Towards autonomous robot control via self-adapting recurrent networks

Abstract

This paper introduces a connectionist architecture for autonomous robot control in which second- order recurrent connections are used to provide a flexible, context-dependent mapping from percepts to actions in order to allow the network to adapt its behaviour continually to its current context and internal state. It is argued that this mechanism, to a higher degree than modular approaches, allows the robot to acquire and adapt complex behaviour autonomously.