Watch this Robot Get Back on its Feet

If you’ve ever watched a three-legged dog run around like nothing was amiss, or hurt your leg and hobbled on through the day, you know that animals can adapt to an injured or lost limb pretty quickly. Robots, though, are much slower at this, if they can do it at all. They can’t “‘think outside the box’ to find a compensatory behavior when damaged,” say robotics engineers Antoine Cully, Jeff Clune, and Jean-Baptiste Mouret. They either have to rely on trial-and-error learning algorithms—which can take hours to find a new gait that works with the injury—or the contingency plans their designers built in, which can’t foresee every possible injury or situation. 

“Damage recovery would be much more practical and effective if robots adapted as creatively and quickly as animals,” the engineers say. And that’s what they’re trying to achieve with a new “intelligent trial-and-error” algorithm.

Instead of leaving a robot to test new behaviors or gaits at random or running through small modifications to ones that seem to work best, the new algorithm has a “behavioral repertoire that contains predicted performances for thousands of different behaviors,” built before the robot is deployed. If it’s damaged, the robot can try a behavior from the repertoire that’s predicted to perform well. After testing that, it can update the predictions for that behavior and similar ones on the fly, narrowing down its choices. The robot’s trial and error process is much faster—taking as little as one to two minutes in some tests—thanks to “intuitions” about how its body works that were developed before it even set off on its business. 

The result, the team says, is a “creative process that adapts to a variety of injuries, including damaged, broken, and missing legs” and could allow for more reliable and robust robots outside of controlled environments. 

In the video below you can see one of the team’s robots adjust to a leg that’s lost power and a leg that’s broken by trying a few gaits from its repertoire.