AI Quadcopter Defeats Human Drone Racing World Champion

2023-09-01

Robots have achieved another victory. After successfully defeating humans in chess, Go, and various video games, AI has now triumphed over the world's top drone racers. Recently, three world champion drone pilots were defeated in a competition by an autonomous AI aircraft called Swift. According to a paper published on Wednesday in the journal Nature, this AI-equipped aircraft, developed by researchers at the University of Zurich, won 15 out of 25 races and set the fastest lap time.

The researchers wrote, "This work represents a milestone in mobile robotics and machine intelligence, and may inspire the deployment of hybrid learning solutions in other physical systems."

Swift has defeated humans in the emerging but rapidly growing first-person view drone racing sport. Human competitors navigate using cameras connected to the drones and head-mounted devices, aiming to pilot the quadcopters through complex obstacles in the fastest time possible while minimizing damage. The speeds of these drones in such races can reach up to 50 miles per hour (80 kilometers per hour).

Swift achieved 15 victories out of 25 one-on-one races against human pilots and recorded an overall fastest lap time of 17.47 seconds. This rapid lap time is nearly half a second faster than the best human performance. The three human competitors, Alex Vanover, Thomas Bitmatta, and Marvin Schaepper, have all previously won championships in drone racing.

In this scenario, human competitors had one week to learn the new track and train for the race. At the same time, Swift trained in a simulation conducted in a digital environment to mimic the track. According to the paper, Swift used deep reinforcement learning in the simulation, as well as additional data collected from the real world. If deep reinforcement learning sounds familiar, it's because it is the training method previously used by AI models to surpass humans in games like chess, Go, and even StarCraft.

However, drone racing is different from previous examples because it involves an object moving in the real world, not constrained by a game board or digital setting. In actual races, Swift receives video collected by cameras and sends it to a neural network that can identify the gates it must pass through. Then, a combination of onboard sensors assists the drone in positioning, speed, and direction. All of this happens autonomously at high speeds.

The researchers noted some interesting differences between Swift and its human competitors in handling the track. They observed that the autonomous system was more consistent in each lap and seemed to take tighter turns. These tight turns can lead to repeatedly shorter lap times, giving the drone an advantage in the race.

Swift's performance seemed to astonish the human champions.

During the race against Swift, Bitmatta exclaimed, "It's insane." "The possibilities are endless, this is a beginning that could change the whole world," Bitmatta added in the research paper.

However, others, like Schaepper, seemed to feel a sense of frustration.

He said, "Racing against a machine feels different because you know the machine won't get tired."

RECENT AI NEWS

RECENT AI TOOLS