RL-Trained Drone Racers Beat a Five-Time National Champion — Trained in 27 Hours on a Single RTX 4090

University of Zurich / Google DeepMind (via Import AI 460)·medium signal

University of Zurich and Google DeepMind researchers trained multi-agent RL quadcopter racers that outpace champion human pilot Marvin Schaepper at speeds over 22 m/s while cutting collision rates 50% versus single-agent baselines. The agents won 100% of 1v1 races (human averaged 53.3%), learned blocking and wake-aware maneuvers through self-play, and required just 200M environment interactions / ~27 hours on one RTX 4090 — a striking demonstration of cheap, real-world-transferable RL. Caveat: policies ran off-board over the network, not on-device.