Skills' learning in an autonomous mobile robot using continuous reinforcement

This paper describes a reinforcement learning algorithm which allows a mobile robot, equipped with a vision system formed by a single CCD camera mounted on a pan-tilt platform, to learn simple skills such as watching skill and orientation skill. Once these skills are learned, they are combined to obtain a more complex skill called approaching which permits the robot to go towards a visual target. The vision system is the only mean by which the robot is able to obtain information from environment. The used learning algorithm allows to work with continuous state and output spaces. When the robot performs an action, it receives a reinforcement signal, so that, at each moment the robot knows wether it is performing the action correctly. Copyright (C) 2 001 IFA.