In August 2023 a team at the University of Zurich published “Champion-level drone racing using deep reinforcement learning” in Nature. The system, called Swift, is an autonomous racing drone that flies a physical first-person-view racing quadcopter using only an onboard camera and inertial sensor, with no external motion-capture or off-board computing. The authors are Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Mueller, Vladlen Koltun, and Davide Scaramuzza.
Swift won real head-to-head races against three world-class human pilots: Alex Vanover, the 2019 Drone Racing League world champion; Thomas Bitmatta, a multi-time MultiGP International Open World Cup champion; and Marvin Schaepper, a multi-time Swiss national champion. Swift won several races against each pilot and set the fastest recorded lap, beating the best human time by about half a second.
The method combines model-free, on-policy deep reinforcement learning trained entirely in simulation with empirical noise models estimated from real flight data. Because a policy trained in a perfect simulator transfers poorly to a real drone, the researchers measured the gap between simulation and reality for both perception and dynamics, then folded those residual error models back into the simulator before deploying the policy unchanged on hardware.
For a general reader, this was a landmark demonstration that a learned controller can outperform expert humans not just in board games but in a fast, physical, real-world skill where split-second perception and control decide the outcome.