In an extraordinary milestone, artificial intelligence has shattered a significant barrier by triumphing over three world-renowned drone-racing champions in multiple races. This marks a historic moment as it signifies the first instance where AI has outperformed human drone racing champions.
This groundbreaking AI system, named Swift, emerged from a collaborative effort between researchers at the University of Zurich in Germany and Intel. Swift demonstrated its prowess in first-person view (FPV) drone racing, a sport where pilots navigate quadcopters at staggering speeds, often exceeding 100 kilometres per hour.
A New Challenge for AI
Historically, AI has primarily limited its conquests to strategic board and video games, such as IBM’s Deep Blue defeating chess grandmaster Gary Kasparov in 1996 and Google’s AlphaGo prevailing over the top Go champion Lee Sedol in 2016. Physical sports present a more formidable challenge for AI due to their inherent unpredictability, unlike the structured nature of board 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.
Davide Scaramuzza – One of the Study’s Authors from the University of Zurich
Swift’s Real-Time Advantage
Until now, AI-driven autonomous drones took twice as long as human-piloted drones to navigate racetracks unless an external position-tracking system was employed for precise trajectory control. However, the recently unveiled Swift AI drone, featured in a new study published in the journal Nature, has demonstrated its real-time responsiveness to data captured by an onboard camera, akin to human racers in the sport.
Researchers revealed that Swift autonomously trained itself in a simulated environment through iterative trial and error. This simulation-based approach prevented the destruction of numerous drones during the early learning stages when system crashes were frequent.
To ensure that the consequences of actions in the simulator closely mirrored those in the real world, we developed a method to optimize the simulator using real data.
Elia Kaufmann – The Study’s Lead Author
Swift quickly prepared to challenge some of the globe’s top human drone pilots, such as 2019 Drone Racing League champion Alex Vanover, three-time Swiss champion Marvin Schaepper and 2019 MultiGP Drone Racing champion Thomas Bitmatta.
A Historic Race
In races conducted between June 5 and June 13, 2022, on a specially designed track within a hangar at Dübendorf Airport near Zurich, Swift achieved the fastest lap, securing a half-second advantage over the best lap achieved by a human pilot.
The unique track, measuring approximately 25 by 25 meters, had seven square gates that a participant had to navigate in a specific sequence to complete a lap. Drones faced demanding manoeuvres, including the acrobatic “Split-S” move, involving a mid-air roll and a high-speed descending half-loop.
While Swift recorded the fastest lap, researchers noted that humans exhibited greater adaptability than the AI drone, which struggled when faced with conditions differing from its training environment.
Nevertheless, this groundbreaking achievement in AI-powered flight extends far beyond the realm of drone racing. “Drones have limited battery capacity; they require most of their energy to stay airborne.
Dr. Scaramuzza
