Faster Than a Baby: Self-Taught Robot Walks by Itself In Mere Hours

Babies start walking at 14-15 months, generally speaking. Some take more time to get bold and brave the world on two feet. So it’s no wonder robots wobble, trip and fall over long periods of time before they learn to do the same and are always helped by humans. But to be truly useful to us in the future, they have to be more independent. So researchers found a way to make a robot teach itself how to walk.

Scientists from the University of California, Berkeley and Google Brain, published a paper describing the process they went through to make a robot teach itself to walk and its results. No human intervention in sight.

In just two hours, a four-legged bot managed to walk forward and backward, to the left and to the right. Researchers claim they intervened minimally, with just a few algorithm tweaks here and there.

From the start, they opted to teach the robot in the real world, instead of a simulated environment. The algorithm they used was more efficient than what was available until that point, meaning there were fewer trials and errots. Although the robot adapted quickly to inclines, steps and obstacles on a flat ground, at first it still required some babystitting. So, they tried another parenting method – bounding the terrain so the robot would be first to walk backwards if he hit an obstacle as it went forward.

By constraining the robots movements, the machine was forced to learn by itself and fast. Soon, it was able to walk on the ground, a memory foam mattress, and an uneven doormat.

This is a huge milestone for robotics, taking us closer to a future where robots will outgrow their child-like needs, becoming independent. Until then, there’s still a way to go. This achievement, for example, was possible in a controled environment with a motion capture system above the robot to keep an eye on its position at all times which, of course, doesn’t happen in real life situations.