Researchers at the University of Zurich have created Swift AI which has beaten expert drones in high speed racing for the very first time. It managed to beat three top-level human pilots, including Alex Vanover, the 2019 Drone Racing League world champion, 60% of the time.

Created by Leonard Bauersfeld and his team, the AI, known as Swift, won 15 out of 25 races, setting a new course record with a half-second lead over the closest human competitor. According to Bauersfeld, previously AI could only outperform human drone racers by relying on an unfair advantage: numerous cameras strategically placed around the racecourse and an external computer providing real-time instructions to the drone. In contrast, Swift works with just a single camera and computer, making it an autonomous system.

“With millimetre precision and really high update rates, like 400 times a second, you know exactly where the drone is located in space and also how it is oriented,” said Bauersfeld in a recent interview.

The model employs deep reinforcement learning to determine the most effective commands to navigate the circuit at high speeds. Given the technique’s reliance on trial and error, the drone crashed several times during its training phase within a simulation.

During an actual race, Swift sends live video from its onboard camera to a neural network designed to identify the racing gates. This visual data is then combined with other information from an inertial sensor to calculate the drone’s precise position, orientation, and velocity.  After that a second neural network uses these estimations to determine the optimal commands to send to the drone.

Analysis of the races showed that Swift consistently outperformed human pilots at the race’s beginning point, executing tighter turns and achieving speed. Its fastest lap clocked in at 17.47 seconds, surpassing the fastest human pilot by half a second. Nevertheless, Swift faced challenges, losing 40% of its races against humans and crashing multiple times. It is also sensitive to varying environmental factors such as changes in lighting conditions.