Swarms Can be Rational

A fundamental challenge in multi-robot systems is spatial coordination (avoiding collisions) between robots, each under its own control. Swarm methods, where by robots coordinate ad-hoc and locally, offer a promising approach. However, while empirically demonstrated to be viable in practice, no guarantees of performance are known. This paper formalizes a class of multi-robot cooperative tasks as differential extensive-form games. We show that the system coordination overhead is a differential function, forming a connection between the theoretical maximum-payoff equilibrium of the system, and the rational self-interested choices of individual robots during task execution: robot swarms can be rational in theory. We then show how to approximate the rational decision-making in practice using reinforcement learning, using internal measures for rewards. We empirically show this leads to consistent optimal performance in with physical and simulated robots.