Imitating Swarm Behaviors by Learning Agent-Level Controllers

Abstract

A main challenge in swarm robotics is the unknown mapping between simple agent-level behavior rules and emergent global behaviors. Currently, there is no known swarm control algorithm that maps global behaviors to local control policies. This paper proposes a novel method to circumvent this problem by learning the agent-level controllers of an observed swarm to imitate its emergent behavior. Agent-level controllers are treated as a set of policies that are combined to dictate the agent’s change in velocity. The trajectory data of known swarms is used with linear regression and nonlinear optimization methods to learn the relative weight of each policy. To show our approach’s ability for imitating swarm behavior, we apply this methodology to both simulated and physical swarms (i.e., a school of fish) exhibiting a multitude of distinct emergent behaviors. We found that our pipeline was effective at imitating the simulated behaviors using both accurate and inaccurate assumptions, being able to closely identify not only the policy gains, but also the agent’s radius of communication and their maximum velocity constraint.