This video looks like a blooper real from a cartoon version of Terminator, but it is in fact a demonstration of a new algorithm from Georgia Tech that teaches robots how to fall properly.
Usually when a robot falls, it topples face-first into the dirt. Give it some stairs, or a bumpy yard, and most robots will eventually trip up and take a tumble. And if a robot has no good way to deal with the fall, it can’t take any measures to avoid damaging itself. Like this:
“A fall can potentially cause detrimental damage to the robot and enormous cost to repair,” says Georgia Tech PhD graduate student Sehoon Ha in a release about the project. “We believe robots can learn how to fall safely.”
Ha and his professor, Karen Liu, have worked out how to teach a robot to stumble instead, so it can fall without sustaining serious damage. The video shows these “planned falls” in action. When given a push, the algorithm-taught robot braces for impact with its hands, where the dumb robot slams face-first into the dirt. When the shoves get really nasty, the robot opts to roll, similar to a judo break fall. The result is still messy, but you can see that there’s no big impact in one spot.
“From previous work, we knew a robot had the computational know-how to achieve a softer landing, but it didn’t have the hardware to move quickly enough like a cat,” Liu says. “Our new planning algorithm takes into account the hardware constraints and the capabilities of the robot, and suggests a sequence of contacts so the robot gradually can slow itself down.”
We humans don’t really think about how we fall. We just do it. Or rather, we assess the situation so quickly, and so automatically, that we don’t notice until we’re laying on the ground with a few scratches instead of a few broken bones. Liu and Ha’s research teaches the robot to react as quickly.
Their robot can detect how many points of impact there are when it is pushed, what order they come in, how hard they are, and the timing between them. Using this information, it computes the best fall possible.
Liu’s previous attempts have failed, she says, because the hardware couldn’t keep up with what the software was telling it to do, something with which any aging human can sympathize.
Ha’s work adds these limitations to the software model, effectively teaching the robot about its own shortcomings. This makes it easy to imagine a future where robots will also need psychiatrists.