AI-Powered FPV Drone Outpaces Human Racers

The field of artificial intelligence (AI) is showcasing its prowess, with robots besting humans in increasingly complex tasks. A groundbreaking example emerged from the University of Zurich, where engineers unveiled an AI-driven drone, named “Swift“, that dominated human opponents in competitive drone racing.

Previously, human skills and intuition were unmatched in the exhilarating realm of drone races, which saw its inception back in 2011 at the hands of German hobbyists. The recent developments, however, signal a paradigm shift. Swift was meticulously crafted using deep reinforcement learning (DRL), a sophisticated AI technique.

Through a reward-based system—where the ‘reward’ is victory in the race—the AI learns optimal actions via repeated trials. Real-world data was incorporated into simulations, enabling Swift to replicate real-life racing scenarios.

Elia Kaufmann, the lead author and autonomy engineer at the University of Zurich, conveyed the significance of this achievement, noting, “Our result marks the first time that a robot powered by AI has beaten a human champion in a real physical sport designed for and by humans.”

Kaufman further elaborated that such advancements in hybrid learning could propel innovations in other physical systems, including self-driving vehicles and personal robots.

The most commendable feature of Swift is its autonomy, as Kaufmann points out, “Swift uses onboard sensors, therefore it is fully autonomous. You don’t need to first instrument the environment with external sensors; instead, just turn on the robot and fly!”

To test Swift’s capabilities, a racetrack was masterminded by an expert drone-racing pilot. Here, Swift faced off against three eminent human drone racers. Given a week to familiarize themselves with the track, the human contenders were still bested by Swift in 15 of the 25 head-to-head contests. Astonishingly, the AI-powered drone set a track record, outpacing the fastest human participant by half a second.

Kaufmann highlights the broader implications of this success, suggesting its potential in “autonomous infrastructure inspection, autonomous warehouse management, or search and rescue scenarios.”

Yet, before visions of AI drones ruling the skies surface, it’s crucial to recognize Swift’s limitations. Human pilots, for instance, recover rapidly from mishaps such as collisions. Swift, in contrast, fumbles significantly when faced with unexpected physical changes, like abrupt weather shifts.

As Guido de Croon, a non-participant robotics researcher from Delft University of Technology in the Netherlands, reportedly noted, “The real world is also much more unpredictable than a virtual one.”

Despite these challenges, the strides made with Swift provide a tantalizing glimpse into the future of autonomous robotics. This milestone not only redefines the boundaries of competitive drone racing but also hints at the untapped potential of AI across myriad applications.

It will probably not come as a surprise that Kaufman, who has a Ph.D. in robotics, joined the US-based drone maker Skydio in March 2022.

Photos courtesy of Leo Bauersfeld.