See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/359324218

Solving Kinematics of a Parallel Manipulator Using Artificial Neural Networks

Conference Paper · March 2022

DOI: 10.1109/ICCTA54562.2021.9916234

CITATION READS

1 172

3 authors:

Yasmin Khattab Mohamed El-Habrouk

Arab Academy for Science, Technology & Maritime Transport Alexandria University
4 PUBLICATIONS   4 CITATIONS    46 PUBLICATIONS   1,065 CITATIONS   

SEE PROFILE SEE PROFILE

Sohair Rezeka
Alexandria University
73 PUBLICATIONS   275 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Renewable Energy View project

Load Balancing View project

All content following this page was uploaded by Sohair Rezeka on 18 March 2022.

The user has requested enhancement of the downloaded file.

Solving Kinematics of a Parallel Manipulator Using
Artificial Neural Networks
Yasmin Khattab Iham F. Zidane
Mechanical Engineering Department Mechanical Engineering Department
College of Engineering and Technology, Arab Academy of College of Engineering and Technology, Arab Academy of
Science Technology and Maritime Transport, AASTMT Science Technology and Maritime Transport, AASTMT
Alexandria, Egypt Alexandria, Egypt
[email protected] [email protected]

Mohamed El-Habrouk Sohair Rezeka

Electrical Engineering Department Mechanical Engineering Department
Faculty of Engineering, Alexandria University Faculty of Engineering, Alexandria University
Alexandria, Egypt Alexandria, Egypt
[email protected] [email protected]

Abstract— Artificial Neural Networks (ANNs) are known manipulator. Forward kinematics of a flight simulator
for their ability to map nonlinear relations between inputs and parallel system was solved using iterative ANN in [3]. A
outputs. This paper presents ANN-based kinematic modeling hybrid computational method for solving forward
of a recently developed parallel manipulator. The kinematics of a parallel machine tool using ANN was
manipulator has 3 limbs of prismatic-universal-universal (3- proposed in [4]. Zhang and Lei [5] applied different
PUU) structure. To avoid the computational complexity of intelligent approaches to investigate the forward kinematics
solving the kinematics problem in real-time application, two problem of a redundant parallel manipulator. Elsheikh [6]
artificial neural networks are trained to estimate the forward proposed forward kinematics solution of 3 different
and inverse kinematics solutions. Simulation results show that
architectures for planar parallel manipulators passing
the developed ANNs have great prediction capabilities,
providing accurate kinematic solution and can then be
through singular configurations using neural networks.
applied in the control design of the proposed manipulator. Forward kinematics solution of a 3 degrees of freedom
(DOF) parallel manipulator was estimated using an ANN
Keywords— Parallel Manipulator, Forward Kinematics, with simulated information of the end-effector’s cartesian
Inverse Kinematics, Artificial Neural Network coordinates as the training dataset [7]. A more recent study
discussed implementing a Neuro-Fuzzy Inference System
I. INTRODUCTION model to predict the end-effector’s position of a 3-link
Kinematics solution is a crucial problem for the design parallel manipulator with revolute-revolute-revolute
and control of robotic manipulators, establishing a relation structure and the results showed success in capturing the
between joint space and cartesian space. Inverse kinematics kinematics behavior of the manipulator [8, 9]. Applying
solution is the problem of solving the manipulator’s joints’ ANNs for the forward kinematics problem has proven to
variables, given the end-effector’s cartesian space position reduce the computation time, allowing a faster, and thus
and orientation. While forward kinematics solution provides more accurate real-time measurement of the manipulator’s
the end-effector’s position and orientation, knowing the position and orientation [10].
manipulator’s joint angles. Solving inverse kinematics is This paper discusses the kinematic solution of parallel
one of the most important topics in robotic control because manipulator having 3 limbs with prismatic-universal-
it provides the required joints’ variables that will move the universal structure (3-PUU). Two feed-forward supervised
manipulator’s end-effector to reach the desired position. ANNs are created using MATLAB, one for estimating the
Moreover, forward kinematics is important in practical forward kinematics and the other estimates the inverse
applications to know the exact position of the robot’s end- kinematics of the manipulator. This paper is organized as
effector. follows. Section II provides a brief description of the
Parallel manipulators which offer multiple advantages manipulator’s design and presents its kinematics. Section III
over the serial configuration, including higher structural studies the forward and inverse kinematics solution using
stiffness, payload capacity and accurate positioning suffer ANNs, showing the structure and training results of the
from complicated forward kinematics problem. It is developed ANNs. Simulation results of the developed
considered a computationally complex, time-consuming ANNs are presented in IV. Finally, section V concludes the
process due to the highly coupled nonlinear equations that paper and suggests future work.
are almost impossible to solve analytically. Recently, efforts II. KINEMATICS OF THE 3-PUU MANIPULATOR
have been made to apply Artificial Neural Networks
(ANNs) for solving forward kinematics of different parallel Schematic diagram of 3-PUU parallel manipulator
manipulators. ANNs are known for their ability to represent developed by [11] is shown in Fig. 1. It consists of a fixed
non-linear relationships between inputs and outputs, they base, a moving base to which the end-effector is attached,
have great learning abilities and are capable of learning from and 3 symmetrically arranged limbs having identical PUU
examples. This makes them great candidates for the kinematic structures connecting the fixed platform with the
mapping process between joint space and cartesian space movable one [11]. This kinematic structure provides the
[1]. Xu and Li [2] developed an ANN to estimate the manipulator with 2 rotational and 1 translational (2R1T)
forward kinematics of a 3-link prismatic-revolute-spherical DOFs.
 Assuming that Ln is the fixed length between each 2
consecutive universal joints in the nth limb as shown in Fig.
1 (where n=1,2,3), vector MB represents cartesian
n
coordinates of the second universal joint with respect to the
moving base frame M-UVW. While Pn represents
coordinates of the prismatic joint with respect to the fixed
base frame F-XYZ. The closed-form solution of the inverse
kinematics obtained in [11] is shown in (2).
dn = RM MB + FPM
F
Pn Z
n
2
RM MB + FPM
F
Pn Z (2)
n
2
RM MB + FPM
F
Pn Ln 2
(A) n

III. ARTIFICIAL NEURAL NETWORKS FOR SOLVING

KINEMATICS
In this study, two feed-forward ANNs with Levenberg-
Marquardt back propagation training algorithm [12] have
(B) been employed to solve the forward and inverse kinematics
Fig. 1. Schematic diagram of the 3-PUU manipulator. (A) Elevation of the 3-PUU manipulator. Supervised learning is carried
view of the manipulator. (B) Shows top view of the fixed base out using previously generated datasets containing inputs
and target outputs. Training datasets were generated from
simulation of the manipulator under pre-programmed
The manipulator has 3 active prismatic joints connecting
trajectories providing the end-effector’s cartesian space
the fixed platform with the first link of each limb. All three
position and orientation in addition to the analytical inverse
prismatic joints have parallel axes which are perpendicular
kinematics solution obtained in [11] providing the
to the plane of the fixed platform. For the passive universal
corresponding prismatic displacements. Since the 3-PUU
joints, the first axis of the first universal joint in each limb
manipulator has 2R1T DOFs, only 3 variables are required
is perpendicular to the plane formed by the axis of the limb’s
to describe its end-effector’s position and orientation which
prismatic joint and the line drawn from the fixed base’s
are rotational angles about x-axis (Ex), y-axis (Ey) and
center to that prismatic joint. The second axis of the first
translational displacement along the z-axis (Pz) that can be
universal joint and the first axis of the second universal joint
obtained from the position vector and the orientation matrix
are parallel to each other and perpendicular to the line
according to [13].
connecting the centers of the 2 universal joints. This unique
arrangement of the manipulator's joints enables The process of training an ANN involves tuning the
unprecedented large bending angles, ±90˚ in all directions, values of the weights and biases of the network to optimize
and almost free-interior-singularity workspace [11]. its performance. A common performance indicator is the
mean square error (MSE) that is the average squared error
Kinematic analysis of the 3-PUU manipulator is
between the network outputs and the target outputs
presented in the previous work of Khalifa et al. [11], where
calculated according to (3).
inverse kinematics solution is obtained analytically
providing a closed-form solution for the active joints’ ∑ (ANN outputs target outputs)2
variables d1, d2 and d3. While forward kinematics solution is MSE = (3)
number of samples
obtained numerically using MATLAB after studying the
The Levenberg-Marquardt backpropagation training
kinematic closure of each limb of the manipulator,
algorithm was employed as it provides fast and stable
providing 3 representations for the position vector ( FPM ) convergence of the training error [1, 12, 14]. It combines the
and orientation matrix ( FRM ) of the manipulator’s end- steepest descent method and the Gauss-Newton training
effector, one according to each limb [11]. General algorithm [15].
representation of FPM and FRM is given according to (1).
A. Structure of The Forward Kinematics ANN
Equating the 3 representations of FPM and the 3
The developed forward kinematics estimator ANN has
representations of FRM results in 24 non-linear equations 3 inputs which are d1, d2 and d3. And 3 outputs
with 12 unknowns that are the angles of the universal joints’ corresponding to Ex, Ey and Pz. Its structure is illustrated in
angles. These nonlinear equations are solved using Fig. 2. This structure has been settled on after several trial
MATLAB program. After solving for the universal joints’ and error procedures. It has 2 hidden layers with 50 and 20
angles, these values are substituted in the obtained neurons respectively and hyperbolic tangent sigmoid
representations of FPM and FRM . function as the activation function, and the output layer
having 3 neurons and a linear activation function. To avoid
Px Ux Vx Wx overfitting, data were randomly divided for training,
F F
PM = Py , R M = Uy Vy Wy (1) validation and test, with 70% provided for training, 15% for
Pz Uz Vz Wz validation and 15% for test.
 Fig. 2. Structure of the Forward kinematics estimator neural network

The created ANN has best validation performance of

1.0508 x 10-5 MSE at epoch 812 as illustrated in Fig. 3. The
error histogram depicted in Fig. 4 shows a good normal
distribution where the network has a maximum frequency
of errors around zero. A perfect fit between the network’s
output and target values indicated by an overall regression
R value of 1 is shown in Fig. 5.

Fig. 5. Regression plot of forward kinematics neural network

Fig. 6. Structure of the Inverse kinematics estimator neural network

B. Structure of The Inverse Kinematics ANN

Fig. 3. Validation performance of forward kinematics neural network
Similarly, the inverse kinematics ANN is trained to have
3 inputs: Pz, Ex and Ey, and 3 outputs: d1, d2 and d3.
Trial and error design procedures resulted in creating a
network having 1 hidden layer with 70 neurons and
hyperbolic tangent sigmoid function as the activation
function, illustrated in Fig. 6. This network structure along
with the adequate number of dataset samples resulted in best
validation performance of 7.0226 x 10-5 MSE at epoch 2119
as shown in Fig. 7. The error histogram shown in Fig. 8
shows a good normal distribution where the network has a
maximum frequency of errors around zero. The regression
plot in Fig. 9 shows a perfect fit between the network’s
output and target values indicated by an overall regression
R value of 1.

Fig. 4. Error histogram of forward kinematics neural network

Fig. 7. Validation performance of inverse kinematics neural network

 Output of the inverse kinematics neural network is
compared to the closed-form solution of inverse kinematics
obtained in [11] after applying new dataset as their input.
Fig. 13 depicts the comparison. The created neural network
has great prediction capability, with maximum deviation
from the analytical solution with 0.4128, 0.1314 and 0.3592
mm in d1, d2 and d3 respectively.

Fig. 8. Error histogram of inverse kinematics neural network

(A)

Fig. 9. Regression plot of inverse kinematics neural network

IV. SIMULATION RESULTS

The ANNs were trained using large number of
previously generated data samples. After training,
comparison between the output targets and the ANN’s
actual outputs was held. Fig. 10 and Fig. 11 depict this
comparison for the forward kinematics and inverse
kinematics ANNs displaying random values of the datasets.
It is seen that the created ANNs provide accurate kinematics
solution with almost zero error.
To study the prediction capability of the created ANNs,
they are tested with new datasets. For the forward
kinematics ANN, the manipulator is simulated, following a
certain path that the ANN has not been trained with, while
observing the simulated moving base’s position and
orientation and comparing it to the ANN’s output. Fig. 12
compares the position and orientation of the manipulator
obtained from simulation and from the developed ANN. It
is clear that the created ANN has reliable prediction
(B)
capability, with maximum errors of 0.1072 mm, 0.01485 rad Fig. 10. Forward kinematics neural network output versus target values.
and 0.007936 rad in Pz, Ex and Ey respectively. (A) First random dataset (B) Second random dataset
 Fig. 12. Testing the forward kinematics neural network with new dataset

(A)

Fig. 13. Testing the inverse kinematics neural network with new dataset

V. CONCLUSION AND FUTURE WORK

(B)
Fig. 11. Comparing target outputs and actual outputs of the inverse This paper presented an artificial intelligence-based
kinematics ANN. (A) First random dataset (B) Second random dataset method for solving kinematics of a 3-DOF parallel
manipulator that was recently developed by Khalifa et al.
[11]. Two supervised feed-forward ANNs were trained to
provide real-time, fast and accurate forward and inverse
kinematic solutions of the manipulator. Training results
showed best validation performance of 1.0508 x 10-5 mean
squared error for the forward kinematics ANN, and 7.0226
x 10-5 mean squared error for the inverse kinematics ANN.
Moreover, testing the ANNs with new datasets proved that
they have great prediction capabilities and can be then
employed in the control system of the manipulator.
Future work will include testing other neural network
configurations to solve the manipulator’s kinematics. In
addition, dynamic modeling of the laparoscopic
manipulator using neural networks such as NARX ANN
will be investigated.
REFERENCES

[1] A.-V. Duka, "Neural network based inverse kinematics solution for
trajectory tracking of a robotic arm," Procedia Technology, vol. 12, pp.
20-27, 2014.
[2] Q. Xu and Y. Li, "A 3-prs parallel manipulator control based on neural
network," in International Symposium on Neural Networks, 2007, pp. 757-
766.
[3] P. J. Parikh and S. S. Lam, "Solving the forward kinematics problem
in parallel manipulators using an iterative artificial neural network
strategy," The International Journal of Advanced Manufacturing
Technology, vol. 40, pp. 595-606, 2009.
[4] L. Guan, J. Wang, and L. Wang, "Hybrid computational intelligence
based kinematic analysis for parallel machine tool," in SMC'03
Conference Proceedings. 2003 IEEE International Conference on
Systems, Man and Cybernetics. Conference Theme-System Security and
Assurance (Cat. No. 03CH37483), 2003, pp. 2763-2768.
[5] D. Zhang and J. Lei, "Kinematic analysis of a novel 3-DOF actuation
redundant parallel manipulator using artificial intelligence approach,"
Robotics and Computer-Integrated Manufacturing, vol. 27, pp. 157-163,
2011.
[6] A. Elsheikh, E. Showaib, and A. Asar, "Artificial neural network based
forward kinematics solution for planar parallel manipulators passing
through singular configuration," Adv Robot Autom, vol. 2, p. 2, 2013.
[7] A. Youssef, A. M. Bayoumy, M. Rostom, and F. A. Tolbah,
"Modelling and simulation of 3DOF parallel manipulator using artificial
neural network," in IOP Conference Series: Materials Science and
Engineering, 2019, p. 012080.
[8] A. S. Sayed, A. T. Azar, Z. F. Ibrahim, H. A. Ibrahim, N. A. Mohamed,
and H. H. Ammar, "Deep Learning Based Kinematic Modeling of 3-RRR
Parallel Manipulator," in AICV, 2020, pp. 308-321.
[9] A. S. Sayed, N. A. Mohamed, A. A. Salem, and H. H. Ammar,
"Modeling of Nonlinear 3-RRR Planar Parallel Manipulator: Kinematics
and Dynamics Experimental Analysis," Int. J. Mech. Mechatronics Eng.,
vol. 20, pp. 175-185, 2020.
[10] J. Ghasemi, R. Moradinezhad, and M. A. Hosseini, "Kinematic
synthesis of parallel manipulator via neural network approach," arXiv
preprint arXiv:1904.04668, 2019.
[11] A. Khalifa, M. Fanni, A. M. Mohamed, and T. Miyashita,
"Development of a new 3-DOF parallel manipulator for minimally
invasive surgery," The International Journal of Medical Robotics and
Computer Assisted Surgery, vol. 14, p. e1901, 2018.
[12] J. J. Moré, "The Levenberg-Marquardt algorithm: implementation
and theory," in Numerical analysis, ed: Springer, 1978, pp. 105-116.
[13] G. G. Slabaugh, "Computing Euler angles from a rotation matrix,"
City University London, Technical Report. Retrieved on August 1999.
[14] M. T. Hagan and M. B. Menhaj, "Training feedforward networks with
the Marquardt algorithm," IEEE transactions on Neural Networks, vol. 5,
pp. 989-993, 1994.
[15] H. Yu and B. M. Wilamowski, "Levenberg–marquardt training," in
Intelligent systems, ed: CRC Press, 2018, pp. 12-1-12-16.

View publication stats