This chapter deals with the combination between multi-agent systems and evolutionary computation. The first section describes an experiment of the evolutionary learning of an autonomous obstacle avoidance behavior. This first experiment proves the possibility of distributing a genetic algorithm into a real robot population. In the proposed architecture, new crossovers and mutation techniques have been proposed allowing the distribution of the algorithm. The second part of the chapter deals with the decomposition of a unique robot into a multiagent system. In the distributed proposed architecture, the behavior of the robot is dependant from a set of coefficients influencing the motion of each agent. These coefficients are evolutionary optimized. Several fitness functions have been experimented: time, distance and energy. Simulation results show that minimizing the energy is the best strategy, the system may be fault tolerant and the computation of the algorithm is very fast. Future works will be oriented on a survey on the behavior of the robot around singular configurations and the implementation of a collision avoidance module, preventing the robot from hurting himself. So as to keep the advantages of the presented architecture, a nice strategy would be to distribute the obstacle avoidance module. This implementation looks like being really challenging, because it will highly increase the communication between the agents. Once the collision avoidance is functional, the implementation on a real robot will be planned.