One Camera, One Tiny Brain: The AI Drone That Outflew World Champion Pilots

Drone racing is one of the hardest tests you can set for a flying robot. Pilots thread tiny quadcopters through gates at highway speeds, reacting in fractions of a second. Machines have beaten champion humans before, but only on tracks built for them and with help from outside tracking systems. A team at Delft University of Technology in the Netherlands now reports a drone that did it on a course the robot's builders did not control, using only what it carried on board. The study appeared in Science Robotics on October 14.

What the Delft team built

The system, called MonoRace, comes from TU Delft's Micro Air Vehicle Lab (MAVLab). The paper's authors include Stavrow Bahnam, Robin Ferede, Christophe De Wagter and Guido de Croon, among other members of the group. Their drone has no GPS and no external cameras watching it. It relies on one forward-facing camera and an inertial measurement unit, the same kind of motion sensor found in phones. De Wagter told Popular Science the goal was to see "how much intelligence we could extract from very limited resources."

A neural network reads the camera feed to spot the racing gates, and that estimate is blended with the motion sensor data so the drone knows where it is on the track. A second, very small network then sends commands straight to the four motors about 500 times per second, skipping the usual step of planning a path and then following it. That controller was trained with reinforcement learning, a trial-and-error method, entirely in simulation before it ever flew for real.

How it did in Abu Dhabi

The team entered MonoRace in a 2025 autonomous drone racing competition in Abu Dhabi, flying on a track and hardware supplied by the organizers. According to TU Delft, it reached speeds of up to 100 kilometers per hour (about 62 mph), beat all the other autonomous teams and posted faster runs than three former world champion human pilots. The researchers say it is the first AI-piloted drone to beat human pilots at an international race.

That matters because the best-known earlier result, the University of Zurich's Swift drone in 2022, beat humans on an indoor track it was designed and trained for. As Popular Science noted, Swift's performance suffered when conditions such as lighting changed.

The fine print

One camera is cheap and light, but at high speed its images become blurry and distorted, which is why the motion sensor is essential. The earlier preprint version of the work is open about other gaps: the drone has no way to avoid other drones, and a collision cost it in a multi-drone race. Its vision system also expects rectangular gates, and some fixes were tuned to this specific track and aircraft. Racing success has not yet carried over to larger robots in messier settings.