Robot exploration using the expectation-maximisation algorithm

Abstract

Adaptability is an important attribute for any robotic system operating in an unstructured environment. The paper describes the first steps towards an adaptable robotic platform, capable of learning behaviours. This involves learning a new low-level behaviour 'on the fly' and integrating it into the existing set of behaviours. The first task selected for the robot to learn is obstacle avoidance. The paper will introduce an innovative and structured method of building knowledge acquired during robotic explorations. The aim is to make direct use of sensory information to construct abstractions of 'perceptions' and build strategies based on constructed knowledge to solve simple navigation tasks.