Real-Time Implementation of Neural Network Learning Control of a Flexible Space Manipulator

Abstract

This paper discusses a neural network approach to on-line control learning and real-time implementation for a flexible space robot manipulator. We at first overview the system development of the Self-Mobile Space Manipulator and discuss the motivations of our research. Then, we propose a neural network to learn control by updating feedforward dynamics based on feedback control input. We address in great detail the implementation issues associated with on-line training strategies and present a simple stochastic training scheme. A new recurrent neural network architecture is proposed, and the performance is greatly improved in comparison to the standard neural network. By using the proposed learning scheme, manipulator trajectory error is reduced by 85%. At last, we discussed the implementation of the proposed scheme in teleoperated control. The approach possesses a high degree of generality and adaptability in various applications and will be a valuable method in learning control for robots working in unconstructed environments.