This New Robot Is Able to Teach Itself How to Walk

Most animals don’t need to be taught how to walk, because they’re able to figure it out on their own. Many animals are abandoned by their parents immediately, and they’d never survive if they couldn’t learn important skills without help. Robots, on the other hand, almost always need to be given explicit instructions on how to walk — or to do anything else. Clever programming and artificial intelligence can help them respond to unpredictable circumstances, but that doesn’t mean they can learn to walk without assistance. This new robot built by Google is different, and can teach itself how to walk within a matter of hours.

This is a quadruped robot with four simple legs, and — mechanically at least — it doesn’t really stand out. Many other quadruped robots, such as Boston Dynamics’ Spot, are far more impressive from a mechanical engineering perspective. This robot is special because it taught itself how to walk without any explicit programming on how to do so. Its sophisticated machine learning algorithms help it to learn through trial and error, just like an animal would. At first it stumbles, falters, and often falls over. But eventually it figures out what not to do, and is able to walk around its little learning area.

This isn’t the first time that researchers have tried to give a robot this ability, and many approaches have been taken. One common method is to perform the trial and error in a simulation, which means the robot can very quickly try many walking techniques without risking damage. But those simulations have to be basic, and don’t do a great job of representing the real world. This robot does all of its learning in reality. It also doesn’t require human intervention during the learning process. It has been explicitly programmed to do two things: avoid leaving the learning area and pick itself back up again if it falls. Other than that, the robot gains the ability to walk purely through reinforcement learning. Currently, the robot does require an external motion camera system to perceive what it is doing, but the team hopes to make that unnecessary in the future.