Redundant Robot Kinematics Control with HCMAC Neural Network Manipulability Enhancement

Abstract

The kinematics problems of redundant robots have been investigated for many years. A plenty of different robot redundancy usages were successfully implemented. In practice, e.g. redundant robot manipulability improvement, and robot obstacle avoidance are commonly used. Conventional methods use the manipulability gradient computation to improve the end-effector manipulability. However, the computational effort of this approach brings about many difficulties in real-time control when used with plenty of constraints. Recently, neural networks have found wide application in robotics, because of their feature to learn any complicated system model. This paper deals with new numerically efficient procedures and an application of HCMAC (Hierarchical Cerebellar Model Arithmetic Controller) neural network for manipulability gradient computation and consecutive robot control improvement. The conventional CMAC can be viewed as a basis function network (BFN) with supervised learning well-performing in the terms of its fast learning speed and local generalization capability for approximating nonlinear functions. Nevertheless, due to the redundant robot high-dimensional function approximation the conventional CMAC is enormously memory-demanding. HCMAC indeed has low memory requirements and very good learning ability. Furthermore, it shows a better performance compared with the conventional approach and conventional CMAC.