Learning Decentralized Controllers for Segregation of Heterogeneous Robot Swarms with Graph Neural Networks

In this paper, we studied the problem of finding decentralized controllers for large-scale heterogeneous robot swarms exhibiting segregative behaviors. As seen in nature, Segregative behaviors involve sorting a group of robots into groups based on their type. Our approach involves learning controllers that utilize local information at test time by imitating the policy of a centralized controller based on a differential potential concept at training time. We parameterized our policy using a time-varying aggregation graph neural network with multi-hop communication. This incorporates information not only from immediate neighbors but distant neighbors. We showed that our controller outperformed a local controller that considers only immediate neighbors and achieved similar performance to the centralized controller through varied experiments. In addition, we demonstrated the scalability of our method by exploring larger swarms and different groups.