Swift, an AI Drone Pilot, Beats the Best Human Pilots in the World of High-Speed Drone Racing
Beating human champions by a full half-second, Swift was trained through trial-and-error in just one hour on a relatively modest desktop PC.
Researchers at the University of Zurich and Intel Labs say they have reached a major milestone in high-performance autonomous vehicle operation: Swift, an artificial intelligence (AI) which can beat the best human pilots in championship-level drone racing.
"Physical sports are more challenging for AI because they are less predictable than board or video games. We don’t have a perfect knowledge of the drone and environment models, so the AI needs to learn them by interacting with the physical world," senior author Davide Scaramuzza explains of the challenges which have meant physical AI systems lag behind their board- and videogaming counterparts in besting humans.
Swift, though, aims to resolve that. Previous attempts to race autonomous drones against human pilots have seen the AIs taking twice as long to complete the course — unless bolstered by an external position-tracking system. Swift, by contrast, beats the best using only internal sensor systems: an inertial measurement unit (IMU) and a camera.
The Swift model was trained using simulation, learning to pilot a drone through trial and error — a key step to avoid burning through vast quantities of physical drones during the initial training period, the team explains. "To make sure that the consequences of actions in the simulator were as close as possible to the ones in the real world," notes first author Elia Kaufmann, "we designed a method to optimize the simulator with real data."
The simulated drone system was given, like its predecessors in the field, the advantage of an external position-tracking system — but after an in-simulation month, which passed by in just an hour using a relatively modest desktop PC system, the AI was ready to be moved to a real drone without training wheels.
Competing against three championship-winning human pilots on a course which included a challenging split-S maneuver, Swift achieved the fastest lap overall by a full half-second — though proved less able to accommodate environmental changes, losing its advantage when the lighting was too bright for the camera system.
While competing in races is one thing, Scaramuzza believes the Swift technology has potential for real-world impact too. "Drones have a limited battery capacity," he explains. "They need most of their energy just to stay airborne. Thus, by flying faster we increase their utility."
The team's work has been published in the journal Nature under open-access terms.