An autonomous drone called Swift has beaten world-class human pilots on a racing course, showing how AI can navigate a fast-moving physical environment. Built by researchers from the University of Zürich and Intel, Swift won multiple races and set the event’s fastest lap, while also revealing limits in how well it handled unfamiliar conditions.
How drone racing puts AI to the test
In first-person view, or FPV, drone racing, pilots steer high-speed drones through an obstacle course while watching a live camera feed from the drone. The perspective puts the pilot virtually in the aircraft, and the goal is to complete the course as quickly as possible.
That makes the task more than simply recognizing a route. A system has to locate itself, spot the gates, and make rapid control decisions as the drone moves through the course. Swift was designed to do those jobs using onboard information, rather than relying on an external motion-capture system that earlier work by the researchers had needed to win.
What Swift sees and how it learns
A camera supplies real-time visual data, while an integrated inertial measurement unit tracks the drone’s acceleration and speed. An artificial neural network uses these inputs to estimate the drone’s position and identify the race gates.
A second deep neural network acts as the control unit, choosing actions intended to get the drone around the circuit quickly. The team trained the system with reinforcement learning in a simulation. Through trial and error, Swift learned how to navigate before competing on the physical course.
This combination matters because the system has to connect perception with action. Seeing a gate is only useful if the drone can estimate where it is in relation to that gate and adjust its flight accordingly.
A demanding course and experienced rivals
Swift faced three world-class pilots: Alex Vanover, the 2019 Drone Racing League champion; Thomas Bitmatta, the 2019 MultiGP Drone Racing champion; and Marvin Schaepper, a three-time Swiss champion. Their races took place between June 5 and June 13, 2022.
The specially designed course covered a 25-by-25-meter area and had seven square gates to pass through in a set order. It also included a “Split-S,” a maneuver involving a half-roll followed by a descending half-loop at full speed, according to UZH.
Swift won multiple races and recorded the fastest lap, finishing with a half-second lead over the best human pilot. That result marks a significant step for autonomous flight: the system could compete directly with skilled pilots using the fast visual judgments the sport demands.
Speed comes with a trade-off
The race also exposed a weakness. Swift struggled when conditions differed from those in its training, including when lighting changed. Its strong performance on the course did not mean it could adapt as readily as a human pilot to variation.
That distinction matters beyond racing. A system can be extremely fast in a known setting, yet less reliable when its surroundings shift. The contest showed progress in machine vision and autonomous control, while highlighting that responding to unexpected conditions remains a human strength.
The researchers see possible uses for faster drones in forest monitoring, space exploration, film production, and search-and-rescue missions. Flying quickly could help a drone cover more ground, a practical benefit when battery capacity is limited. These are potential applications described by the team, rather than results demonstrated by the race itself.
The researchers published their results in a Nature article titled “Champion-level drone racing using deep reinforcement learning.” Their work suggests that lessons from a tightly controlled sporting challenge may help guide autonomous flight, while the system’s limits point to the need for better adaptation when real conditions vary.