in hospitals,
schools, and eventually homes, are almost within reach. The Roomba, shown
earlier, is an example of one such system where many are expected to follow.
A SSISTIVE R OBOTICS Assistive robotics refers to robot systems capable of helping people with
special needs, such as individuals convalescing from an illness, needing re-
habilitation following an accident or trauma, learning or training in a special
setting, or aging at home or in a managed care facility. For example, Pearl
the Nursebot (shown in Figure 22.10), designed at Carnegie Mellon Univer-
sity, roamed around an elder care home and helped people find their way to
the dining room, the TV room, and other important places. Did you know
that older people who have a pet or even a plant tend to live longer and re-
port that they are happier, but often those pets and plants go unfed? Perhaps
22.6 Service, Assistive and Rehabilitation Robotics 281
Figure 22.10 Nursebot, developed at Carnegie Mellon University. (Photo courtesy

of Carnegie Mellon University)
robots can be effective companions to people, and make them happy and live
longer without requiring food in return. Who knows?
R EHABILITATION There are many different ways in which people can receive assistance. Re-
R OBOTS habilitation robots provide hands-on help by moving the parts of the body of
the patient in order to guide prescribed exercises and recovery. Figure 22.11
shows a rehabilitation robot and computer system; the robot moves the pa-
tient’s arm while the screen provides instructions.
Another way to help people involved in rehabilitation is not to guide their
S OCIALLY A SSISTIVE arms and legs, but to provide social interaction through HRI. Socially assistive
R OBOTICS robotics studies robots that are capable of monitoring their users and pro-
viding coaching, motivation, and encouragement without physical contact.
Figure 22.12 shows Thera, a robot developed at the University of Southern
California, which helped stroke patients perform their exercises. During the
two to three months immediately after a stroke, the patient has by far the best
chance to avoid permanent disability, but only if she properly performs very
carefully designed exercises. Because over 750,000 people have strokes each
year in the United States alone, and that number is expected to double in the
next twenty years with the growing elderly population, robots can play a key
role in helping a lot of people recover from stroke.
Figure 22.11 A rehabilitation robot in action. (Photo courtesy Dr. Hermano Krebs)
Figure 22.12 A socially assistive robot helping a stroke patient. (Photo courtesy of
the author)
22.7 Educational Robotics 283
Figure 22.13 Another socially assistive robot.
Socially assistive robotics is focused on developing robots for a variety of

users beyond stroke patients. For example, robots are being used as tools
for studying human social behavior and various social anomalies, ranging
from pathological lying to extreme shyness. Such robots can also be used
to help children with autism spectrum disorder (ASD), who are particularly
interested in machines and are often very talented, but have serious social
and developmental problems. Socially assistive robots may also be able to
help kids with attention deficit and hyperactivity disorder (ADHD). By cre-
ating robots that can be intelligent and autonomous companions and helpers,
we can discover entirely new areas where such robots can improve human
quality of life.
22.7 Educational Robotics
What better place to improve than in school? As you saw in Chapter 21,
robots can learn. More importantly, they can teach, and serve as excellent
tools for teaching. Children and people of all ages and cultures love to
play with robots, program them, and push them to their limits. As you no
doubt noticed, understanding how robots work and programming them to
do what you want is far from simple, and as you learned how to do it, you
also learned a great deal more. Working with robots provides experiential,
hands-on education which most people who try it find very interesting and
motivating. Imagine working with the robot in figure 22.14; would that make
Figure 22.14 A robot can be an interesting learning tool and buddy. (Photo courtesy
of the author)
class more fun? Few people are bored by robots they can program and con-
trol, because the challenge is open and endless.
People love challenges, so robotics has been shown to be a superb tool for
hands-on learning, not only of robotics itself (as we have done in this book)
but also of general topics in science, technology, engineering, and math, also
STEM T OPICS called STEM topics. As it happens, STEM topics are not as popular with
children and students as they should be, considering what great jobs, careers,
and incomes are available in those areas. This creates an undersupply of
trained people to take up those great jobs, so increasing attention has been
paid to developing innovative tools for improved teaching of STEM topics.
Robotics is at the top of the list.
In addition to robotics courses in and after school and as part of univer-
sity curricula, a great many robotics contests have been popping up. These
include FIRST Robotics, LEGO Mindstorms (shown in figure 22.15), Botball,
Micromouse, and many others.
Because robotics is a great field to learn about and a great tool for learn-
E DUCATIONAL ing in general, the field of educational robotics, which creates such learning
R OBOTICS materials, is expected to grow in the future. We can look forward to all ele-
Figure 22.15 Some creative LEGO robots being put to good use. (Photo courtesy of
Jake Ingman)
mentary students having an opportunity to explore robotics in order to get

excited about learning.
22.8 Ethical Implications

Before we finish this book, we need to think about one very important aspect
of any rapidly changing technology, including robotics. The media, in par-
ticular the movies, have often portrayed robots as evil. Why not? After all,
they make such great bad guys. In all that bad publicity, it is easy to forget the
major motivations of the field of robotics, which are aimed at helping people.
However, any technology has the potential to be abused, and robotics is no
exception. As we develop more intelligent and capable robot systems, we
need to take great care about many important ethical issues. Here are some
of them:
• Safety: No robot should hurt a human being. This is Asimov’s First Rule
of Robotics, and it’s a great one to follow. Robots must naturally be safe
when they interact with people, such as in the applications we listed ear-
lier in the chapter. But people should also be safe when they design robots,
and stay away from human-killing machines. People often worry about
robots “going bad.” But the real concern is people going bad; robots are
only as good or bad as the people who create them.
• Privacy: Robots that become effective at working with people and in hu-
man environments will see and hear a great deal of information that they
could, potentially, pass on to somebody who should not see or hear it.
Privacy has to be taken seriously when the robot is designed, not after the
fact, as a result of users’ complaints.
• Attachment: If robots succeed in being engaging and useful, their human

users will become attached to them, and will not want to throw them out
or let go of them when the lease/lifetime/warranty expires. (Consider
how many people still drive very old cars in spite of the cost, or stick to
old computers or obsolete technology due to habit.) But what happens
when an outdated robot can no longer be maintained? Roomba users
already refuse to have their Roombas replaced when they need repair,
insisting on getting the same one back. What happens when the robot is
much more interesting, intelligent, and engaging than the Roomba?
• Trust and autonomy: Choosing to use a robot implies trusting the tech-
nology. What if the robot, out of uncertainty, an error, or simply its own
best judgment, acts in a way that the user finds unacceptable and there-
fore untrustworthy? Nobody would want a pushy or bossy robot, but
presumably everybody would want a robot that can save their life. How
does a robot earn and keep the user’s trust? How much autonomy should
the robot have, and who is really in charge?
• Informed use: Users of robots need to be aware of all of the issues above
and any others that are involved in the system they are using. As technol-
ogy becomes more complex, it is increasingly difficult to know everything
about the system (Do you know all the features of your phone?), yet those
features may have very serious implications for safety, privacy, attach-
ment, and trust. Designing systems that are easy to use and allow the
user to know them completely is a major open challenge in technology.
In Summary
The future of robotics is in your hands! Go do something great, but remem-

ber that innovation and discovery involve risks. Think about the potential
uses of what you are developing, beyond the obvious, and to the unintended.
Be careful, be smart, and do some good with robotics.
Food for Thought
• While the potential for robots to help people is vast, some people believe
that instead of creating technology that replaces human labor, we should
focus on training people as providers of those services. With the excep-
tion of jobs that are inherently dangerous or undesirable, it is important
to be able to properly justify robots in roles that could be filled by people,
if such people were available. How do you think this should be justified?
Are robots taking these roles away from people or are they filling other-
wise unfilled roles? What if those roles are unfilled now, but once robots
fill them, they could never again be filled by people? These are just some
of the ethical and economic questions that have been posed. How would
you answer them?
• Some philosophers have argued that once robots become sufficiently in-
telligent and self-aware, they will be so much like people that it will be
unethical to have them working for us, as it will create a race of slaves.
What do you think about that?
Looking for More?
Well, there is no more in this book, except for the glossary of terms and the
index.
If you haven’t yet tried all the robot programming exercises in the the
Robotics Primer Workbook, you should:
http://roboticsprimer.sourceforge.net/workbook/.
Since the workbook on the web is on a public site, a growing community will
be able to add new exercises and solutions, so it’s worth returning to the site
to both contribute and learn more.
There are numerous books and articles, popular and scientific, that you
can read about the future robotics. And of course there is science fiction, the
field that produced the term “robot” (remember Chapter 1?) and all manner
of robot characters.
The best place to start in science fiction about robotics is with Isaac Asimov,
who wrote a truly impressive number of books. I, Robot is a natural first read;
but read the book, don’t just see the movie, since the former is incomparably
better than the latter.
Based on what you have learned so far, you should be able to weed out
the wheat from the chaff in your future readings about robotics. Best of luck,
and enjoy!
Bibliography
Edwin A. Abbott. Flatland: A Romance of Many Dimensions. Dover Publications, Mi-

neola, NY, 1952.
Harold Abelson and Gerald Jay Sussman. Structure and Interpretation of Computer
Programs. The MIT Press, Cambridge, MA, 1996.
Ronald Arkin. Behavior-Based Robotics. The MIT Press, Cambridge, MA, May 1998.
W. R. Ashby. Introduction to Cybernetics. Chapman and Hall, London, England, 1956.
Isaac Asimov. I, Robot. Gnome Press, United States of America, 1950.
Dana Ballard and Christopher Brown. Computer Vision. Prentice-Hall Inc., Englewood
Cliffs, NJ, May 1982.
R. Beckers, O. E. Holland, and J_L Deneubourg. From local actions to global tasks:
Stigmergy in collective robotics. In R. Brooks and P. Maes, editors, Artificial Life IV,
pages 181–189. MIT Press, Cambridge, MA, 1994.
George Bekey. Autonomous Robots. The MIT Press, Cambridge, MA, 2005.
Rodney Brooks. Cambrian Intelligence. The MIT Press, Cambridge, MA, July 1999.
H. Choset, K. M. Lynch, S. Hutchinson, G. Kantor, W. Burgard, L. E. Kavraki, and
S. Thrun. Principles of Robot Motion: Theory, Algorithms, and Implementations. The
MIT Press, Cambridge, MA, June 2005.
H. Choset, K. M. Lynch, S. Hutchinson, G. Kantor, W. Burgard, L. E. Kavraki, and
S. Thrun. Principles of Robot Motion: Theory, Algorithms, and Implementations. The
MIT Press, Cambridge, MA, June 2005.
Jonathan Connell and Sridhar Mahadevan, editors. Robot Learning. Kluwer, Boston,
MA, June 1993.
Thomas Corman, Charles Leiserson, and Ronald Rivest. Introduction to Algorithms.
The MIT Press, Cambridge, MA, September 2001.
J. J. D’Azzo and C. Houpis. Linear Control System Analysis and Design: Conventional
and Modern. McGraw-Hill, New York, NY, 4 edition, 1995.
290 Bibliography
J. L. Deneubourg, S. Goss, N. R. Franks, A. Sendova-Franks, C. Detrain, and L. Chre-

tien. The dynamics of collective sorting: Robot-like ants and ant-like robots. In
J-A Meyer and S. Wilson, editors, Simulation of Adaptive Behaviour: from animals to
animats, pages 356–365. MIT Press, Cambridge, MA, 1990.
Richard C. Dorf and Robert H. Bishop. Modern Control Systems. Prentice-Hall, Engle-
wood Cliffs, NJ, 10 edition, 2004.
Hugh Durrant-Whyte and John Leonard. Directed Sonar Sensing for Mobile Robot Nav-
igation. Springer, Dordrecht, The Netherlands, May 1992.
H. R. Everett. Sensors for Mobile Robots: Theory and Applications. AK Peters Ltd., Natick,
MA, 1995.
James Jerome Gibson. The Perception of the Visual World. Houghton Mifflin, Boston,
MA, 1950.
Clive Gifford. How to build a robot, 2001.
Madan M. Gupta and Naresh K. Sinha, editors. Intelligent Control Systems, Theory and
Applications. IEEE Press, New York, 1995.
Simon Haykin and Barry Van Veen. Signals and Systems. Wiley & Sons, New York,
NY, 1999.
Berthold Klaus Paul Horn. Robot Vision. The MIT Press, Cambridge, MA, March 1986.
Paul Horowitz and Winfield Hill. The Art of Electronics. Cambridge University Press,
Cambridge, MA, 2 edition, July 1989.
B. C. Kuo. Automatic Control Systems. Prentice-Hall, Englewood Cliffs, NJ, 1962.
Steven M. LaValle. Planning Algorithms. Cambridge University Press, Cambridge,
MA, May 2006.
Steven Levy. Artificial Life. Random House Value Publishing, New York, NY, May
1994.
Fred Martin. Robotic Explorations: A Hands-on Introduction to Engineering. Prentice
Hall, Englewood Cliffs, NJ, 1 edition, December 2000.
Maja J Matarić. A distributed model for mobile robot environment-learning and nav-
igation. Master’s thesis, Massachusetts Institue of Technology, Cambridge, MA,
January 1990.
Maja J Matarić. Integration of representation into goal-driven behavior-based robots.
IEEE Transactions on Robotics and Automation, 8(3):304–312, June 1992.
Tom Mitchell. Machine Learning. McGraw-Hill Education, New York, NY, October
1997.
Gordon E. Moore. Cramming more components onto integrated circuits. 38(8):114–
117, April 1965.
Robyn Murphy. Introduction to AI Robotics. The MIT Press, Cambridge, MA, 2000.
Bibliography 291
Alan V. Oppenheim, Alan S. Willsky, and Nawab S. Hamid. Signals and Systems.
Prentice-Hall, Englewood Cliffs, NJ, August 1996.
Paolo Pirjanian. Multiple Objective Action Selection & Behavior Fusion using Voting. PhD
thesis, Aalborg University, Denmark, April 1998.
Julio Rosenblatt. DAMN: A Distributed Architecture for Mobile Navigation. PhD thesis,
Carnegie Mellon University, Pittsburgh, PA, 1997.
Stuart Russell and Peter Norvig. Artificial Intelligence, A Modern Approach. Prentice-
Hall, Englewood Cliffs, NJ, December 2002.
L. Sciavicco and B. Siciliano. Modeling and Control of Robot Manipulators. Springer,
Great Britian, 2 edition, 2001.
Roland Siegwart and Illah Nourbakhsh. Autonomous Mobile Robots. The MIT Press,
Cambridge, MA, April 2004.
Michael Sipser. Introduction to the Theory of Computation. PWS Publishing Company,
Boston, MA, 1996.
Mark Spong, Seth Hutchinson, and M. Vidyasagar. Robot Modeling and Control. John
Wiley and Sons Inc., United States of America, November 2005.
Rich Sutton and Andy Barto. Reinforcement Learning: An Introduction. The MIT Press,
Cambridge, MA, 1998.
Ellen Thro. Robotics, the marriage of computers and machines, 1993.
Sebastian Thrun, Wolfram Burgard, and Dieter Fox. Probabilistic Robotics. The MIT
Press, Cambridge, MA, September 2005.
W. Grey Walter. An imitation of life. 182(5):42–45, 1950.
W. Grey Walter. A machince that learns. 185(2):60–63, 1951.
Norbert Wiener. Cybernetics or Control and Communication in the Animal and the Ma-
chine. The MIT Press, Cambridge, MA, 2 edition, 1965.
E. O. Wilson. Ants. Belknap Press, Cambridge, MA, March 1998.
John Yen and Reza Langari. Fuzzy Logic: Intelligence, Control and Information. Prentice-
Hall, Englewood, CA, November 1998.
Glossary
Achievement goal A state the system tries to reach, such as the end of a
maze.
Actuator The mechanism that enables an effector to execute an action or

movement.
Actuator uncertainty Not being able to know the exact outcome of an action
ahead of time, even for a simple action such as “go forward three feet.”
Amplifying Something that causes an increase in size, volume, or signifi-

cance.
Artificial intelligence (AI) The field that studies how intelligent machines
(should) think. AI was officially “born” in 1956 at a conference held at
Dartmouth University, in Hanover, New Hampshire.
Assistive robotics Robot systems capable of helping people with special needs,
such as individuals convalescing from an illness, rehabilitating from an
accident or trauma, learning or training in a special setting, or aging at
home or in a managed care facility.
Autonomy The ability to make one’s own decisions and act on them.
Behavior coordination Deciding what behavior or set of behaviors should

be executed at a given time.
Bit The fundamental unit of information, which has two possible values: the
binary digits 0 and 1; the word comes from “binary digit.”
Bottom-up Progression from the simpler to the more complex.
Broadcast communication Sending a message to everyone on the commu-

nication channel.
294 Glossary
Calibration The process of adjusting a mechanism to maximize its perfor-

mance (accuracy, range, etc.).
Cartesian robots Robots that are similar in principle to Cartesian plotter

printers, and are usually used for high-precision assembly tasks.
Command arbitration The process of selecting one action or behavior from

multiple possibilities.
Continuous state State that is expressed as a continuous function.
Control architecture A set of guiding principles and constraints for organiz-

ing a robot’s control system.
Cooperation Joint action with a mutual benefit.
Coordination Arranging things in some kind of order (ord and ordin mean
”order” in Latin).
Cybernetics A field of study that was inspired by biological systems, from

the level of neurons (nerve cells) to the level of behavior, and tried to im-
plement similar principles in simple robots, using methods from control
theory.
Direct current (DC) motor A motor that converts electrical energy into me-
chanical energy.
Damping The process of systematically decreasing oscillations.
Data association problem The problem of uniquely associating the sensed

data with absolute ground truth.
Degrees of freedom The dimensions in which a manipulator can move.
Deliberative control Type of control that looks into the future, so it works
on a long time-scale.
Demodulator A mechanism that is tuned to the particular frequency of the

modulation, so it can can be decoded.
Desired state The state the system wants to be in, also called the goal state.
Detector A mechanism that perceives (detects) a property to be measured.
Differential drive The ability to drive wheels separately and independently,

through the use of separate motors.
295
Diffuse reflection Light that penetrates into the object, is absorbed, and then
comes back out.
Discrete state Different and separate state of a system (from the Latin discre-
tus meaning “separate”), such as down, blue, red.
Dynamics The study of the effects of forces on the motion of objects.
Effector Any device on a robot that has an effect (impact, influence) on the
environment.
Embodiment Having a physical body.
Emitter A mechanism that produces (emits) a signal.
Error The difference between a desired value and the measured value.
Exo-skeleton The outer skeleton (from the Greek ex meaning “outside”).
External state The state of the world, as the robot can perceive it
Internal state The state of the robot, as the robot can perceive it.
Exteroception The process of sensing the world around the robot, not in-
cluding sensing the robot itself.
Feedback The information that is sent back, literally “fed back”, into the
system’s controller.
Fixed hierarchies Hierarchies whose ordering does not change, like royal
families where the order of power is determined by heredity.
Focus A point toward which light rays are made to converge.
Foraging The process of finding and collecting something items from some
specified area, such as harvesting a crop or removing trash from a stretch
of highway.
Fusion The process of combining multiple possible possibilities into a single

result.
Ganged gears Gears positioned next to each other so their teeth mesh and,
together, they produce a change of speed or torque of the attached mech-
anism. Also called gears in series.
296 Glossary
Grasp points Locations where the fingers or grippers should be placed in

order to best grasp an object. Grasp points are computed relative to the
center of gravity, friction, location of obstacles, etc.
Goal state The state the system wants to be in, also called the desired state.
Haptics The study of the sense of touch.
Heuristics Rules of thumb that help guide and hopefully speed up the search.
Hierarchies Groups or organizations that are ordered by power (from the

Greek hierarkh meaning “high priest.”)
Holonomic Being able to control all available degrees of freedom (DOF).
Human-robot interaction (HRI) A new field of robotics focusing on the chal-

lenges of perceiving and understanding human behavior in real time (who
is that talking to me, what is she saying, is she happy or sad or giddy
or mad, is she getting closer or moving away?), responding in real-time
(what should I say? what should I do?), and doing so in a socially appro-
priate and natural way that engages the human participant.
Hybrid control Approaches to control that combine the long time-scale of

deliberative control and the short time-scale of reactive control, with some
cleverness in between.
Image plane The projection of the world onto the camera image; it corre-
sponds to the retina of the biological eye.
Inhibitory connection A connection in which the stronger the sensory in-

put, the weaker the motor output.
Inverse pendulum problem Controlling a system that is like an upside-down

pendulum, such as balancing in one-legged robots.
Joint limit The extreme of how far a joint can move.
Kinematics The correspondence between actuator motion and the resulting

effector motion.
Lens The structure that refracts the light to focus it on the retina or image
plane.
Level of abstraction Level of detail.

297
Life-long learning The idea of having a robot learn continually as long as it

is functional.
Linear actuator An actuator that provides linear movement, such as getting

longer or shorter.
Maintenance goal A goal that requires ongoing active work on the part of
the system, such as “keeping away from obstacles.”
Manipulation Any goal-driven movement of a manipulator.
Manipulator A robotic manipulator is an effector.
Manipulator links Individual components of the manipulator.
Model-based vision An approach to machine vision which uses models of

objects (prior information or knowledge about those objects) represented
and stored in a way that allows comparison and recognition.
Natural language processing (NLP) The field that studies understanding of

written and spoken language. A part of Artificial Intelligence.
Negative feedback Feedback that gets smaller in response to the input, re-
sulting in damping.
Non-holonomic Not being able to control all available degrees of freedom

(DOF).
Odometry Keeping track/measuring how far one has gone (from the Greek
hodos meaning “journey” and metros meaning “measure”).
Optimization The process of improving a solution to a problem by finding

a better one.
Optimizing search The process of looking for multiple solutions in order to

select the best one.
Passive actuation Utilizing potential energy in the mechanics of the effector

and its interaction with the environment to move the actuator, instead of
using active power consumption.
Perception See Sensing.
Photocell A cell that is sensitive to the amount of light that hits it.
298 Glossary
Photosensitive elements Elements that are sensitive to the light, such as

rods and cones in the biological eye.
Pixel A basic element of the image on the camera lens, computer, or TV

screen.
Polarized light The light whose waves travel only in a particular direction,
along a particular plane.
Polarizing filter A filter that lets through only the light waves with a partic-
ular direction.
Polygon of support The area covered by the ground points of a legged ob-
ject or robot.
Position control Controlling a motor so as to track the desired position at all

times.
Positive feedback Feedback that gets larger in response to the input, result-
ing in amplification.
Proprioception The process of sensing the state of one’s own body.
Pulse-width modulation Determining the duration of the signal based on

the width (duration) of the pulse.
Putting through Having the learner experience the task directly, through its
own sensors.
Quadrature shaft encoding A mechanism for detecting and measuring di-

rection of rotation.
Reactive systems Systems that do not use any internal representations of

the environment, and do not look ahead at the possible outcomes of their
actions; they operate on a short time-scale and react to the current sensory
information.
Reactive control A means of controlling robots using a collection of priori-

tized rules, without the use of persistent state or representation.
Reconfigurable robots Robots that have modules or components that can

come together in various ways, creating different shapes and differently
shaped bodies.
Redundancy The repetition of capabilities within a system.

299
Redundant Robot or actuator that has more ways of control than the degrees
of freedom it has to control.
Rehabilitation robots Robots that provide hands-on help by moving the

parts of the body of the patient in order to guide prescribed exercises and
recovery.
Representation The form in which information is stored or encoded in the

robot.
Resolution The process of separating or breaking something into its con-

stituent parts.
Robustness The ability to resist failure.
Rotational DOF The possible ways in which a body can rotate (turn).
Segmentation The process of dividing up or organizing the image into parts

that correspond to continuous objects.
Sensing The process of receiving information about the world through sen-
sors.
Sensor fusion Combining multiple sensors to get better information about

the world.
Sensor preprocessing Processing that comes before anything else can be done
in terms of using the data to make decisions and/or take action.
Sensor-actuator networks Groups of mobile sensors in the environment that

can communicate with each other, usually through wireless radio, and can
move around.
Servo motor A motors that can turn its shaft to a specific position.
Signal to symbol The problem of going from the output of a sensor to an

intelligent response.
Situated To exist in a complex word and to interact with it.
Situated automata Abstract (not physical) computing machines with par-

ticular mathematical properties that interact with their (abstract) environ-
ments.
300 Glossary
Smoothing Application of a mathematical procedure called convolution, which

finds and eliminates isolated peaks in the signal/data.
Social gaze Making eye contact in social and inter-personal interaction.
Socially assistive robots Robots that are capable of monitoring their users
and providing training, coaching, motivation, and encouragement with-
out physical contact.
Sonar Ultrasound (from SOund Navigation And Ranging).
Space All possible values or variations of something.
Spatial credit assignment The general problem of assigning the credit or

blame to actions taken by members of a team.
Specular reflection The reflection from the outer surface of the object. The
sound wave that travels from the emitter bounces off multiple surfaces in
the environment before returning to the detector.
State The description of a system. A general notion from physics that is bor-
rowed by robotics (and by computer science and AI, among other fields).
State estimation The process of estimating the state of a system from mea-
surements.
State space Consists of all possible states a system can be in.
Static stability Being able to stand without having to perform active control
to prevent from falling over.
Statically stable walking A robot that can walk while staying balanced at
all times.
Statistical neural networks A set of formal mathematical techniques from

statistics and probability applied to neural networks.
Stereo vision The ability to use the combined points of view from the two
eyes or camers to reconstruct three-dimensional solid objects and to per-
ceive depth. Formally called binocular stereopsis.
Stigmergy The form of communication in which information is conveyed

through changing the environment.
301
Supervised learning An approach to learning in which an external teacher

provides the answer or at least tells the robot what it did wrong.
Tele-operated Operated from afar (from the Greek tele meaning “far”).
Temporal credit assignment The general problem of assigning the credit or

blame to actions taken over time.
Time series A sequence of data provided over time.
Topological map A map represented as a collection of landmarks connected

with links.
Torque control An approach to control which tracks the desired torque at

all times, regardless of its specific motor shaft position.
Trajectory and motion planning A computationally complex approach that

which involves searching through all possible trajectories and evaluating
them, in order to find one that will satisfy the requirements.
Transducer A device that transforms one form of energy into another.
Translational DOF DOF that allows the body to translate, meaning move
without turning (rotation).
Tripod gait A statically stable gait in which three legs stay on the ground,
forming a tripod, while the other three lift and move.
Uncertainty The inability to be certain, to know for sure, about the state of
itself and its environment, in order to take absolutely optimal actions at
all times.
Uncontrollable DOF The DOF that are not controllable.
Universal plan A set of all possible plans for all initial states and all goals
within the state space of a particular system.
Vestibular occular reflex (VOR) The reflex that keeps the eyes fixed while
the head is moving.
Wikipedia A great ever-growing free encyclopedia found on the Web at

http://en.wikipedia.org/wiki/ that has a lot of information about every-
thing, including various topics covered in this book.
Index
Action Selection, 167 Behavior-Based, 141, 187–205

Actuator, 24, 29, 40 Centralized, 248
Analog, 10 Deliberative, 141, 151–159
Analog Electronic Circuit, 10 Derivative, 128–129
Anthropomorphic, 64 Distributed, 248
Arbitration, 207 Feedback, 121
Artificial Intelligence, 13 Feedforward, 131–132
Artificial Life, 11 Hybrid, 141, 177–185
Assistive Robotics, 280 Integral, 129–130
Autonomous, see Robot PID, 130–131
Autonomy, 26 Proportional, 126–127
Reactive, 141, 161–176
Backlash, 36 Control Architectures
Behavior coordination, 207–213 Herbert, 172–173
Behavior Fusion, 209 Sense-Plan-Act, 156
Biomimetic, 9, 108 Subsumption, 169–172
Bit, 71 Universal Plan, 183
Braitenberg’s Vehicles, 11 Control Policy, 258
Control Theory, 7–8
Calibration, 89 Controllers, 26
Cartesian Robots, 94 Cybernetics, 8, 11
Center of Gravity, 48
Command Arbitration, 167 DC Motor, 32–34
Command Fusion, 167 Degrees of Freedom, 25, 39, 48
Communication, 243 Controllable DOF, 40
Compliance, 66 Uncontrollable DOF, 40
Computation, 74 Demodulator, 90
Connectionist Learning, 260 Differential Drive, 54
Continuous state, 22 Digital, 95
Control Discrete state, 22
Architecture, 136 Dynamics, 66
304 Index
Educational Robotics, 284 Imitation, 262–265

Effector, 24, 29, 40 Lifelong, 267
Embodiment, 20 Reinforcement, 256–260
Emergent Behavior, 10, 216, 219 Supervised, 260–261
Endeffector, see Manipulation Light
Excitatory Connection, 12 Ambient, 88
Exoskeleton, 62 Infra Red, 89–90
External state, 23 Modulated, 90
Exteroception, 69 Polarized, 86
Visible, 89
Foraging, 234 Light-Emitting Diode (LED), 87
Linear Actuator, 32
Gait, 51
Localization, 225–227
Alternating Tripod Gait, 52
Locomotion, 24, 47
Ripple Gait, 53
Tripod Gait, 52 Machine Vision, 107
Gears, 34–35 Manipulation, 59
Ganged, 37 Endeffector, 59
Series, 37 Free Space, 60
Gedanken Experiments, 11 Grasp Points, 66
Links, 59
Haptics, 276
Problems, 65
Hidden State, 22
Motion Planning, 55
Holonomic, 42, 54
Human Information Processing, 279
Nanorobotics, 276
Human-Robot Interaction, 113, 263, 278
Natural Language Processing, 279
Navigation, 223–232
Image Plane, 109
Coverage Problem, 224
Inhibitory Connection, 12
Neural Networks, 260
Internal State, 23
Niche, 21
Inverse Kinematics, 65
Nonholonomic, 42
Inverse Pendulum, 50
Inverse Problem, 249
Observable, 22
Joint Limit, 60 Odometry, 225
Joints, 62
Partially Observable, 22
Ball-and-Socket, 42
Passive Actuation, 30
Prismatic, 62
Passive walker, 30–31
Rotary, 62
Path Integration, 225
Kidnapped Robot Problem, 202 Perception, 21
Kinematics, 62 Perceptual Space, 23
Perceptual System, 69
Learning, 255–268 Photocell, 84
Index 305
Photophilic, 12 Emitter, 81
Photophobic, 12 Laser, 104–107
Pitch, 39 Light, 84–94
Pixel, 71 Limit, 83
Polarizing Filter, 86 Passive, 81
Position Control, 38 Potentiometers, 95
Power issues, 25–26 Reflectance, 87–89
Proprioception, 69 Shaft Encoders, 83, 91–94
Proxemics, 279 Simple, 72
Pseudocode, 26 Sonar, 97
Pulse-Width Modulation, 38 Service Robots, 280
Servo Motors, 37
Quadrature Shaft Encoding, 93 Shakey, 14
Signal Processing, 73
Reactive Control, 10
Signal-to-Symbol, 73
Reconfigurable Robotics, 276
Situated Automata, 183
Redundant, 42
SLAM, 229–230
Rehabilitation Robots, 281
Social Gaze, 279
Remote Control, 60
Socially Assistive Robotics, 281
Representation, 23
Spatial Credit Assignment, 260
Retina, 72
Specular Reflection, 101–104
Robot
Stability, 48
Autonomous, 2
Dynamic, 50–51
History, 1, 7, 9
Polygon of Support, 48–50
Simulation, 3
Static, 48
What is, 2–5
Robotics, 5 Statically Stable Walking, 50
Roll, 39 State, 21
Rotational DOF, 39 State Estimation, 225
State Space, 22
Self-Assembly, 276 Stigmergy, 243
Sensing, 21
Sensor Fusion, 78 Teleoperation, 60
Sensor Space, 23 Robot, 2
Sensor-Actuator Networks, 235 Robot-Assisted Surgery, 61
Sensors, 4, 21, 23, 70 Temporal Credit Assignment, 259
Preprocessing, 73 Topological Map, 147
Active, 81 Torque, 35
Break Beam, 87 Torque Control, 38
Camera, 108–110 Trajectory, 41
Complex, 72–73 Optimal, 56
Contact, 83 Planning, see Motion Planning
Detector, 81 Transducer, 99
306 Index
Translational DOF, 39
Types of Robot Control, 16
Ultrasound, 97
Uncertainty, 70
Vision
Blob Tracking, 117
Edge Detection, 110
Model-Based, 112
Motion, 113
Stereo, 114
Vision-Based Navigation, 15
Waveform, 38
Yaw, 39

The Robotics Primer (Maja Mataric) - 2007

Uploaded by

Copyright:

Available Formats

The Robotics Primer (Maja Mataric) - 2007

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Robotics Primer (Maja Mataric) - 2007

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What topics does the book cover?

What topics does the book cover?

The

Illustrations by Nathan Koenig

Typeset in 10/13 Lucida Bright by the authors using LATEX 2ε .

Library of Congress Cataloging-in-Publication Information

2 Where Do Robots Come From? 7

4 Arms, Legs, Wheels, Tracks, and What Really Drives Them 29

7 What’s Going On? 69

8 Switch on the Light 81

9 Sonars, Lasers, and Cameras 97

9.3.7 Biological Vision 116

10 Stay in Control 121

11 The Building Blocks of Control 135

12 What’s in Your Head? 145

13 Think Hard, Act Later 151

14 Don’t Think, React! 161

15 Think and Act Separately, in Parallel 177

15.3 Avoiding Replanning 181

16 Think the Way You Act 187

17 Making Your Robot Behave 207

18 When the Unexpected Happens 215

19 Going Places 223

20 Go, Team! 233

20.5.2 Work It Out as a Team: Distributed Control 248

21 Things Keep Getting Better 255

Upon arriving at USC in 1997 as an assistant professor in Computer Science, I

compelling case for studying robotics to a wide variety of readers, because

Welcome to The Robotics Primer! Congratulations, you have chosen a very

A robot is an AUTONOMOUS system

Figure 1.1 An example of a robot.

erated; tele means “far” in Greek, so teleoperation means operating a system

• Robotics is a fast-growing ﬁeld whose deﬁnition has been evolving over

• Robotics involves autonomy, sensing, action, and achieving goals, all in

Food for Thought

• Is HAL, from the movie 2001, the Space Odyssey, a robot?

Looking for More?

• Robotics, The Marriage of Computers and Machines by Ellen Thro is another

• For a truly comprehensive review of modern robotics, complete with a

2 A Brief but Gripping History of

2.1 Control Theory

2.2.1 Grey Walter’s Tortoise

Let’s look at what Machina Speculatrix consisted of:

• One rechargeable battery

• Find the light

• Head toward the light

• Back away from bright light

• Turn and push to avoid obstacles

2.2.2 Braitenberg’s Vehicles

Figure 2.2 An example of Braitenberg’s vehicles, in simulation. (Photo courtesy of

2.3 Artiﬁcial Intelligence

• Internal models of the world

• Search through possible solutions

• Planning and reasoning to solve problems

• Symbolic representation of information

• Hierarchical system organization

• Sequential program execution.

S HAKEY So what is AI-inspired robotics? Shakey, shown in ﬁgure 2.3, is a good