Analytical and experimental results on multiagent cooperative behavior evolution

The paper addresses the problem of automatically programming cooperative behaviors in a group of autonomous robots. The specific task that we consider here is for a group of distributed autonomous robots to cooperatively push an object toward a goal location. The difficulty of this task lies in that the task configurations of the robots with respect to the object do not follow any explicit, global control command, primarily due to certain modeling limitations as well as planning costs as in many real life applications. In such a case, it is important that the individual robots locally modify their motion strategies, and at the same time, create a desirable collective interaction between the distributed robot group and the object that can successfully bring the object to the goal location. In order to solve this problem, we have developed an evolutionary computation approach in which no centralized modeling and control is involved except a high level criterion for measuring the quality of robot task performance. The evolutionary approach to distributed robot behavioral programming is based on a fittest preserved genetic algorithm that takes into account the current positions and orientations of the robots relative to the object and the goal, and a weak global feedback on the collective task performing effect in relation to the goal if some new local motion strategies are employed by the robots.