Constructive RBF network based iterative learning controller for manipulators

Abstract

Aiming at the slow convergence speed of iterative learning controller for trajectory tracking of manipulator, a new iterative learning controller based on a constructive RBF neural network is proposed by well considered the past experience of tracking various trajectories to select the initial control input of an iterative learning controller properly. A new desired trajectory can be decomposed into many query points at first, and then RBF network is applied to construct inverse dynamics of manipulator by fitting the nearest k data points near every query point and subsequently is used to predict the initial control input. Moreover, the structure of RBF network can be constructed dynamically according to the change of the manipulator's dynamics during the control process, which ensures network size is economical. Therefore, the proposed iterative learning controller has advantages of quick learning, high accuracy and compact structure. At last, the control method is verified by computer simulation of trajectory tracking of a planar two-link manipulator.