TU Delft Autonomous Racing Drone Outpaces Former World Champions Using Single Camera and No GPS

Researchers at Delft University of Technology (TU Delft) in the Netherlands have developed an autonomous drone navigation and control system that outpaced former human world champion pilots on a physical racecourse using only a single forward-facing camera and no GPS, as reported by TechXplore (https://techxplore.com/news/2026-10-ai-drone-outpaces-world-champions.html).

Developed inside the MAVLab at TU Delft's Faculty of Aerospace Engineering, the framework—named MonoRace—eliminates reliance on multi-camera perception rigs and external motion-capture tracking systems. Instead, the drone pairs its single camera with an onboard inertial measurement unit (IMU) and a physical model of the aircraft to calculate orientation and position in real time.

High-speed autonomous flight presents severe sensing constraints: fast motion produces blurry, distorted video frames and leaves fractions of a second to interpret spatial features. MonoRace uses a neural network to detect racing gates in the camera feed, blending visual detections with IMU telemetry and model dynamics. Main author Stavrow Bahnam noted that the team designed a robust state-estimation pipeline to merge camera inputs with internal motion data, allowing the drone to navigate track environments despite image distortion, sensor saturation, and the absence of GPS signals.

The researchers also redesigned the aircraft's control architecture. Rather than routing perception outputs through multi-tiered controllers that compute explicit flight trajectories before issuing actuator signals, MonoRace runs a compact neural network trained via reinforcement learning directly on the drone's low-power flight controller chip. The network outputs direct motor commands 500 times per second, learning to maneuver the craft within its physical performance limits. Co-author Robin Ferede stated that bypassing intermediate trajectory-planning layers allows the controller to react rapidly while remaining small enough to execute on constrained onboard processors.

To evaluate MonoRace outside controlled laboratory settings, the team entered the platform in the 2025 Abu Dhabi Autonomous Drone Racing Competition. The contest required participants to fly on standardized hardware and unfamiliar tracks configured by the event organizers. MonoRace attained flight speeds of up to 100 kilometers per hour (62 miles per hour), placing first among all autonomous entrants and defeating three former world champion human pilots.

Researcher Christophe De Wagter explained that the project aimed to assess how much computational performance could be extracted from minimal hardware resources under extreme physical conditions. TU Delft professor Guido de Croon noted that while the team optimized MonoRace for racing speed, the underlying architecture of low-power direct neural control could be transferred to safety- or efficiency-focused robotics applications, including domestic mobile cleaners, industrial cranes, and humanoid systems.

The peer-reviewed study, titled "MonoRace: Robust monocular AI for champion-level drone racing," is published in Science Robotics.