Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning

Researchers used league-based self-play to train quadrotor racing agents that can handle overtaking, collision avoidance, and other multi-racer interactions.

The drones reached speeds above 22 m/s, beat a champion-level human in multiplayer races, and cut collisions by 50% versus strong single-agent baselines.

The approach also generalized more safely to human interaction, suggesting interactive multi-agent training can improve both performance and safety in robotics.