NASA drone race pits artificial intelligence against professional human pilot
On Oct. 12, researchers at NASA's Jet Propulsion Laboratory (JPL) in Pasadena, California conducted a drone race in which they timed laps through a twisting obstacle course as they raced UAS controlled by artificial intelligence (A.I.) against a world-class drone pilot named Ken Loo.
The race capped off two years of research into UAS autonomy funded by Google. Google was interested in JPL's work with vision-based navigation for spacecraft, which are technologies that can also be utilized by UAS. To showcase the team’s progress, a timed trial between JPL’s A.I. and Loo was set up.
After building three custom UAS— dubbed Batman, Joker and Nightwing—the JPL team developed the complex algorithms the UAS needed to fly at high speeds while avoiding obstacles. The algorithms were integrated with Google's Tango technology, which JPL also worked on.
Built to racing specifications, the UAS could go as fast as 80 miles per hour (mph) in a straight line. Being that the obstacle course was set up in a JPL warehouse, though, the UAS could only fly at 30 or 40 mph before they needed to use the brakes.
“We pitted our algorithms against a human, who flies a lot more by feel,” says Rob Reid of JPL, the project's task manager. “You can actually see that the A.I. flies the drone smoothly around the course, whereas human pilots tend to accelerate aggressively, so their path is jerkier.”
In comparison to Loo, the UAS flew more cautiously but consistently. Their algorithms are still a work in progress, as exemplified by the fact that sometimes, they moved so fast that motion blur caused them to lose track of their surroundings.
While able to reach higher speeds, and perform impressive aerial corkscrews, Loo was ultimately limited by exhaustion; an issue that he acknowledged afterwards, and an issue that the A.I.-piloted UAS did not encounter during the trial.
“This is definitely the densest track I've ever flown,” Loo says. “One of my faults as a pilot is I get tired easily. When I get mentally fatigued, I start to get lost, even if I've flown the course 10 times.”
The A.I. and human pilot started out with similar lap times, but after dozens of laps, Loo learned the course, allowing him to become more creative and nimble. Loo averaged 11.1 seconds for the official laps, while the autonomous UAS averaged 13.9 seconds.
The autonomous UAS were more consistent overall, though, as Loo’s times varied more, and the A.I was able to fly the same racing line every lap.
“Our autonomous drones can fly much faster,” Reid adds. “One day you might see them racing professionally!”
Below: JPL engineers recently finished developing three drones and the artificial intelligence needed for them to navigate an obstacle course by themselves. As a test of these algorithms, they raced the drones against a professional human pilot. Photo: NASA/JPL-Caltech.