Enabling Stochastic Dynamic Games for Robotic Swarms

This article scales stochastic dynamic games to large swarms of robots through selective agent modeling and variable partial belief space planning. We formulate these games as a continuous partially observable Markov decision process (POMDP) to find a local Nash equilibrium by using a belief space variant of iterative linear quadratic Gaussian (iLQG). Prior work focused on small team sizes, making it intractable for swarms. We break this notion by selectively choosing agents to model, based on the estimated influence between agents and integrated into the ego agent's action value function. The ego agent prioritizes modeling influential agents with stochastic game‐theoretic reasoning, while less‐influential agents are modeled without game‐theoretic reasoning and with assumed deterministic states. Simulations show that our selection method both improves computation times and improves performance compared to ego agents attempting to plan over all agents. We validate our approach in simulation studies over three case studies that illustrate the flexibility and scalability across heterogeneous multirobot swarms.