UZH’s AI-piloted drone beats world champions

The Swift AI system developed at the UZH has won several drone races against three world champions. Its drone competed against Drone Racing League Champion and World Champion 2019 Alex Vanover, MultiGP Drone Racing Champion 2019 Thomas Bitmatta, and three-time Swiss champion Marvin Schaepper, winning several drone races in a hangar at Dübendorf airport in the canton of Zurich. These races involved flying through several targets in a certain order, including challenging maneuvers and an acrobatic exercise.

According to a press release from the UZH, until recently autonomous drones without an external positioning system took twice as long as those flown by humans. In contrast, Swift reacts in real time to data collected by an onboard camera.

The system was trained in a simulated environment. Swift taught itself to fly using a special type of machine learning called reinforcement learning. Elia Kaufmann, first author of the paper, commented: “To make sure that the consequences of actions in the simulator were as close as possible to the ones in the real world, we designed a method to optimize the simulator with real data.”

The UZH drone was half a second faster than the others but failed when conditions were different to those it was trained on, for example when it was too bright in the room. The researchers believe that increasing the efficiency of drones is important for rescue operations, enabling large areas to be traversed in as short a time as possible, such as when monitoring forests or in space. ce/mm