AI beats pilots in drone racing

Using cameras and sensors, AI-controlled drones beat pilots at high speeds through obstacle-filled tracks.

The Swift autonomous system beat three professional drone pilots in 15 of 25 races on a track full of curves and obstacles designed by a professional drone racing pilot, Science Alert reported on August 31. The system combines AI algorithms with a camera and many built-in sensors to detect the surrounding environment as well as the drone's movements.

Swift was designed by Elia Kaufmann, a robotics engineer at the University of Zurich, and researchers at Intel Labs. They aimed for a system that would not rely on input from external moving cameras like previous autonomous racing drones.

“Reaching professional pilot level with an autonomous drone is challenging because it needs to fly within its physical limits, while estimating speed and position in the track using only onboard sensors,” the team said.

Pilots wear special goggles that provide a “first-person” view (like sitting inside the drone) through a camera mounted on the drone. The drone can reach speeds of 100 km/h.

Similarly, the Swift has a built-in camera and inertial sensors to measure the drone's acceleration and rotation. This data is analyzed by two AI algorithms to determine the drone's position relative to obstacles and issue control commands accordingly.

Despite losing 40% of the races, Swift beat the pilot several times and achieved the fastest recorded race time, half a second faster than the best human time.

“Overall, averaged over the entire race, the autonomous drone achieved the highest average speed, found the shortest route, and successfully maintained its performance close to its limits throughout the race,” Kaufmann and his colleagues said.

Swift’s real innovation, says Guido de Croon, a roboticist at Delft University of Technology in the Netherlands, is its second neural network, which uses deep reinforcement learning. Swift isn’t the first drone system to fly around obstacles, but it does so with remarkable accuracy. The new research is published in the journal Nature.

Thu Thao (According to Science Alert )

Source link