Hindawi Publishing Corporation

Computational Intelligence and Neuroscience

Volume 2016, Article ID 5720163, 10 pages
http://dx.doi.org/10.1155/2016/5720163

Research Article
A New Artificial Neural Network Approach in Solving Inverse
Kinematics of Robotic Arm (Denso VP6242)

Ahmed R. J. Almusawi,1,2 L. Canan Dülger,1 and Sadettin Kapucu1

1
Mechanical Engineering Department, University of Gaziantep, Gaziantep, Turkey
2
Mechatronics Engineering Department, University of Baghdad, Baghdad, Iraq

Correspondence should be addressed to Ahmed R. J. Almusawi; [email protected]

Received 15 May 2016; Accepted 18 July 2016

Academic Editor: Dong W. Kim

Copyright © 2016 Ahmed R. J. Almusawi et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

This paper presents a novel inverse kinematics solution for robotic arm based on artificial neural network (ANN) architecture. The
motion of robotic arm is controlled by the kinematics of ANN. A new artificial neural network approach for inverse kinematics is
proposed. The novelty of the proposed ANN is the inclusion of the feedback of current joint angles configuration of robotic arm as
well as the desired position and orientation in the input pattern of neural network, while the traditional ANN has only the desired
position and orientation of the end effector in the input pattern of neural network. In this paper, a six DOF Denso robotic arm
with a gripper is controlled by ANN. The comprehensive experimental results proved the applicability and the efficiency of the
proposed approach in robotic motion control. The inclusion of current configuration of joint angles in ANN significantly increased
the accuracy of ANN estimation of the joint angles output. The new controller design has advantages over the existing techniques
for minimizing the position error in unconventional tasks and increasing the accuracy of ANN in estimation of robot’s joint angles.

1. Introduction were done on the implementation of artificial intelligence

on a robotic arm to overcome the singular configuration
Artificial intelligence has become the most modern technol- problem of robotic arm. The inverse kinematics of three
ogy of robotic control. It has many advantages in performance DOF robotic arm was solved by multilayer network inversion
such as precise control and less computing time, in addition method; the joint angles were estimated for given end effector
to overcoming some mathematical problems in motion and position in a simulation of three-link robotic arm. The results
path generation. The main problem of motion control in showed an approximation solution for inverse kinematics [1].
robotic arm is to find the accurate and reliable solution for Singularities and uncertainties in arm configurations are the
inverse kinematics. The calculation of inverse kinematics main complications in the kinematics of robot control, in
is necessary in real-time control; the solving of inverse order to have a realistic solution based on one of the heuristic
kinematics is computationally complex and requires a very methods; artificial neural network (ANN) was suggested for
long processing time. Most applications of motion in robotic a nonsurgical robot. The main idea of this approach was the
manipulation require Cartesian space motion control. In use of ANN to learn the robot system characteristics rather
inverse kinematics, desired position and orientation of the than having to specify an explicit robot system model [2]. A
end effector in Cartesian space are given while the set of neural network and genetic algorithms were used together
robot’s joint angles in joint space is calculated. to solve the inverse kinematics problem of the nonsurgical
In general, the solutions of inverse kinematics of a robotic robotic manipulator to minimize the error at the end effector
manipulator are geometric, iterative, analytic, or algebraic and improve the precision of the inverse kinematics solution
approaches. Recently, a high focus has been applied on [3]. The inverse kinematic of redundant manipulators was
artificial intelligence based methods for inverse kinematics presented by neural networks (NNs) to obtain the joint angles
problem solution of general purpose robot. Many studies of the robot using the Cartesian coordinate of the end effector.
2 Computational Intelligence and Neuroscience

Position errors of end effector and feasibility of the joint used to solve the kinematics of the autonomous positioning
angles were obtained [4]. An artificial neural network was of a robotic arm. This modeling and analysis approach was
used for controlling 3 DOF robotic manipulator. The methods tested by using a five DOF arm with a gripper mounted to
introduced a nonlinear relation between Cartesian and joint the iRobot mobile platform [15]. Online robot kinematics
coordinates using multilayer perceptron in artificial neural parameter errors estimation based on inertial measurement
network. A simulation test was implemented [5]. A neural unit (IMU) was presented. It obtained the orientation of the
network architecture was introduced to solve the inverse manipulator with the orientation of the IMU in real time.
kinematics problem for robotics manipulators with two This approach incorporated Factored Quaternion Algorithm
degrees of freedom. The neural networks were multilayered (FQA) and Kalman Filter (KF) to the orientation of the
perceptron (MLP) with a backpropagation training algorithm IMU [16]. An analytical solution of inverse kinematics for a
for reducing the complexity of the algorithm and calculation five DOF spatial parallel micromanipulator was presented. A
(matrix inversion) of inverse geometric of robotic arm. The geometrical mode and structural of system were introduced
result showed a mean squared error (MSE) in performance for the microrobot’s task [17]. Forward kinematics and inverse
near to 10−5 [6]. A high-order-logic theorem was used for kinematics were calculated and simulation was done for
solving the kinematics analysis of six-axis revolute joint joints and link parameters of six-axis robotic arm. Trajectory
robot. The approach required an enormous amount of user planning was described for the requisite motion of the
intervention to overcome the limitation of kinematic analysis manipulator as a time sequence task [18]. Forward and inverse
[7]. The inverse kinematics problem of the 6 DOF robot kinematics of a KUKA robotic arm in the application of
is solved by using curved-surface scanning to carry the a simple welding process were introduced. A general DH
ultrasonic testing task. Many results for the joint angles were representation of forward and inverse matrix was obtained.
acquired; the method of the shortest distance was assumed A movement flow planning was designed and developed for
to solve the inverse problem of the robot system. A 3D the programming of the robot [19]. The mobile robot with
application software was introduced to simulate the motion arm (KUKA youBot) and the solving of inverse kinematics
of ultrasonic trajectory and path planning [8]. The kine- problem were introduced. The robot was presented as 8
matics and singularities of an asymmetrical parallel robotic DOF. The kinematics redundancy of the holonomic platform
wrist were investigated by using the method of Lagrange was presented. Including redundancy parameters, the inverse
multipliers and considering all the mobile components. kinematics solution was suggested [20]. The end effector
The designed model was numerically illustrated to show position and orientation error of a space robot were studied.
its computation accuracy [9]. A technique for solving the A geometric parameter identification method was presented
inverse kinematics problem using artificial neural networks based on a laser ranger attached to the end effector. The
was introduced for a PUMA 560 robot. An inverse kine- independence of the geometric parameters was analyzed.
matic solution was studied by training the neural network Identification equations were derived by simulation which
with the robot’s end effector Cartesian coordinates and its was implemented for different types of robot configuration
corresponding joint configurations. Results showed mean [21].
square error (MSE) of 1.2178, the regression value obtained This paper introduces a novel solving method for six-
was 0.87527, and the position errors in 𝑥-, 𝑦-, and 𝑧-axis axis manipulator robot based on ANN to be used in motion
were 4.93%, 7.29%, and 3.73%, respectively [10]. A solution control. It is clear from previous survey that no study has
to the inverse kinematics was required for generating desired included current joint angles of robot in their ANN. In this
trajectories in the Cartesian space (2D). A feedforward neural study, the ANN architecture has included current joint angles
network was used for planar of three-link manipulators. of robot in the input pattern and it improves the performance
The result showed the best performance at epoch 9 with of proposed ANN in solving inverse kinematics. It is the first
mean squared error (MSE) of 0.0054387 [11]. The kinematics ANN that fulfils the requirement of robot precise motion and
of three DOF was introduced for the lower limb of the reduces joint angles error and outcomes in some aspects of
humanoid robot. Decoupled closed-form solution for the the robot tasks.
position and orientation was the solution of kinematics; This paper is organized as follows. Section 2 presents the
the joint sequences were presented by Denavit-Hartenberg kinematics analysis and the required parameters for motion
(DH) transformation matrices. Swing phase equations were control followed by an explanation of problem formula-
developed to avoid matrix inversion problems [12]. The tion. The proposed artificial neural network is described in
kinematic parameters on industrial robot were affected by Section 3. Section 4 illustrates the system setup. Section 5
vibrations disturbance. The error in motion was improved presents the experimental work and discussion of results.
by sensors of accelerometer and gyroscope. The motion Finally, Section 6 concludes this paper.
profile was analyzed for joint; then, the path tracking of
welding task was estimated [13]. Human robot kinematics was
identified by geometry kinematics approach to map human 2. Kinematics Analysis
arm configuration and stiffness controlled index by hand
gesture. The human arm stiffness was estimated within robot The kinematics of serial manipulator describe the relation-
experiential stability region. A moving task was implemented ship between the joint angles and the position and orientation
to test the performance of geometry kinematics approach of its end effector. The kinematics of robot is required in tra-
on Baxter robot simulator [14]. The geometric approach was jectory generation and motion control. The transformation
Computational Intelligence and Neuroscience 3

Table 1: DH parameters of the Denso robot.

Y3 Link 𝑖 𝜃𝑖 𝑑𝑖 𝑎𝑖 𝛼𝑖
Y2 1 𝑞1 0.125 0 pi/2
Z3 2 𝑞2 0 0.21 0
Y4
X2 X3 Z4 3 𝑞3 0 −0.075 −pi/2
Z2
X4 4 𝑞4 0.21 0 pi/2
Z5 5 𝑞5 0 0 −pi/2
Y1
X5 6 𝑞6 0.07 0 0
Y5
Z1 X1

Z0 Y0 Y6 X6

X0 The system has six links and a gripper. The homogeneous

Z6 transformation matrix is calculated by multiplication of
matrices as follows:

0
𝑇gripper = 𝐴 1 𝐴 2 𝐴 3 𝐴 4 𝐴 5 𝐴6 𝐴 gripper , (3)
Figure 1: Reference coordinates for the system.

where
matrices have been used for control. The robotic system is the
Denso robot with 6 revolute joints. The kinematics analysis 1 0 0 0
is done after system coordinate frame has been performed, [ ]
the coordinates 𝑂0 , 𝑥0 , 𝑦0 , 𝑧0 are fixed to the base which is the [0 0 −1 0 ]
𝐴1 = [ ]
[0 1 0 .125] ,
base frame. The other coordinate frames are attached to the [ ]
corresponding links. The reference coordinates of the system
are shown in Figure 1. [0 0 0 1 ]
The homogeneous transformation matrix is stated to 1 0 0 .21
represent the position and orientation of end effector with [ ]
respect to base coordinate; a homogeneous transformation [0 1 0 0 ]
𝐴2 = [
[0 0 1 0 ] ,
]
matrix 𝑇60 for overall system is as follows: [ ]
[0 0 0 1 ]
𝑅60 𝑃60
𝑇60 = [ ], (1) 1 0 0 −.075
0 1 [ ]
[0 0 1 0 ]
[
𝐴3 = [ ],
]
where 𝑅60 is a rotation matrix 3 × 3 and 𝑃60 is a position vector [0 −1 0 0 ]
of the end effector in the base frame coordinate. The Denavit- [0 0 0 1 ]
Hartenberg DH method is used to analyze the kinematics (4)
of Denso robot. The robot transformation matrix has been 1 0 0 0
denoted. The single link homogenous transformation matrix [ ]
[0 0 −1 0 ]
𝐴 𝑖 is 𝐴4 = [ ]
[0 1 0 .21] ,
[ ]
𝐴 𝑖 = Rot𝑧,𝜃𝑖 Trans𝑥,𝑑𝑖 Trans𝑥,𝛼𝑖 Rot𝑥,𝛼𝑖 , [0 0 0 1]

𝐶𝜃𝑖 −𝑆𝜃𝑖 𝐶𝛼𝑖 𝑆𝜃𝑖 𝑆𝛼𝑖 𝑎𝑖 𝐶𝜃𝑖 1 0 0 0

[ ] [ ]
[ 𝑆𝜃𝑖 𝐶𝜃𝑖 𝐶𝛼𝑖 −𝐶𝜃𝑖 𝑆𝛼𝑖 𝑎𝑖 𝑆𝜃𝑖 ] (2) [0 0 1 0]
[
𝐴𝑖 = [ ], 𝐴5 = [ ]
[0 −1 0 0] ,
[ 0 𝑆𝛼𝑖 𝐶𝛼𝑖 𝑑𝑖 ] ] [ ]
[ 0 0 0 1 ] [0 0 0 1]
1 0 0 0
where 𝑖 is link number, 𝑆𝜃𝑖 = sin 𝜃𝑖 , 𝐶𝜃𝑖 = cos 𝜃𝑖 , 𝜃𝑖 is the joint [ ]
rotation angle, 𝑎𝑖 is the length of links, 𝛼𝑖 is the twist angles, [0 1 0 0 ]
[
𝐴6 = [ ].
𝑑𝑖 is the link offsets, and 𝜃 is the joint angles. ]
[0 0 1 .17]
The Denavit-Hartenberg HD parameters of the robot are
shown in Table 1. [0 0 0 1 ]
4 Computational Intelligence and Neuroscience

geometry and the nonlinear trigonometric equations that

0.6
describe the relationship between the Cartesian space and the
6
0.4 4 joint space. Although a closed-form solution to this problem
5 3
0.2 x is preferable in robotics, sometimes it is impossible to find.
Z (m)

y 2 Denso robot
0 z
Therefore, various other ways to determine the solution for
−0.2 1
−0.4 inverse kinematics problem were studied such as geometrical
−0.6 solutions and numerical algorithms. This task depends on the
−0.6−0.4 −0.4 designed structure of the robot while many robots such as
−0.2 0 0 −0.2 redundant manipulators do not have an analytical solution
0.2 0.2
Y (m 0.4 0.6 0.6 0.4 )
) X (m for the inverse kinematics.
In this study, two ANNs are designed for inverse kinemat-
Figure 2: The graphical representation for the robotic system.
ics of robotic arm. The first one is the traditional ANN as used
in serial robotics inverse kinematics analysis, and the second
is the proposed ANN by considering the feedback of current
The transformation matrix of the gripper is 𝐴 gripper ; the joint robot configuration (current joint angles) in the design of
angles and gripper transformation matrix are given; then, the ANN.
transformation matrix 𝑇Home of home position is calculated:

−1 0 0 .21 3. Traditional Design of Artificial

[ ] Neural Network
[0 1 0 0]
𝑇Home = [ ]
[ 0 0 −1 .24] . (5)
[ ] A traditional design for ANN is used in many studies [6,
10, 11]. In order to utilize the advantages of this proposed
[0 0 0 1] method, traditional ANN is designed in this study to solve
the inverse kinematics. In this ANN, the elements in the
The graphical representation of Denso robot is done by MAT-
input layer are six variables, which are the position 𝑃
LAB programming robotics toolbox as shown in Figure 2.
and orientation 𝑅 of gripper in Cartesian coordinates. The
The inverse kinematics might got several solutions pro-
number of hidden layers is ten. The output layer has six
duced for each of the joint angles because these are corre-
elements of the angles of joint 𝑄. MATLAB/neural network
sponding to many robot configurations such as elbow up,
toolbox is used for training, validation, and testing. Figure 3
elbow down, wrist up, wrist down, shoulder forward, and
shows a block diagram for traditional ANN and its model as
shoulder back. The position and orientation of end effector
follows:
are obtained by forward kinematics as follows:
[𝑄] = ANN Traditional Net (𝑃, 𝑅) . (9)
𝑇60 = forward kinematics [𝑞] . (6)
The inputs are uniformly enclosed with the workspace of
The joint angles are calculated by inverse kinematic for specified position; the corresponding inputs/outputs are
desired position/orientation of end effector: computed by solution of forward kinematics. In this way, each
position of the robot has a unique joint configuration in the
[𝑞] = inverse kinematics (𝑇60 ) . (7) neural network inputs/outputs set.
The training algorithm is the Levenberg-Marquardt back-
In this study, a new artificial neural network solution for propagation; it is used to assure fast convergence of the train-
inverse kinematics of (8) is introduced in Section 3: ing error and is also a very popular curve-fitting algorithm.
Figure 4 shows the performance of the traditional neural
𝑞1 𝑝𝑥 network; the MSE of training is decreased until the validation
[ ] error is stopped at epoch 121 and MSE was 1.1892𝑒−5 in the
[𝑞2 ] 𝑝𝑦
[ ]
[ ] ( ) best performance. A closed relation was between the output
[𝑞3 ] (𝑝 )
[ ] = inverse kinematics ( 𝑧 ) . (8) and target samples at correlation of 0.99758.
[𝑞 ] (𝜃 )
[ 4] ( 𝑥)
[ ]
[𝑞 ] 𝜃𝑦 4. Proposed Artificial Neural Network Design
[ 5]
[𝑞6 ] ( 𝜃𝑧 ) Always, robot starts motion from current position, and in
most applications robot moves on trajectory on sequential
2.1. The Problem Statement. In general, the desired motion point’s path. So the inclusion of current joint configuration
of robot is carried out in the Cartesian coordinate, while in ANN has a positive effect in the estimation of joint angles
the robotic arm motion is controlled by joint coordinate; for the next desired position. In this paper, a novel neural
a solution for the inverse kinematics is very important to network design is proposed and used to solve the inverse
be calculated. Solving the inverse kinematics problem for kinematics problem of robotic arm. The proposed method
robotic manipulators is a difficult and also quite challenging relies on the constraints of the kinematics of robotic arm
task. The difficulty of this problem is given by the robot’s to achieve robot’s motion in an intelligent way with high
Computational Intelligence and Neuroscience 5

Robotic arm motion

T
Joint Driving
Position Trained ANN angles unit
Orientation without Qc Qc

Figure 3: Block diagram of robot motion control by traditional ANN.

100 The performance of the neural network was determined

Mean squared error (MSE)

based on the mean squared error (MSE) between the neu-

10−1 ral network’s actual output and the desired output. The
Best validation performance is
performance of the proposed ANN is shown in Figure 6.
10−2 1.1892e − 05 at epoch 121
The differences between the network outputs and target are
10−3 calculated through the mean squared error (MSE); it drops
rapidly through the learning process; the MSE of training
10−4 is decreased until the validation error is stopped at epoch
68 and MSE was 3.3029𝑒−8 in the best performance. A
10−5 closed relationship was between output and target samples
0 20 40 60 80 100 120
Epochs at correlation 0.99999. The new outputs of the network are
checked by the test data, the training samples were 2800, and
Training Test the validated and test samples were 600 for each one.
Validation Best

Figure 4: The performance of the traditional neural network.

5. System Setup
Denso robot VP6242 is a six-axis industrial robot (Quanser
accuracy in position. The contribution of proposed design is Company). Artificial neural network is implemented on this
considering the current joint angles of the robot in solving system for predicting joint angles during real-time Cartesian
the inverse kinematics of robot by ANN. The inclusion of motion. The robot is communicated with MATLAB/Simulink
the current joint angles during training produces a strong via TCP/IP. An open source program (QUARC) control
network and adjusts the weights with very low error. software is supported; the QUARC software is executed in
The forward kinematics is used to generate the input/ Simulink for real-time application. A blockset is used to
target data set which is used in training; the inputs of neural connect Simulink program with Denso driving unit. The joint
network are the desired position/orientation and current positions and joint currents (ampere) are reading on PC while
robot joint configuration while the targets are the required the joint velocities/positions are sending to robot. The joint
joint angles of the robot relative to those points. The proposed PID parameters or joint feedforward gains are adjustable and
neural network has 13 elements in the input layer, which user can deal with them. The total arm length is 420 mm, and
are the gripper position 𝑃 and orientation 𝑅 in Cartesian the payload is 2.5 kg. Figure 7 shows the control parameters
coordinates and current joint angles 𝑄𝑐 of the robot. The between robot and QUARC software SW.
output layer has six elements of joint coordinates, which are The gripping system is an electrical drive gripper WSG 32
the joint angles 𝑄 of the robot. Figure 5 shows a block diagram SCHUNK; it is a very precise handling system for medium
for proposed ANN, the inputs, and outputs of neural network parts weight up to 0.5 kg, and it has integrate high-sensitivity
as follows: sensor to detect parts in a gripping force of 5–50 N and
opening of up to 65 mm. The gripper is required for pick-and-
place part or other handling activities. Figure 8 shows robotic
[𝑄] = ANN Proposed Net (𝑃, 𝑅, 𝑄𝑐 ) ,
system full setup.
𝑄 = [𝑞1 , 𝑞2 , 𝑞3 , 𝑞4 , 𝑞5 , 𝑞6 ] , A virtual model is connected with Simulink to visualize
the system by using Simulink 3D animation. The virtual
𝑃 = [𝑥, 𝑦, 𝑧] , (10) model and the system model are combined to create a
virtual robotic system environment. This virtual model can
𝑅 = [𝑅𝑥 , 𝑅𝑦 , 𝑅𝑧 ] , record and present the motion path of real robot during
experiments. Figure 9 shows the virtual model of robotic
𝑄𝑐 = [𝑞1𝑐 , 𝑞2𝑐 , 𝑞3𝑐 , 𝑞4𝑐 , 𝑞5𝑐 , 𝑞6𝑐 ] . system.
6 Computational Intelligence and Neuroscience

Robotic arm motion

Trocar position Joint Driving
Trained ANN unit
with Qc angles
Orientation Qc

Feedback of current joint angles

Figure 5: Block diagram of robot motion control by proposed ANN.

100
Mean squared error (MSE)

10−1
Best validation performance is
10−2
−3
3.3029e − 08 at epoch 68
10
10−4
10−5
10−6
10−7
10−8
0 10 20 30 40 50 60 70
Epochs

Training Test Figure 9: Virtual model of the robotic system.

Validation Best

Figure 6: The performance of the proposed neural network.

6. Experiment Results and Discussion

Joint velocity/position
PID parameters
After the training of proposed ANN and traditional ANN is
completed, the experiments of movement are carried out by
Joint position the robotic system. The gripper motion is in a helical path
QUARC SW Joint current within the workspace area. The input path for gripper tip
Controller and Denso robot
is generated by 𝑃𝑥 (𝑘) = 𝑐1 sin (𝑘), 𝑃𝑦 (𝑘) = 𝑐2 cos (𝑘), and
Figure 7: The control parameters between robot and QUARC 𝑃𝑧 (𝑘) = 𝑐3 𝑘. In this experiment, the robot has to follow a
software. sinusoidal rotation in 𝑥-axis and 𝑦-axis while the motion
in 𝑧-axis is a linear path; the constants 𝑐1 and 𝑐2 are the
amplitude of radius on the 𝑥- and 𝑦-axis, respectively; the
constant 𝑐3 is the pitch of a helix, and 𝑘 is the sample time.
Denso robot Figure 10 shows the real robotic system and the virtual model
to demonstrate helical motion path by ANN. The inverse
kinematics solution for robot motion is achieved by proposed
ANN and traditional ANN.
The movements of robot are executed in circular and
Gripper linear motion. The pitch of path of the end effector is changed
by varying the position in 𝑧 direction relative to workspace
coordinates. The rotation of end effector is performed by
changing the position to 𝑥 and 𝑦 directions. Two paths are
generated by using proposed ANN and traditional ANN with
desired path. Figure 11 illustrates the paths configuration of
the end effector by the robot movement.
The motion of the robot is in a Cartesian space. In
Computer Figure 12(a), the motion in 𝑥 direction is an enlargement
program of the time period (20 s–25 s). Figure 12(b) is the motion
in 𝑦 direction enlarged for the time period (26 s to 30 s).
Figure 12(c) is the motion in 𝑧 direction enlarged for the
Figure 8: Robotic arm system setup. time period (14 s to 19 s). Figures mentioned above show the
Computational Intelligence and Neuroscience 7

Figure 10: Experiment on the robotic motion demonstration by ANN (movement of end effector through helical path).

Table 2: Performance of robotic system by ANNs.

0.25
Parameters Proposed ANN Traditional ANN
0.2 j
𝑃𝑥 error% 0.17 5.78
Z (m)

0.15 𝑃𝑦 error% 0.36 7.25

i
0.1 k 𝑃𝑧 error% 0.12 1.28
0.05 MSE 3.3029𝑒−8 1.1892𝑒−5
0.1 Regression 0.99999 0.99758
Y( 0 0.3
m) 0.25
−0.1 0.2
0.15 X (m)
According to the ANN based solving inverse kinematics
Desired path results in the literature [2–6, 10, 11], the proposed approach in
Proposed ANN this study has a minimized error in the inverse kinematics
Traditional ANN solution. Table 3 shows system performance comparison
Figure 11: Paths configuration generated by robot in two different between this study and other studies from literature. The high
ANNs. The selected points 𝑖, 𝑗, and 𝑘 are enlarged portions in accuracy and the low MSE that are obtained in this work can
Figure 12. be obviously seen.
The errors are increased in some points and are reduced in
other points; the point of high error is because of estimating
position by ANN while the point of low error is the point that
differences in 𝑥, 𝑦, and 𝑧 directions for motion resulting from is near the samples of training set in ANN.
proposed ANN and traditional ANN and the high accuracy The proposed ANN has given higher accuracy and preci-
and precision of proposed ANN. sion in position than the traditional ANN, and this method
The differences between desired path and artificial neural is applicable in precise robotic motion.
network path are measured by calculating the error in posi-
tion from proposed ANN and traditional ANN with respect 7. Conclusions
to position from desired path. Figure 13 shows the error in 𝑥,
𝑦, and 𝑧 position for proposed ANN, while Figure 14 shows This study introduced a very accurate solution for inverse
the errors in 𝑥, 𝑦, and 𝑧 position for traditional ANN. kinematics by using the artificial neural network to overcome
The motion curves of proposed ANN are much more the drawbacks of traditional ANN controller. A new design
precise than the motion curves of traditional ANN in esti- of artificial neural network ANN has been proposed for the
mating the joint angles of the robot for desired positions. The optimal robot motion control in Cartesian coordinates. In
maximum error in 𝑥 direction for proposed design of ANN order to evaluate the integral performance of the system, the
was near 0.22 millimeters and for traditional design of ANN2 current joint angles information was added to the traditional
was 6.5 millimeters. The maximum error in 𝑦 direction for ANN based inverse kinematics solution. The proposed design
proposed design of ANN was less than 0.3 millimeters and showed improvement in performance of end effector in some
for traditional design of ANN2 was near 6 millimeters. The aspects.
maximum error in 𝑧 direction for proposed design of ANN The motion of robot has been executed; it satisfies the
was 0.35 millimeters and for traditional design of ANN2 was constraint of robotic arm motion by the designed structure.
near 2.5 millimeters. Table 2 shows the error percentages of Robot kinematics are analyzed, and position/orientation of
proposed and traditional ANNs. end effector in different configurations are studied.
8 Computational Intelligence and Neuroscience

0.145
−0.06
0.14 −0.065
Motion in X direction (m)

Motion in Y direction (m)

0.135 −0.07
0.13 −0.075

0.125 −0.08
−0.085
0.12
−0.09
0.115
−0.095
0.11 −0.1
0.105 −0.105
21 22 23 24 25 26 27 28 29 30
Time (s) Time (s)

Desired path Desired path

Proposed ANN Proposed ANN
Traditional ANN Traditional ANN
(a) (b)
0.216
Motion in Z direction (m)

0.214

0.212

0.21

0.208

0.206
15 16 17 18
Time (s)

Desired path
Proposed ANN
Traditional ANN
(c)

Figure 12: Robot motion by two ANNs: (a) the motion in 𝑥 direction, (b) the motion in 𝑦 direction, and (c) the motion in 𝑧 direction.

Table 3: System performance comparison between this study and other studies mentioned in the literature.

Study System DOF Method MSE Hidden layers Errors

𝑥 = 0.17%
Proposed Denso 6 ANN 3.3𝑒−8 10 𝑦 = 0.36%
𝑧 = 0.12%
𝑥 = 6.42%
Luv et al., 2014 [10] PUMA 560 6 ANN 1.217 30 𝑦 = 4.90%
𝑧 = 2.92%
𝑥 = 3.34%
Hasan et al., 2010 [2] FANUC M-710i robot 6 ANN ∼1 1 𝑦 = 6.72%
𝑧 = 0.35%
Toshani and Farrokhi, 2014 [4] Simulation PA-10 robot 7 NNs with optimization 1 — End effector 5 mm
Duka, 2014 [11] Planar simulation 3 NNT 0.0054 1 —
Köker, 2013 [3] Stanford 6 ANN 2.38 25 End effector 4.28 mm
Nanda et al., 2012 [5] Simulation 3 ANN, FLANN >1 20 —
Daya et al., 2010 [6] simulation 2 NNT 5.24𝑒−5 2 —
Computational Intelligence and Neuroscience 9

×10−4 Acknowledgments
6
This study is supported by Scientific Research Projects Gov-
4 erning Unit (BAPYB) and Mechatronics Research Lab in
Error in position (m)

Gaziantep University. The authors would like to thank them

2
for their help and support.

0 References
[1] T. Ogawa and H. Kanada, “Solution for Ill-posed inverse
−2 kinematics of robot arm by network inversion,” Journal of
Robotics, vol. 2010, Article ID 870923, 9 pages, 2010.
−4 [2] A. T. Hasan, N. Ismail, A. M. S. Hamouda, I. Aris, M. H. Marha-
0 10 20 30 40 50
ban, and H. M. A. A. Al-Assadi, “Artificial neural network-
Time (s) based kinematics Jacobian solution for serial manipulator pass-
X error ing through singular configurations,” Advances in Engineering
Y error Software, vol. 41, no. 2, pp. 359–367, 2010.
Z error [3] R. Köker, “A genetic algorithm approach to a neural-network-
based inverse kinematics solution of robotic manipulators
Figure 13: Proposed ANN motion error in 𝑥, 𝑦, and 𝑧 direction.
based on error minimization,” Information Sciences, vol. 222, pp.
528–543, 2013.
×10−3
[4] H. Toshani and M. Farrokhi, “Real-time inverse kinemat-
ics of redundant manipulators using neural networks and
10
quadratic programming: a Lyapunov-based approach,” Robotics
8 and Autonomous Systems, vol. 62, no. 6, pp. 766–781, 2014.
6 [5] S. K. Nanda, S. Panda, P. R. Subudhi, and R. K. Das, “A
Error in position (m)

4 novel application of artificial neural network for the solution of

2 inverse kinematics controls of robotic manipulators,” Interna-
tional Journal of Intelligent Systems and Applications, vol. 4, no.
0
9, pp. 81–91, 2012.
−2
[6] B. Daya, S. Khawandi, and M. Akoum, “Applying neural net-
−4 work architecture for inverse kinematics problem in robotics,”
−6 Journal of Software Engineering & Applications, vol. 3, no. 3, pp.
−8
230–239, 2010.
0 10 20 30 40 50 [7] A. Wu, Z. Shi, Y. Li et al., “Formal kinematic analysis of a general
Time (s) 6R manipulator using the screw theory,” Mathematical Problems
in Engineering, vol. 2015, Article ID 549797, 7 pages, 2015.
X error
Y error [8] Z. Lu, C. Xu, Q. Pan, X. Zhao, and X. Li, “Inverse kinematic
Z error analysis and evaluation of a robot for nondestructive testing
application,” Journal of Robotics, vol. 2015, Article ID 596327, 7
Figure 14: Traditional ANN motion error in 𝑥, 𝑦, and 𝑧 direction. pages, 2015.
[9] G. Wu, “Kinematics and dynamics of an asymmetrical parallel
robotic wrist,” Journal of Robotics, vol. 2014, Article ID 296250,
13 pages, 2014.
A comparison has been carried out between two ANNs.
The parameters of motion and the errors were calculated. [10] A. Luv, A. Kush, and J. Ruth, “Use of artificial neural networks
The results showed that the proposed ANN has superior for the development of an inverse kinematic solution and visual
performance in terms of the joint angles estimation. The identification of singularity zone(s),” Procedia CIRP, vol. 17, pp.
812–817, 2014.
design of ANN compared to other techniques is applicable
for some of the most difficult and challenging problems of [11] A. V. Duka, “Neural network based inverse kinematics solution
kinematics. These results have proved the effectiveness of for trajectory tracking of a robotic arm,” Procedia Technology,
vol. 12, pp. 20–27, 2014.
the proposed ANN. The inclusion of current configuration
of joint angles in ANN increased the accuracy of ANN [12] A. Said, E. Rodriguez-Leal, R. Soto, J. L. Gordillo, and L.
estimation and succeeded in mapping between input position Garrido, “Decoupled closed-form solution for humanoid lower
limb kinematics,” Mathematical Problems in Engineering, vol.
and joint angles output.
2015, Article ID 437979, 14 pages, 2015.
[13] D. A. Elvira-Ortiz, R. de Jesus Romero-Troncoso, A. Y. Jaen-
Competing Interests Cuellar, L. Morales-Velazquez, and R. A. Osornio-Rios, “Vibra-
tion suppression for improving the estimation of kinematic
The authors declare that there is no conflict of interests parameters on industrial robots,” Shock and Vibration, vol. 2016,
regarding the publication of this paper. Article ID 6954012, 15 pages, 2016.
10 Computational Intelligence and Neuroscience

[14] P. Liang, L. Ge, Y. Liu, L. Zhao, R. Li, and K. Wang, “An

augmented discrete-time approach for human-robot collabora-
tion,” Discrete Dynamics in Nature and Society, vol. 2016, Article
ID 9126056, 13 pages, 2016.
[15] K. E. Clothier and Y. Shang, “A geometric approach for robotic
arm kinematics with hardware design, electrical design, and
implementation,” Journal of Robotics, vol. 2010, Article ID
984823, 10 pages, 2010.
[16] G. Du and P. Zhang, “IMU-based online kinematic calibration
of robot manipulator,” The Scientific World Journal, vol. 2013,
Article ID 139738, 10 pages, 2013.
[17] D. Prusak, K. Kobus, and G. Karpiel, “Kinematics analysis of
a novel five-degree-of-freedom spatial parallel micromanipula-
tor,” Journal of Robotics, vol. 2014, Article ID 806294, 13 pages,
2014.
[18] K. Kishore, A. Srinath, G. Jugalanvesh, R. Premsai, and M.
suresh, “Kinematic analysis and simulation of 6Dof KukaKr5
robot for welding application,” International Journal of Engi-
neering Research and Applications, vol. 3, no. 2, pp. 820–827,
2013.
[19] M. Dahari and J.-D. Tan, “Forward and inverse kinematics
model for robotic welding process using KR-16KS KUKA
robot,” in Proceedings of the 4th International Conference on
Modeling, Simulation and Applied Optimization (ICMSAO ’11),
Kuala Lumpur, Malaysia, April 2011.
[20] S. Sharma, G. K. Kraetzschmar, C. Scheurer, and R. Bischoff,
“Unified closed form inverse kinematics for the KUKA youBot,”
in Proceedings of the 7th German Conference on Robotics, pp. 1–
6, Munich, Germany, May 2012.
[21] Y. Liu, Z. Jiang, H. Liu, and W. Xu, “Geometric parameter
identification of a 6-DOF space robot using a laser-ranger,”
Journal of Robotics, vol. 2012, Article ID 587407, 11 pages, 2012.