Robots may not feel pain or embarrassment, but they still have good reasons to avoid a wipeout. Their parts are expensive, for one thing, and they’re lousy at healing their own scrapes. And robots that walk around on two legs are bound to take some spectacular spills. That’s why researchers are working on a way to teach robots to fall a little more gracefully.
At the Georgia Institute of Technology, Sehoon Ha and Karen Liu created an algorithm that plans the safest way for a robot to fall. Pitching directly onto your face, whether you’re a bot or not, isn’t ideal. Hitting the ground in multiple smaller impacts lets you get rid of your momentum with less damage to your body. When humans trip on an uneven sidewalk, say, we do this instinctively by putting out our arms to break our fall.
The algorithm starts with a robot’s velocity at the beginning of its tumble—this would be a human’s “whoops!” moment just after tripping—and calculates the best way for the robot to fall. How can it use its limbs to make the series of contacts with the ground that will cause the least damage? Ha and Liu first used their algorithm to determine the ideal falls for two types of robot under different scenarios. Then they tested some of the algorithm’s pre-computed solutions in an actual robot. They presented their research at a conference earlier this month.
Here’s a robot reacting to a shove as planned by the algorithm (top) and taking the same shove normally (bottom):
Ha and Liu tested several scenarios, asking the algorithm what a robot should do if it were mid-stride and found itself falling forward with a certain momentum. With more momentum, the robot could soften its fall by using more contact points. (In an infinitely soft fall, the robot would be a sphere, touching the ground at an infinite number of points as it rolled away. Math!)
“We were certainly inspired by judo and other human motions that require great agility, such as parkour or gymnastics,” Liu says. Ha adds that what the algorithm does is similar to how a judo student practices breaking falls, a skill called ukemi.
When the researchers tested the results in actual robots, the machines managed to get out of their falls gracefully. The robot below, taking an especially big shove, executed a beautiful forward roll. Although the roll looks a lot like ukemi, Liu says, the strategy emerged from the algorithm naturally.
This tool isn’t ready for robots to use in the real world yet. That’s because in simulations, it takes the algorithm between 1 and 10 seconds to plan the best fall. It’s not a simple math problem.
Yet Liu says it’s “highly possible” that robots could use this type of algorithm once their computing power improves. Even before that, she says, a robot could use the algorithm to plan ahead for a number of likely falling scenarios. Then, when the robot is actually falling, it could quickly choose the solution that best matches the circumstances—whether it’s facing a sidewalk bump, a landmine, or a pesky researcher giving it a shove.
Top image from Ha & Liu, 2015. GIFs from Georgia Tech News Center video.
Sehoon Ha and C. Karen Liu (2015). “Multiple Contact Planning for Minimizing Damage of Humanoid Falls.”
Help do some science! Want to be part of a scientific study without leaving your chair or sniffing a poop stick? I’ve teamed up with researcher Paige Brown Jarreau to create a survey of Inkfish readers. By participating, you’ll be helping me improve Inkfish and contributing Paige’s research on blog readership. You will also get FREE science art from Paige’s Photography for participating, as well as a chance to win a t-shirt and other perks. It should take 10-15 minutes to complete the survey, which you can find here: http://bit.ly/mysciblogreaders. Thank you!!