Teaching drones to race safely by competing with each other

Autonomous systems trained in isolation fail in the real world because they ignore other actors. This paper trains quadrotor drones using multi-agent reinforcement learning through competitive racing, where agents learn to anticipate collisions, overtake strategically, and handle aerodynamic interactions with multiple competitors. The trained agents outperform elite human pilots in multi-drone races and—surprisingly—generalize to safer interaction with untrained human partners without any fine-tuning, suggesting that learning through realistic multi-agent competition builds robustness better than explicit safety constraints.