Inverse Kinematics Problem (IKP) of 6-DOF Manipulator By Locally Recurrent Neural Networks (LRNNs)

Y. I. Al-Mashhadany, MIEEE Electrical Eng.Dept./Al-Anbar University Email: [email protected].

Abstract The paper presents a cognitive architecture for solution of inverse kinematics problem (IKP) of 6DOF elbow manipulator with spherical wrist by Locally Recurrent Neural Networks (LRNNs) and simulated the solution by using MATLAB/Simulink. This design is aimed to allow the manipulator system to perform complex movement operations by solving the Inverse Kinematic Problem (IKP) with LRNNs by using the position and orientation of end-effector which represent by wrist with 3-DOF. The Levenberg-Marquardt back propagation (LMBP) is used in the learning of LRNNs which offered the high computation and accuracy for solving IKP for manipulator. This model permits direct forward dynamics simulation, which accurately predicts wrist position, also present a solution to the inverse problem of determining set of joints angle to achieve a given command for posture of manipulator. The simulation of design achieved by connect the program with Simulink\MATLAB Ver. 2009b to calculate the forward and inverse kinematic and implement the movements manipulator. Satisfactory results are obtained, that explains the ability of implement the posture of 6-DOF manipulator by calculate the kinematic with LRNNs and implement high complex movements. Index Terms Locally Recurrent Neural Networks, Inverse Kinematics Problem

I. INTRODUCTION robot manipulator is composed of a serial chain of rigid links connected to each other by revolute or prismatic joints. A revolute joint rotates about a motion axis and a prismatic joint slide along a motion axis. Each robot joint location is usually defined relative to neighboring joint. The relation between successive joints is described by 4x4 homogeneous transformation matrices that have orientation and position data of robots. The number of those transformation matrices determines the degrees of freedom of robots. The product of these

transformation matrices produces final orientation and position data of a n degrees of freedom robot manipulator. Robot control actions are executed in the joint coordinates while robot motions are specified in the Cartesian coordinates. Conversion of the position and orientation of a robot manipulator end-effector from Cartesian space to joint space, is called as inverse kinematics problem, which is of fundamental importance in calculating desired joint angles for robot manipulator design and control [1]. For a manipulator with n degree of freedom, at any instant of time joint variables is denoted by (qi = q (t), i = 1;2;3;....n) and position variables (xj = x(t), j = 1;2;3;.m). The relations between the endeffector position x(t) and joint angle q (t) can be represented by forward kinematic equation, x(t) = f (q (t)) where f is a nonlinear, continuous and differentiable function. On the other hand, with the given desired end effector position, the problem of finding the values of the joint variables is inverse kinematics, which can be solved by, q (t) = f(x(t)) Solution of (q(t)) is not unique due to nonlinear, uncertain and time varying nature of the governing equations. Fig-1- shows the schematic representation of forward and inverse kinematics[2]. The different techniques used for solving inverse kinematics can be reviewed with some articles where, Wu et.al. [3], a new analytic inverse kinematics (IK) solver is proposed which is suitable for multiple constrained 12-DOF human limbs. By decomposing human skeleton into five parts one head chain , two arm chains and two leg chains, a multiconstrained human skeleton can be solved analytically. Al Faiz et.al. [4], presented a computation of the inverse kinematic model of the human arm for robot based rehabilitation that uses measurements of the hand position and orientation and radial acceleration of the upper arm. Analytical analysis and empirical validation of the method are presented. The

978-1-4244-5326-9/10/$26.00 2010 IEEE

algorithm enables estimation of human arm angles, which can be used in trajectory planning for rehabilitation robots. Drzevitzky [5], introduced Inverse Kinematics problems for anthropomorphic limbs and have shown how to solve those analytically in order to obtain symbolic solutions. The symbolic solutions can be modied and re-computed to match ,for example, other input values that serve as constraints when solving the according Inverse Kinematics problem. II. CASE STUDY An Elbow Manipulator with Spherical Wrist has been taken as the example for applied the solution in analytic method and then by used N.N, where it is shown in Fig-1- [6].

c1 s A1 = 1 0 0 c 2 s A2 = 2 0 0 1 0 A3 = 0 0

0 s1 0 c1 1 0 0 0 0 s2 0 c2 1 0 0 0 0 1 0 0 0 0 0 0 1 d3 0 1

0 0 0 1 0 0 d2 1

(1)

(2)

(3)

3 z2

5 4 6

c 4 s A4 = 4 0 0 c 5 s A5 = 5 0 0 c 6 s A6 = 6 0 0

0 s4 0 c4 1 0 0 0 0 s5 0 c5 1 0 0 0 s6 c6 0 0

0 0 0 1 0 0 0 1

(4)

(5)

Fig.1. Elbow Manipulator with a spherical wrist.

A. ANALYTIC SOLUTION This manipulator has an offset in the shoulder joint that slightly complicates both the forward and inverse kinematics problems. We first establish the joint coordinate frames has been established using the DH convention. The DH parameters are shown in the Table 1. It is straightforward to compute the matrices Ai as shown below:
TABLE 1. DH PARAMETERS FOR STANFORD MANIPULATOR.( * JOINT VARIABLE ).

0 0 0 0 1 d6 0 1

(6)

The forward kinematic represents by given as:

0 A6 and it is

A6 .

Link 1 2 3 4 5 6

di d1 0 0 0 0 d6

ai 0 a2 a3 -/2 /2 0

i /2 0 0 0 0 0

i * * * * * *

r11 r12 r13 R t13 r21 r22 r23 33 0 = A6 = r31 r32 r33 0 0 0 1 0 0 0

tx ty (7) tz 1

The end-effector of the manipulator at the point with coordinates given by (o) and the orientation of the end-effector given by R=(rij), it is necessary and sufficient that the wrist center (oc) have coordinates given by :

0 o = o d 6 R 0 1
o c

(8)

The orientation of the frame o6x6y6z6 with respect to the base be given by R. If the components of the end-effector position (o) are denoted ox,oy,oz and the components of the wrist center o co are denoted xc,yc,zc then equation(8) gives the relationship:

xc xc d 6 r13 y = y d r 6 23 c c zc zc d 6 r33

(9)

The inverse kinematic is achieved by closed solution of above eqns, and the general solution of angles of rotation can be summarized as follows:

3 = a tan(D, 1 D2 ) 2 2 2 2 2 2 x + yc d + ( zc d1 ) a2 a3 where: D = c 2a2a3 2 = a tan 2( xc2 + yc2 d 2 , zc d1 ) (10) 4 = a tan 2(c1c23r13 + s1c23r23 + s23r33, c1s23r13 + s1s23r23 + c23r33 ) 2 5 = a tan 2(s1r13 c1r23, 1 (s1r13 c1r23 ) ) 6 = a tan 2(s1r11 + c1r21, s1r12 + c1r22 ) where: ci = cos( i ), and si = sin(i )
If we substitute the values of (d1=15, d6=10, a2=20, a3=25 (all lengths in cm)) with position and orientation of end-effector at (25,30,15,/2,-/2,/2) respectively, the following results can be obtained: respectively, the following results can be obtained:
xc = 10cm, yc = 30cm, zc = 15cm, D = 25cm, 1 = 0.4636 rad , 2 = 0.8961 rad , 3 a = 3 b = 1.5708 rad , 4 = 2.2455 rad , (11) 5 a = 0.4636 rad , 3 b = 2.6779 rad , 6 = 6.1232 e 17 rad ,

1 = a tan 2( xc , yc )

B. ANN SOLUTION This section introduces the basics of ANN architecture and its learning rule. Inspired by the idea of basing the feedforward and backpropagation network structure. Fig -3- shows this structure , the Learning rule which is used in this paper is fast momentum back-probagation with delta rule structure of network with dimension ( 3-19-6). The inputs are the position of end-effector in (x,y,z) ,the network is single layer with dimension (19) neurons ( this dimension limited by trial and error ). The output dimensions are the angles of rotation and translation displacement in joints. The N.N. is used for identify the system based on network shown in Fig 2 It is represent the LRNNs[7]: It consist of i/p Layer (I), hidden Layer (H) and output Layer (O). The weight matrix consists of WImxi and WOnxm, the characteristic of the network is that the units in hidden layer are connected with themselves and each other. However, the information in such units at one time can only be transferred until being replaced by new data after finite steps, i.e. the units in hidden layer only memorize the information for finite time which is the key difference from the fully recurrent BP. Hence at the time (t) the input to the (ith) hidden units is:
AHi (t ) = WT ik I k +
k =1 I

WH
k =1

ik

i = 1, 2 ..... m

f ( AH k (t 1)); (12)

The activation function (tangent f(x)) which is used in hidden layer. The output of the network is a weighted sum of the hidden unit o/ps:

Oi (t ) = WOik f ( AH k (t ) );
k =1

i = 1, 2, ....n

(13)

The net is trained by minimization of the total error, which is evaluated as that:
E (t ) =

pp

1 p =1 2
pp

k =1

e kp ( t ) 2 ,
kp

&

1 E (t ) = p =1 2

(T
k =1

( t ) O kp ( t ) 2

(14)

Where; (pp) is the sample length, Tkp(t) are the target values that the output of the (Kth) unit in output layer while inputting the (pth) sample match at time (t). This can be fulfilled by a gradient descent procedure adjusting (w) a long the negative of w E(t). In this paper the dimension of LRNNs used (6-196) and the adaptation of weight is achieved by Levenberg-Marquardt modification algorithm. All the details of learning algorithm can be seen by ref.[8,9].

Fig. 3. Analytical Solution of IKP for 6-DOF Elbow Manipulator.

Fig. 2. Scheme of LRNN with one hidden layer.

III. SIMULATION RESULTS The simulation of N.N and the solving of equations of analytic solution is achieved by Matlab Var.R2009b. Fig-3- presents the solution of IKP by analytical solution with Matlab/Simulink and Fig-4represents the solution of IKP by using LRNNs. This simulation is solved by using the connection between the MATLAB/Editior program and Simulink to run every input of anllytical solution and NN solution and then comparison between them. Figs-5,6represent the N.N. identification for inverse kinematic calculation with maximum normalized values for an angle of joints. The solution of IKP is achieved by using the LRNNs as forward direction after the neuro-identifer data base to be making used in calculation of N.N (with dimension (6-19-6 )) inverse kinematic by LMBP learning algorithm.

Fig. 4. NN Solution of IKP for 6-DOF Elbow Manipulator.

Fig. 5. NN identification with maximum normalized values for angle joints.

[4] M. Z. Al Faiz, Y. I. Al Mashhadany, Analytical Solution for Anthropomorphic Limbs Model, (IK of Human Arm) , Will published in ISEIA 4-6 October, 2009. [5]. S. Drzevitzky, Symbolic Solutions for Inverse Kinematics Problems for Anthropomorphic Limbs, Diplomarbeit Thesis, Germany, March, 2008. [6]. M W. Spong, M. Vidyasagar. Robot Modeling and Control. First Edition, 2003 [7]. L Wu, P Baldi. Learning to play Go using recursive neural networks science direct neural network 21 February 2008

[8]. M.T. Hagan, M.B. Menhaj, Training Feedforward Network with the Marquardt Algorithm, IEEE transaction on neural network VOL. 5, NO. 6, November 1994. [9]. Z. Liu and I. Elhanany, A Fast and Scalable Recurrent Neural Network Based on Stochastic Meta Descent. IEEE transaction on neural network, vol. 19, no. 9, september 2008

Fig 6.Error values for LRNNs identification. Yousif Ismial Mohammed AL Mashhadany (MIEEE) was bon in Baghdad 1973. He received the B.Sc. degree in Electrical and Electronic Engineering Department from the MEC, Baghdad (1995) Iraq. M.Sc degree in Control Engineering from MEC, Baghdad (1999). Ph.D degree in Control Engineering from the University of Technology, Iraq in (2009). Since 2004, he has been working at the University of Anbar Iraq, as an Asst. Lecturer in the Electrical Engineering Department. His research interests include biomedical, robotic and control system.

IV. CONCULITION From the obtained results, it can be concluded that, the ability of FRBP-NN to identify and solve the IKP with high accuracy and it can use NN to identify any envelop for movement of manipulator with short time of computation when it compared with analytical solution by closed form using the Jacobean matrix. The sensitivity of N.N in calculation depends on the band of identified sampling. This point can be overcome by training the network for every point with many trails to get optimal results. Any additional movement constraints to the IKP may be resolved easily considering the proposed NN approach compared to that one of analytical solution.. REFERENCES
[1]. P. Baerlocher, Iverase Kinematic Techniques of the Interactive Posture Control of Articulated Figures. Thesis 2001. [2]. S Alavandar, M.J. Nigam. Inverse Kinematics Solution of 3DOF Planar Robot using ANFIS, Int. J. of Computers, Communications & Control, ISSN 1841-9836, E-ISSN 18419844. Vol. III (2008), Suppl. issue: Proceedings of ICCCC 2008, pp. 150-155 [3] X. Wu, L. Ma, Z. Chen, Y. Gao, A 12-DOF Analytic Inverse Kinematics Solver for Human Motion Control , Journal of Information & Computational Science, PP 137141, 30-August 2004.