Pilotage reactif d'un robot mobile : etude du lien entre la perception et l'action

摘要

The objective of this work is to compare the main existing techniques to synthesize perception to action mappings for reactive control of a mobile robot towards a goal in an unknown environment. This work is motivated by the limitations of the navigation system created for the European Research Project Mithra Eureka EU:110 which is based on the execution of a planned path from a parametric model of the observed environment. The first mapping we consider is based on force fields derived from potential fields. We propose an heuristic approach generating continuously translation movement to try to reduce the number of local minima. As all the problems are not solved, we have then considered a second family of transformation generators~: fuzzy logic. The experiments we have made leads to the conclusion that problems are of the same nature as force fields problems. The second part of the manuscript deals with learning control techniques in mobile robotics. We have more precisely considered two main families of approaches : supervised and reinforcement learning. We first remind the robotic particular constraints that a selected learning algorithm must be able to deal with. We then propose for supervised learning a (\em Grow and Learn) type network coupled with a principal components analysis of sensory data. We then propose a new learning algorithm for fuzzy rules which can be used in an Adaptive Heuristic Critic approach on reinforcement learning.