To make robots more graceful, scientists shove them to the ground, repeatedly

Scientists at Georgia Tech created an algorithm that allows robots to minimize the amount of damage they suffer in a fall. The team leader, Karen Liu, says that one day hardware advances will allow robots to land on their feet the way cats do.

Humanoid robots have made incredible progress over the past several years. Robots have learned to navigate uneven terrain, been taught karate kicks, and even shown signs of self-awareness. Now, scientists at the Georgia Institute of Technology are trying to teach humanoid robots a different kind of skill: how to fall gracefully.

Humans and animals don’t usually think consciously about how to fall. When we find ourselves teetering off-balance, instinct takes over and we automatically assume positions that cushion our heads and internal organs.

Georgia Tech professor Karen Liu and her research team wanted to see whether robots could be taught the same instinctive behavior, allowing them to fall more gracefully – or at least in a way that would minimize damage to sensitive components.

The researchers studied the way cats fall, and the way divers and astronauts orient themselves in microgravity, to better understand how robots can recover from stumbles. Then they shoved robots to the ground, over and over again, to help them learn how to mitigate falling damage.

“It’s not the fall that kills you. It’s the sudden stop at the end,” Dr. Liu says, in a statement. “One of the most important factors that determines the damage of the fall is the landing angle.”

The team developed an algorithm that allows robots to determine the angle and speed at which they are falling, and to quickly orient themselves to minimize damage on impact. The major limitation is hardware – many humanoid robots simply can’t move their limbs quickly enough when their computers recognize that a fall is imminent. So the team adapted the algorithm to recommend movements that would slow or cushion the fall, rather than trying to prevent it altogether.

One of those movements is an awkward somersault, in which the robot tries to roll out of a fall rather than crash into the ground. The result isn’t particularly acrobatic, but it’s a big improvement over simply remaining motionless and risking damage to sensitive components in a spill.

Liu hopes that her team’s research will help search-and-rescue robots to navigate hazardous conditions more gracefully.

The research being performed by the Georgia Tech team is an intermediate step. Right now, their goal is to allow robots to recover from a fall with as little damage as possible. But as motors and other hardware continue to advance, robots will eventually gain the ability to reorient themselves almost instantaneously, so that they suffer no damage at all from falls.

Cats have this ability naturally – they can twist around in midair to land on their feet – but human joints aren’t as flexible, so we sometimes get injured from falls.

“If we believe that one day we will have the capability to build robots that can do this kind of highly dynamic motion, we also have to teach robots how to fall — and how to land, safely, from a jump or a relatively high fall,” Liu says.