Drone AI beats world champion on the track

On April 14, two drone racing events took place simultaneously: the Falcon Cup Final for pilots and the A2RL Drone Championship for autonomous drones controlled by AI, according to Phys.org. Excellent AI drones also competed with the best pilots. Drone AI developed by Delft University of Technology first won the A2RL Grand Challenge, then went on to win the knockout tournament against three DCL world champions, reaching flight speeds of up to 95.8 km/h on a winding track.

A team of scientists and students from Delft University of Technology achieved the above achievement by developing an efficient and powerful AI system that can control quickly and accurately. While previous AI breakthroughs such as defeating world champions in chess or Go often took place in virtual environments, this achievement took place in the real world.

Two years ago, the Robotics and Perception group at the University of Zurich was the first to defeat drone racing champions with autonomous drones. However, that impressive achievement took place in a flying laboratory environment, where conditions, hardware and tracks were still controlled by researchers, very different from this world championship where the hardware and tracks were completely designed and managed by the organizers.

The goal of the 2025 A2RL Drone Championship in Abu Dhabi is to push the boundaries of physical AI by encouraging research in AI robotics under extreme time pressure, with very limited computational and sensor resources. The drone only uses a forward-facing camera, a big difference from previous autonomous drone races. It operates more like how a pilot flies, creating additional cognitive challenges for the AI.

The AI ​​that beat three former world champions DCL was developed by a team of scientists and students from the MAV Laboratory of the Department of Aerospace Engineering of Delft University of Technology. One of the core new elements of AI drones is the use of a deep neural network that does not send control commands to a traditional controller but directly to the motor. These networks were initially developed by the Advanced Concepts Group at the European Space Agency (ESA) under the name "Guidance and Control Networks".

Human-designed optimal control algorithms consume too much computational resources to run on limited systems such as drones or satellites. ESA found that deep neural networks can mimic the results of traditional algorithms but require significantly less processing time. Testing whether these networks work well on real hardware is difficult, so ESA teamed up with the MAV Lab at Delft University of Technology.

"We train deep neural networks with reinforcement learning, a form of learning through trial and error," said team leader Christophe De Wagter. "This allows the drone to get closer to the physical limits of the system. To achieve that, we had to redesign not only the control training process but also how we learn about the drone's dynamics from onboard sensor data."

Highly efficient AI developed for robust perception and optimal control is not only important for autonomous racing drones but can also be used for other robots.

According to De Wagter, AI robots are limited by the computational resources and energy required. Autonomous drone racing is an ideal test case for developing and testing highly efficient and robust AI. Flying drones faster is important for many economic and social applications, from timely delivery of blood samples and defibrillators to searching for people in natural disaster situations. Furthermore, we can use development methods to achieve not only optimal time but also other criteria such as energy or safety. This will impact many other applications, from robot vacuum cleaners to self-driving cars.

An Khang (According to Phys.org)