Reactive navigation by fuzzy control

Abstract

Reactive navigation can improve the performance of robotic systems by solving simple subtasks without requiring the update of internal representations or plans: it involves only the mapping of perceptual situations onto control commands. Fuzzy systems and neural networks are both efficient methods for learning such a mapping from training data, avoiding its annoying and error-prone programming. Moreover fuzzy systems allow the use of a-priori knowledge about the task and the verification of the strategy acquired during the learning phase. This paper presents a control system for wall-following which is intended to be the first brick of a more articulated control system including all the low-level behaviors needed for the safe navigation of an autonomous mobile vehicle. The wall-follower works on data supplied by an ultrasonic sensor ring. The mapping between the input space and the control command space is generated automatically from training data acquired during operator-driven runs of the vehicle. The first experimental results, obtained by controlling a TRC Labmate with the obtained fuzzy system inside an indoor environment, point out that even very simple training sessions allow the derivation of fuzzy rules which have proved effective in solving the task at hand.