Boston Dynamics Puts Wheels on Atlas, Lands a 360-Degree Jump

How the robot learned

The sideways gliding came from AI: the system learned on its own how to use the wheels. The 360-degree jump relied on recorded human motion data, and according to the researchers it succeeded on the first attempt.

The team trained the robot through millions of simulations in a virtual environment. The movements learned there were then reproduced on the physical robot almost unchanged.

A test, not a product

The purpose of the experiment is not to commercialize a wheeled robot. Instead, it was designed to probe the limits of Atlas's motors and drive systems and to validate AI-based control techniques.

Boston Dynamics says it is also applying what it learned in this work to improve the control technology of the standard Atlas. That makes the wheeled version less a preview of a new model than a testbed whose results feed back into the existing humanoid.

The demonstration also shows how far simulation-trained control has come: behavior learned entirely in a virtual world, including a jump with a full rotation, transferred to hardware with little adjustment.