Animals are amazingly adept at adapting to injury, but limping does not come so naturally to robots. A broken part can render a robot useless without a tech on standby.
With the help of a new computer algorithm, robots can now learn how to deal with damage all on their own. It's so effective that a robot with a broken leg can actually learn to run even faster than it did before the injury in a matter of minutes, according to a study published in the journal Nature on May 27.
"You can step on it and then in the course of one or two minutes, it figures out a new way to walk and then it's up and off to the races again," study co-author Jeff Clune of the University of Wyoming said in an interview.
Such resilient robots would be particularly useful in situations that involve dangerous conditions where it isn't safe to send in people. But robots run a greater risk of injury in these environments as well.
"Robots will provide tremendous benefits to society by doing jobs that are very dangerous for humans to do, like putting out forest fires and finding survivors after earthquakes, but they're not going to be very helpful unless they are able to soldier on when something goes wrong," Clune says.
The algorithm basically allows a broken robot to run a set of experiments on itself. If the six-legged robot shown in the video below loses a leg, for example, the pattern of movements that it employed to get around before will not work so well. But there are still thousands of possible different patterns of movements that use only five legs. Simple trial and error would eventually lead the robot to the best new strategy, but with so many possibilities this would take quite a while. This algorithm helps the robot streamline the trial and error process so that it can recover in minutes or even just seconds.
"You can already use our algorithm on some of the robots that are already deployed in the world," says Clune. "For example, in our paper we experimented on a robotic arm that has to place things in bins, so if some of the motors break it can still do its job."
Robots that can perform extremely complex tasks like looking for disaster survivors are still years away, but once they become a reality this algorithm will be able to help them use their advanced abilities in the face of damages or unforeseen conditions, according to Clune.
"The real payoff will be once we have robots that are more advanced," he says. "But to see it adapt in front of your eyes in a short amount of time, that's practical for the field and it's tremendously exciting."
The findings are detailed in the journal Nature.
Photo: Antoine Cully | Pierre and Marie Curie University
