The Swift AI has beaten expert drone racers in high-speed races using an on-board computer that fuses artificial intelligence and classical algorithms – a method that could speed up delivery drones.
An artificial intelligence has consistently beaten champion drone pilots in races for the first time, achieving lap times no human was able to match. The technology could be used to speed up drones carrying out everyday tasks.
 |
| Swift (blue) races against the drone (red) of Alex Vanover, the 2019 Drone Racing League world champion Leonard Bauersfeld |
The sport of drone racing involves humans piloting small quadcopters around a course at speeds of more than 100 kilometres per hour, with the vehicles subject to g-forces of up to 5 g. The drones must fly through a series of gates in the correct direction and order, and the people steering them wear headsets streaming video captured by cameras on the drones.
Leonard Bauersfeld at the University of Zurich in Switzerland and his colleagues have created an AI called Swift to compete in these races. It managed to beat three top-level human pilots, including Alex Vanover, the 2019 Drone Racing League world champion, 60 per cent of the time.
Human pilots were given a week to practise on the specific race track and Swift was also allowed to train on it before they raced. The AI won 15 out of 25 races and achieved the fastest recorded time over the course, beating the closest human time by half a second.
Bauersfeld says AIs have only beaten human performance in drone racing before when they have benefited from dozens of cameras placed around the course and an external computer transmitting instructions to the drone in real time.
“That’s a bit of an unfair advantage,” says Bauersfeld. “With millimetre precision and really high update rates, like 400 times a second, you know exactly where the drone is located in space and also how it is oriented.”
Swift uses only a single camera and an on-board computer, making it a totally encapsulated system. Bauersfeld says Swift is designed to fly at about 99.5 per cent of the speed it could manage, in order to leave a safety margin and improve reliability.
Swift is also unusual in that it fuses AI and classical computer algorithms. First, an AI takes video from its camera and identifies where the gates are. Then, a classical algorithm determines the drone’s orientation and position given those gate locations. Another AI takes that data and determines where the drone should fly next and what control inputs are needed to achieve that. “We fuse the two [computational approaches] in order to get something that is better than just one,” says Bauersfeld.
Guido de Croon at Delft University of Technology in the Netherlands says the work is impressive because all the sensors and computer power are physically carried. AIs that have beaten people at other games, such as Go, StarCraft or Atari games, could use unlimited computational resources. They could also operate virtually, with people moving pieces for them with the physical games.
De Croon says there are few applications that require the high speeds involved in these experiments – outside possible military use – but the work will have other applications.
“I think, mostly, you don’t need that super high speed, ever. But you do need to fly faster than drones have been flying. If you fly slowly, you waste your battery,” he says. “If you want to do real-world missions, to use your battery to the best, to use the energy for actually doing your task as a robot… you need to fly a bit faster. And this work helps solve those problems.”
Journal reference