Kinematics control of redundant manipulators using a CMAC neural network combined with a genetic algorithm

SUMMARY A method is proposed to solve the inverse kinematics and control problems of robot control systems using a cerebellar model articulation controller neural network combined with a genetic algorithm. Computer simulations and experiments with a 7-DOF redundant modular manipulator have demonstrated the effectiveness of the proposed method. 1. INTRODUCTION The design of a redundant manipulator involves an inverse kinematics problem in determining the joint variables corresponding to any desired end-effector position and orientation. In order to control the manipulator, the joint trajectories must be found which yield the desired end-effector trajectory. Since a redundant manipulator has more degrees of freedom (DOF) than necessary, the solution is not unique. The Jacobian pseudoinverse algorithm has been widely used for solving such inverse kinematics problems, because it is able to satisfy additional constraints through mapping the velocities corresponding to the additional constraints onto the null space of the Jacobian while tracking the desired workspace trajectory. 1 However, the inversion of the Jacobian matrix causes difficulties at or near singularities. Neural networks (NNs) 2 have been widely applied in robotic control in recent years. They are normally composed of many neurons, which make them robust and fault-tolerant. Since NNs are able to solve highly nonlinear problems, they are considered very promising for application to robotic control problems. Among the many kinds of NNs, CMAC (Cerebellar Model Articulation Controller) NNs will be discussed in this paper.