Researchers: AI Outflies Human Operators in Drone Races on Indoor Course
In yet another case of AI outperforming humans in a contest, an AI-operated quadcopter drone won in 15 of 25 races against drones piloted by top-rated human operators. This comes after AI has won at chess, Go and in poker and video games, according to a story in the academic journal “Nature.”
Npr.org reported on the contest, set in Switzerland, that paired drones with AI “brains” against drones flown by pilots with controls on an indoor race course.
“This is the first time that an AI has challenged and beaten human champions in a real-world competitive sport,” says Elia Kaufmann, an autonomy engineer at Skydio, a drone company based out of Redwood City, California, who worked on the drone while at the University of Zurich in Switzerland.
Kaufmann said this is significant because it’s harder to simulate real-world conditions in flying a drone than in playing a game on a board.
“This is called the sim-to-real gap,” he says.
Supervised and Reinforcement Learning
Researchers overcame the difficulties through “supervised learning,” or hand-training the drones with tens of thousands of images of the racing gates. The team also used “reinforcement learning” by putting the drone’s control code into a virtual version of the race course and training it virtually for the equivalent of 23 days (or one hour of computing time). The code practiced until it learned the fastest route.
While the success of the AI brain is unprecedented, it’s also only for a course it’s been trained on and under constrained conditions. Human pilots in the great outdoors would likely do much better than AI-driven drones. If another drone bumps into it, for instance, it gets confused and crashes.
According to Guido de Croon, a researcher at Delft University in the Netherlands, in the “Nature” article, the new technology has a way to go.
“To beat human pilots in any racing environment, the drone will have to deal with external disturbances such as the wind as well as with changing light conditions, gates that are less clearly defined, other racing drones and many other factors,” he writes.
The AI drones have a long way to go before they can be used in the field by the military, Kaufmann admitted.
read more at npr.org