An AI pilot called Swift has triumphed over professional human drone pilots in a head-to-head championship.

Swift, a creation of international researchers, demonstrated its prowess by winning 15 out of 25 races against three seasoned human champions, including two world champions from international leagues.

The competitive arena in question is first-person view drone racing, a high-speed sport where pilots navigate drones through complex 3D circuits using live video feeds from onboard cameras. 

The challenge lies in pushing the drones to their physical limits while simultaneously processing data from onboard sensors to gauge their speed and position within the racecourse.

To bridge this gap, Elia Kaufmann and a team of researchers devised a system that blends deep reinforcement learning in simulations with real-world data. 

The real-world testing pitted Swift against human champions who had one week of practice on a track designed by a professional drone racer. 

Swift not only outperformed its human rivals but also set a record by completing the course half a second faster than the fastest human pilot.

This milestone highlights the potential of hybrid learning-based solutions in other physical systems, such as self-driving ground vehicles, aircraft, and personal robots. 

However, experts caution that further advancements are necessary to handle unpredictable real-world elements like wind, shifting light conditions, less-defined gates, and competing drones. 

These challenges remain substantial hurdles for existing AI techniques in the pursuit of conquering human pilots in various racing environments.

More details are accessible here.