Atlas got scooter wheels for feet. The real test was the motors underneath

The clip is built to be shared: a humanoid gliding across a warehouse floor, dashing sideways, spinning a full 360 degrees in the air and landing on wheels. Boston Dynamics posted it on November 9 under the title Atlas on a Roll. The more useful part is the write-up beside it, because it explains why a company that has spent years teaching Atlas to walk would take its feet away.

“A wheeled Atlas isn’t a product. It’s a research testbed,” Boston Dynamics wrote in the post, which was authored by Xiaoyu (Andy) Wu, a controls engineer on the Atlas team.

Why wheels make a good torture test

Wu’s job is to make Atlas’ motors and actuators behave predictably. He wanted a setup that forced them to deliver high torque and high speed continuously while carrying the robot’s weight. Rolling does that well: every launch, turn and stop sweeps the motors across their full range.

The hardware change was small. One custom adapter fits an ordinary scooter rim onto the existing ankle actuator, and that adapter is the only new part. Wu compared the retrofit to swapping a wheel on a bicycle and credited the ease to Atlas being designed for high-volume production and easy servicing. On the software side, each wheel is treated as an ankle that can spin forever, so the team kept the standard robot model and simply changed the feet.

Simulation has to be honest

The behaviors were learned with reinforcement learning, which means the controller practiced across millions of simulated runs before it touched the real robot. That only works if the simulated motors act like the real ones. If the simulation is too generous, the policy learns moves the hardware cannot actually perform. Wu describes a lot of unglamorous work, such as matching low-level control settings across every software setup, running characterization tests and tuning the motor model to what the team measured on Atlas. Boston Dynamics also thanked the MuJoCo team at Google DeepMind for the open-source physics engine the work ran on.

The payoff, by Wu’s account, was a near-perfect transfer from simulation to hardware. The spinning jump follows a motion-capture recording of a person and worked on the first try. Wu said he picked human clips with lots of foot sliding on purpose, and in the side dashes the policy figured out on its own to steer and roll where a person would slide. “Give the policy wheels, and it figures out how to use them,” Wu wrote. He added that this was his first real project in reinforcement learning, coming from a classical motor-control background.

Where this fits for Atlas

Boston Dynamics says the lessons are already going back into the standard, legged Atlas. Wu’s next project is how the controller should react when an actuator fails, which matters a lot more on a factory floor than in a demo. Atlas is being trained on real car parts at the Robotics Metaplant Application Center at Hyundai’s Georgia plant, and it is getting a new four-fingered gripper. Hardware that behaves the same way in simulation and on the floor is what lets those factory skills get trained in software first.

It is also not the first time the company has played with wheels. Nearly a decade ago it showed Handle, a robot that balanced on two wheels.

The wheels-versus-legs argument

The demo arrives in the middle of a real industry debate. Agility Robotics has said it is exploring wheeled designs alongside Digit, though Jonathan Hurst said Digit itself likely won’t get wheels. AGIBOT already sells a wheeled industrial model, the G2 Max, which it showed off when it pitched a new way to score robot productivity. And on the same day as the Atlas video, Swiss startup Embodied AI, a spinout from EPFL, unveiled AIR, a two-armed robot on a wheeled base that it says reached a 95 percent average task success rate in paid factory pilots, without saying how that figure was measured.

Boston Dynamics is not joining that camp. It is using wheels to learn about the motors it already has. But the video does show how cheaply a well-designed humanoid can be reconfigured, and that flexibility could matter more over time than any one body shape.

Our take

The spin is the headline, but the message for anyone building robots is the sim-to-real work. A learned controller is only as trustworthy as the simulation it trained in, and Boston Dynamics is treating motor accuracy as the foundation. You can follow the company on its Boston Dynamics page, compare Atlas with other machines on our humanoids page, and find controls and reinforcement-learning roles on our jobs board.