Learning from NN output feedback control of robot manipulators

Abstract

In this paper, based on recently developed deterministic learning (DL) theory, we investigate the learning issue in neural network (NN) output feedback control of robot manipulators with unknown system dynamics and disturbance. Our objective is to learn the unknown closed-loop robot system dynamics while tracking to a periodic or periodic-like reference orbit with only joint angle measurements. Firstly, a high-gain observer (HGO) is used to estimate the joint velocities. An adaptive NN output feedback controller is then designed to guarantee the stability of the closed-loop robot system and the tracking performance when tracking a periodic or periodic-like reference orbit. Based on DL theory, when a partial persistence of excitation (PE) condition of the regression subvector is satisfied, part of the neural weights of the employed radial basis function (RBF) NN will converge to their optimal values. The unknown dynamics of robot manipulators can be learned by NN in a local region along the estimated state trajectory and the learned knowledge is stored in constant RBF networks. Secondly, the peaking phenomenon generated by the use of HGO and its adverse effect on learning are analyzed. If the gain of HGO is not chosen too high, the peaking phenomenon will be weakened and the accuracy of the estimated system states can still be guaranteed for learning from robot manipulators control. Thirdly, when repeating same or similar control tasks, the learned knowledge can be recalled and reused to achieve the guaranteed stability and better control performance with little effort. Finally, simulation studies are included to demonstrate the effectiveness of the proposed method.