Researchers at the Delft University of Technology developed an autonomous quadcopter called MonoRace that completed an Abu Dhabi racing course faster than three former world-champion human pilots, achieving speeds up to 100 km/h using a single camera and an onboard computer.
An autonomous racing drone built in the Netherlands has outraced former world-champion human pilots during an international competition, marking a significant milestone for physical artificial intelligence. The aerial vehicle, developed at the Delft University of Technology, navigated a challenging course at speeds reaching 100 kilometers per hour without relying on external motion-tracking systems or GPS.
MonoRace Design at the Delft University of Technology
Engineers at the university’s Micro Air Vehicles Lab designed the quadcopter, named MonoRace, around a minimalist hardware philosophy. Rather than loading the aircraft with heavy sensor arrays, the team built a lightweight machine weighing 966 grams and measuring approximately 11.8 by 11 by 3.9 inches.
The drone relies on a single forward-facing rolling shutter camera and an inertial measurement unit to gauge acceleration and rotation. This limited hardware feeds data directly to an onboard NVIDIA Jetson Orin NX computer and flight controller. According to study author Stavrow Bahnam, processing high-speed video through a single lens creates immense perceptual hurdles.
Flying at very high speed with only one camera is challenging because the images are blurry, deformed and the AI system has very little time to precisely determine where the drone is.
Stavrow Bahnam, study author via Livescience
To overcome sensor limitations during aggressive maneuvers, researchers combined visual gate detection with a mathematical motion model. When rapid acceleration caused the motion sensor to exceed its measurement limits or electrical interference distorted the camera images, the system filtered out the erroneous data or substituted predicted acceleration values.
Direct Neural Network Control at Abu Dhabi
MonoRace broke from conventional autonomous drone designs by discarding multi-layered intermediate controllers that map out flight paths before issuing commands. Instead, a compact neural network running on the flight controller generates motor commands directly at a rate of 500 times per second.
Co-author Robin Ferede noted that instead of first planning a desired flight path and conventional commands, our neural network learns to directly control the motors
to execute rapid adjustments.
The team trained this controller using reinforcement learning inside a simulation environment that introduced variations in aircraft weight, drag, voltage, and propeller condition. That training allowed the aircraft to perform reliably when transferred to physical hardware.
At the 2025 Abu Dhabi Autonomous Drone Racing Competition, the quadcopter defeated every rival autonomous team and three former world-champion human pilots in a knockout tournament. Covering 11 gates across two laps in a time of 16.56 seconds, the aircraft reached a maximum velocity of roughly 63 mph.
Broader Significance for Physical AI and Robotics
This feat represents the initial occasion where artificial intelligence piloting a drone has triumphed over human operators within an uncontrolled international racing setting. While a University of Zurich team won a similar race three years earlier using an AI system named Swift, that victory occurred inside an indoor lab setting measuring 82 by 82 feet with optimized lighting and conditions.
Project leader Guido de Croon emphasized that drone racing serves as an extreme proving ground for physical artificial intelligence. He suggested the technology could eventually optimize everyday robotics for minimal energy consumption rather than maximum speed.
MonoRace could become a training tool for human drone pilots, a bit like powerful chess programs have also inspired humans playing the game.
Guido de Croon, professor at Delft University of Technology via Livescience
Despite its success, the system currently requires a previously mapped course and cannot yet detect or avoid other airborne drones, a limitation that caused a collision during the multi-drone event. Researchers believe the underlying principles will eventually benefit autonomous vehicles, inspection robots, and mobile systems operating without external positioning infrastructure.