Optimal multi-agent path finding algorithms computing collision-free space-time trajectories for over 1,000 concurrent autonomous agents.
MassSwarm splits path planning into a high-level conflict tree that identifies vertex and edge collisions between robot trajectories, and low-level single-agent A* searches that resolve collisions by imposing space-time reservation constraints.
To handle unforeseen wheel slip or mechanical delays, paths are continuously replanned over a sliding 20-second time window, guaranteeing real-time response times under 50 milliseconds across the entire mega-fleet.