Neural network architectures for the forward kinematics problem in robotics

Abstract

Various neural models are considered for solving the robot forward kinematics problem. It is demonstrated that a three-layer backpropagation network is capable of learning the forward kinematics of a rigid-link, open-chain manipulator without knowledge of the manipulator's kinematic structure. Simulation results show that, by properly training such a network, it is possible to model the forward kinematics with an acceptable degree of accuracy. However, it is also shown that, if information about the kinematic structure of a manipulator is available, a functional link network gives, by far, the most accurate results