RL-trained drone racers beat a national champion, trained in 27 hours on one RTX 4090.

Zurich and DeepMind trained multi-agent RL quadcopter racers that outpace champion pilot Marvin Schaepper above 22 m/s while cutting collisions 50% versus single-agent baselines. The agents won 100% of 1v1 races (the human averaged 53.3%), learned blocking and wake-aware maneuvers through self-play, and needed just 200M interactions, about 27 hours on a single 4090. The caveat keeps it honest: policies ran off-board over the network, not on-device. Still, champion-beating real-world RL on consumer hardware is the number that sticks.