Visual tracking skill reinforcement learning for a mobile robot

In this paper, we implement a reinforcement learning algorithm that allows an autonomous mobile robot, with a single CCD camera mounted on a pan-tilt platform, to learn simple skills such as watch and orientation, and to obtain the skill called approach combining the learned skills previously. The results obtained show the advantages of 1) decomposing complex skills in simpler skills due to the learning rate improves, 2) using the reinforcement as a learning mechanism because of the possibility to learn on-line without the robot having a previous knowledge of what it has to do for a given situation. We also present the neural network architecture used for the learning mechanism implementation. Copyright (C) 2001 IFAC.