Researchers at Delft University of Technology (Netherlands) have created an autonomous drone capable of competing on equal terms with the best human pilots. Their MonoRace system navigates space using only the image from a single camera and data from a built-in motion sensor - an inertial measurement unit (IMU, a device that records acceleration and rotation). No GPS, no external cameras and no powerful onboard computer are required.
The work was published in the journal Science Robotics. The main idea is to test how far artificial intelligence can be pushed under strict constraints on sensors, computing power and energy. At high speeds, even a small delay in recognizing the environment or in control leads to a collision or loss of control.
A neural network onboard recognizes racing gates in the image, while the IMU together with a model of the drone itself helps estimate its position and motion. All of this works without satellite navigation and without external tracking systems. As one of the authors, Stavros Bachnam, explains, the difficulty is that at speed the image becomes blurred and distorted, and the AI has very little time to determine its location accurately. The researchers taught the system to reliably combine camera data with information about the drone's own motion, as well as to cope with interference and sensor overload that occur in real flight.
Control was also restructured. Instead of the classic multi-level scheme with intermediate controllers, a small neural network issues commands directly to the motors. It runs on a miniature chip - the flight controller - and updates commands 500 times per second. The neural network was trained using reinforcement learning: a method in which a program learns by trial and error, receiving a conditional reward for successful actions. The drone's physical constraints were taken into account during training.
The system was tested under harsh conditions - at the international autonomous drone racing competition in Abu Dhabi in 2025. Importantly, both the track and the racing drone itself were prepared by the organizers, not the researchers. That is, the system had to work on equipment unfamiliar to it and on a track the team did not control. Despite this, the TU Delft drone accelerated to 100 km/h, beat all other autonomous teams and three former human world champions.
According to Christophe De Wagter, the goal was not to increase the power of sensors and computers, but to understand how much intelligence can be extracted from very modest resources. Drone racing provided an extreme venue for testing ideas outside the laboratory.
Professor Guido de Croon emphasizes: autonomous racing is a way to push the limits of robotic AI. In this work, the AI was optimized for speed, but the same approach, according to the authors, can be applied for efficiency or safety - for example, for smooth, energy-optimal control of robot vacuum cleaners or huge port cranes. The researchers believe the principles will help many autonomous systems - from vehicles and mobile robots to future flying and humanoid machines - make decisions quickly with limited onboard resources.
It should be noted: this is a laboratory development, not a finished product. The source materials say nothing about when such systems will go on sale.