Reinforcement Learning based Omnidirectional Vision Agent for Mobile Robot Navigation

This paper presents an Omnidirectional Vision Agent able to learn to navigate a mobile robot in its working environment. The novelty of the work is the application of Reinforcement Learning paradigm to Vision Agents aiming to develop a totally autonomous system able to learn control policies by on-line learning, to deal with changing environment and to improve its performance dur- ing lifetime. SARSA(λ) method is used by the Vision Agent to learn the con- trol policy for the robot. The LEM strategy is also applied to speed up learning. The knowledge acquired by one Vision Agent is then “copied” to another Vision Agent in a network of cameras implementing a Distributed Vision System (DVS). By copying the knowledge the aim is both reducing learning time and exploiting the knowledge already learned. Since our prime interest was to investigate how the Vision Agent learns the knowledge by using SARSA(λ) and to evaluate its performance, we carried out the experimentation in simulation. The good results obtained during the experimental phase are very encouraging to transfer all the experimentation in a real context.