A game theoretical approach to self-assembly in swarm robotics

This paper focuses on coordination mechanisms for large multi-agent systems (MAS) where the agents have limited ability in sensing and computation. To address this problem, we propose an approach inspired by mean field games where each agent can compute wonderful life utility without simulating the behavior of other agents. This provides a scalable solution that enables a cooperative MAS to make rational decisions and to achieve a local optimum of the global objective. We also provide a general method to enable the self-organization of agents into groups when the set of subtasks is not given. Experimental results on a formation problem show that the division of groups emerges successfully, and our method achieves higher efficiency than classic reinforcement learning.