Automatic generation of fuzzy rules for reactive robot controllers

Basic tasks for navigation of autonomous vehicles can be performed as reactive behaviors, that directly map sensory data into control commands with no need of internal representations. Fuzzy systems can efficiently realize such a direct mapping by means of comprehensible linguistic rules. Automatic learning of rules from representative data simplifies the programming of the control system especially when dealing with dynamically changing environments. We present two methods for the automatic extraction of fuzzy rules from numerical data acquired by recording the choices of a human operator driving the vehicle in representative training situations. They are applied to build a reactive wall-follower, the first component of a more articulated control system for indoor navigation of a TRC Labmate mobile robot. The produced wall-followers are compared in terms of complexity and quality of behavior in actual navigation runs of the robot along arbitrarily shaped walls.