Decentralized coordination

We tackle hard coordination problems like multi-agent path finding, where hundreds of robots must plan collision-free routes at once. Central planners do not scale to these team sizes, so we instead train policies where each agent acts on its local observations alone — approaching the performance of optimal solvers while running in real time on-board.

A recurring finding is latent resilience: when identical robots learn distinct, complementary roles rather than converging to a single strategy, the team absorbs disruptions much more gracefully. Getting these decentralized policies off the simulator and onto real hardware is a research problem in its own right, and our sim-to-real pipeline — validated on the Cambridge RoboMaster platform and on our indoor multirotor swarm — is how we close that loop.