Beruflich Dokumente
Kultur Dokumente
Proc. of the 5 International Conference on Advances in Mechanical Engineering (ICAME-2011), June 06-08, 2011
S.V. National Institute of Technology, Surat – 395 007, Gujarat, India
1. Introduction
A Robot manipulator is made up of rigid links serially connected to each other by
prismatic or revolute joints. The position and orientation data of robots is stored in a 4x4
homogenous transformation matrix that is known as the forward kinematics model. The
control of a robot is executed in the joint space while robot motions are specified in the
Cartesian coordinates. The Inverse kinematics (IK) problem involves the conversion of the
position and orientation of end-effectors of a robot manipulator from Cartesian space to joint
space. The conventional solutions to the Inverse kinematics problems involve tedious
calculations and are time intensive. Moreover closed form geometric solutions are not
available for every kind of robot owing to matrix singularities, existence of multiple solutions
and difficulty of solving trigonometric equations analytically.
To overcome the above problems, neural network solution to inverse kinematics
problems has been proposed by many researchers. The ability of neural networks to learn by
example makes them an efficient tool to model complex problems. Allon Guez and Ziauddin
Ahmad [1] employed a neural network model in the solution of the inverse kinematics problem
in robotics and found that the neural network can be trained to generate a fairly accurate
solution which when augmented with local differential inverse kinematic methods will result in
minimal burden on processing load of each control cycle and thus enable real time robot
control. Joseph A. Driscoll [2] chose Radial basis function (RBF) neural networks over other
neural network architectures because they usually exhibit good performance in function
approximation. Pei-Yan Zhang, Tian-Sheng Lu and Li-Bosong [3] used the Levenberg–
Marquardt (LM) algorithm for training networks used for the solution of IK problem of the 6-
DOF MOTOMAN manipulator. Pablo J. Alsina and Narpat S. Gehlot [4] studied the modular
approach for solving the Inverse kinematics in which neural modules are assigned to each
link in order to realize its own inverse kinematics and then the inverse neural modules are
concatenated in a global scheme for the updating of the inverse kinematics of a 3-DOF and
4-DOF SCARA manipulator. Z. Bingul, H.M. Ertunc and C. Oysu Bingul [5] used Back
propagation algorithm to train the networks for solution of the IK problem of a 6-DOF robot
manipulator with offset wrist, whose geometric features do not allow solving inverse
kinematics problem analytically.
In the present work, a MLNN model is made for a 6-DOF Stanford arm, with a fixed
wrist and LM algorithm has been used to train the networks. An approach to choose the
training data, has been suggested which minimizes the error in the Inverse Kinematics
solution.
2. Methodology
2.1 Generation of training data
The forward kinematic model of the Stanford Arm is formulated according to the D-H
convention. Sample of the various joint variables (qi) are generated using incremented values
936
th
Proc. of the 5 International Conference on Advances in Mechanical Engineering (ICAME-2011), June 06-08, 2011
S.V. National Institute of Technology, Surat – 395 007, Gujarat, India
with in the presumed domain of the manipulator. Substituting the different qi values in the
transformation matrix the end effecter’s position and orientation is computed taking help of the
Peter Corke’s Robotics toolbox in MATLAB [6]. At first the qis are increased with a constant
gradient and imported to the robotics toolbox to generate 1000 datasets of position and
orientation of the end effector. This dataset corresponds to the points of a localized trajectory
of the manipulator. Subsequently the 3D workspace was plotted [Figure 1-3] and another data
set was generated by randomly choosing points from it. It is a more representative dataset for
training the network.
937
th
Proc. of the 5 International Conference on Advances in Mechanical Engineering (ICAME-2011), June 06-08, 2011
S.V. National Institute of Technology, Surat – 395 007, Gujarat, India
3. Discussion of Results
Table.1 shows the comparison of various network-training algorithms. The performance
measure of the algorithms is the mean square error. The Levenberg-Marquardt (LM)
algorithm gave shows the least mean square error of 9.45e-18. Based on the performance
measure, LM algorithm is used for training the network used for the IK solution. Further,
increasing the number of hidden layer neurons from 10 to 20 increased the performance from
8.98e-16 to 9.45e-18.
Table 1 Comparison of different Learning Algorithms
Learning Algorithms in MATLAB Epoch Time Performance Gradient
Levenberg–Marquardt (auto) 1000 1:48 8.98E-16 1.15E-09
Bayesian Regularization (SSE) 100 0:15 5.40E-11 5.64E-05
Conjugate Gradient with Beale-Powell 139 0:02 1.79E-05 0.000844
Restarts
Conjugate Gradient Back propagation 242 0:03 9.62E-06 0.000958
with Fletcher-Reeves Restarts
Conjugate Gradient Back propagation 198 0:03 1.08E-05 0.000271
with Polak-Ribiere Restarts
Gradient Descent Back propagation 1000 0:10 3.36E-02 0.061
with momentum
Gradient Descent Back propagation 107 0:01 3.28E-03 0.0525
with Adaptive Learning Rate
The error of the generated solution is analyzed using two different sets of training data.
Figure.2 shows an error plot using the localized training data, while Figure.3 shows an error
plot using generalized training data. It is observed that when the link parameters are
938
th
Proc. of the 5 International Conference on Advances in Mechanical Engineering (ICAME-2011), June 06-08, 2011
S.V. National Institute of Technology, Surat – 395 007, Gujarat, India
increased linearly they generate only a specific trajectory and not the entire workspace of the
manipulator. Hence, when the network is trained with such a dataset the results are accurate
only when it is simulated for a point near to the trajectory, otherwise for distant points the error
is considerable. When the entire generalized workspace is plotted and the training dataset is
selected from it, the overall error of simulation for the trained network is minimized.
4. Conclusions
References
[1.]Guez. A, Ahmad, Z. “Solution to the Inverse Kinematics problem in Robotics by Neural
Networks”, Proceedings of the IEEE International conference on Neural Networks, San
Diego-1988, Vol 2, 617-624
[2.]Driscoll. J. A. “Comparison of Neural Network Architectures for the Modeling of Robot
Inverse Kinematics”, Proceedings of the IEEE, Southeaston-2000, 44-51
[3.]Zhang.P, Lu.T, Bosong.L. “RBF networks-based inverse kinematics of 6R manipulator”,
International Journal of Advanced Manufacturing Technology, 2005, Vol-26,144-147
[4.]Pablo J. Alsina Narpat S. Gehlot , “Robot Inverse Kinematics: A Modular Neural Network
Approach”, Proceedings of the 38th Midwest Symposium on Circuits and Systems,Vol-2,
631-634, August 13-16,1995,Rio-de-Janerio,Brazil
[5.]Z. Bingul, H.M. Ertunc and C. Oysu ,“Comparison of Inverse Kinematics Solutions Using
Neural Network for 6R Robot Manipulator with Offset”, Proceedings of the IEEE Computer
Intelligence Methods and Applications, ICSC Congress, 2005
[6.]P.I. Corke, "A Robotics Toolbox for MATLAB", IEEE Robotics and Automation Magazine,
Volume 3(1), March 1996, 24-32.
939