Abstract:

We propose a decentralized, multi-agent motion planner that guarantees probabilistic safety of a team subject to stochastic uncertainty in agent model and environment. Our scalable approach generates safe motion plans in real-time using off-the-shelf, single-agent reinforcement learning rendered safe using distributionally-robust, convex optimization and buffered Voronoi cells. We guarantee recursive feasibility of the mean trajectories and mitigate the conservativeness using a temporal discounting of safety. We show in simulation that our approach generates safe and high performant trajectories as compared to existing approaches, and further validate these observations in physical experiments using drones.