Practical connection between potential fields and neural networks

Abstract

In this paper we examine the underlying similarities and differences between two major computational formalisms in developing intelligent robots; namely, artificial potential fields, which are often implemented for real-time robot planning and control, and artificial neural networks, which are usually considered as one of the biologically-inspired powerful learning techniques. Such comparisons will offer us new insights into how the two can complement each other in learning and control during robot-environment interaction. (This paper is based on an article by Liu and Khatib for the forthcoming The Handbook of Brain Theory and Neural Networks, Michael Arbib, Editor, MIT Press, 2000.) 1 INTRODUCTION The problem of robot motion planning was traditionally treated as an optimization problem, in which the configuration of a robot is represented in a parameter space, and a solution to this problem is computed by searching the parameter space in an attempt to satisfy a predefined cost funct...