Résumé
The aim of this work is to build fault tolerant cooperative multi-robots systems. Our approach uses self-learning techniques to control groups of reactive robots. This work focuses on learning low level sensory-motor behaviors. It seems important that the proposed methods may be implemented on real robots. The constraint of such real systems is sometime far from simulated worlds. This is why we imagined, designed and build an experimental platform composed of four mobile robots, one miniature mobile manipulator and one stereoscopic vision system. This study is composed of two parts. The first one is applied to homogeneous systems. Evolutionist techniques are studied. The second one, applied to heterogeneous systems, focuses on using simulated annealing procedure to optimize the synaptic weights of a neuro-controller. Another method is also experimented, based on reinforcement learnin