ADAPTIVE NONLEARNING AND LEARNING CONNECTIONIST CONTROL OF MANIPULATION ROBOTS WITH FLEXIBLE LINKS

摘要

This paper considers the problem of adaptive control in the case of manipulation robots with flexible links using two different techniques: the first considers conventional nonlearning adaptive indirect control based on a model of rigid robot dynamics, and the second is based on hybrid learning method using integrated connectionist structures for fast and robust online learning of robot dynamic uncertainties. The conventional adaptive nonlearning approach uses the basic principles of indirect adaptive control (estimation of unknown robot dynamic parameters and self-tuning strategy). Using the theory of indirect adaptive control systems, the authors propose new robust algorithms for adaptive control with varying ability to adapt in feedforward or feedback loop. In order to compensate unmodelled dynamics of elastic modes, adaptive nonlearning control is upgraded with hybrid learning approach that uses model-based control part and neural network structures. The multilayer perceptrons or recurrent neural networks as a part of hybrid learning control algorithms use fast learning rules and available sensor information in order to progressively improve robotic performance for the smallest possible number of learning epochs through the process of synchronous training. Some simulation results of trajectory tracking with single-link flexible robot and unknown dynamic parameters of payload are presented in order to verify the effectiveness of the proposed approach.