Boston Dynamics Puts Atlas on Scooter Wheels for RL Testing

Boston Dynamics mounted an off-the-shelf scooter rim on each of Atlas's ankle actuators using a single custom adapter — the only new hardware in the experiment — and the resulting wheeled humanoid rolled, drifted and landed a 360-degree jump on its first attempt. The company frames the retrofit not as a product but as a testbed for how far a learned whole-body controller can push real actuators.

"A wheeled Atlas isn't a product. It's a research testbed," Atlas controls engineer Xiaoyu (Andy) Wu wrote in a company blog post published alongside a video of the robot in action.

How small was the hardware change?

Small enough to fit in one sentence. One custom adapter attaches a scooter rim, bought off the shelf, onto the ankle actuator Atlas already carries. That adapter was the only new part in the entire build, according to the company's post.

In software, the modifications ran deeper. The team modeled each wheel as an ankle that can spin continuously, then replaced the feet in the baseline robot model with wheels. Wu compared the retrofit to swapping a bicycle wheel, and credited Atlas's design for high-volume production and easy servicing for making the swap that straightforward.

That production-minded architecture is the quiet point of the demo. A humanoid whose feet come off like bicycle wheels is a humanoid a customer can service without an engineer on site. The wheels are a stunt; the serviceability is the strategy.

How does the controller work?

Control is split into two layers. Classical model-based motor control provides the baseline stability the robot needs to stay upright at all. The team characterized the motors and tuned the simulated motor model until it matched what they measured on the actual hardware — a painstaking but conventional step.

The interesting layer sits on top. A whole-body reinforcement learning policy takes proprioceptive feedback plus velocity commands, and outputs pose and force adjustments that absorb disturbances and imperfect wheel contact. The team trained the policy in simulation using Google DeepMind's MuJoCo physics engine, adapting it from Atlas's existing RL frameworks with changes to rewards, command ranges and posture.

Then they put it on real hardware. Atlas rolled, launched, turned and stopped. It slid through sideways dashes. And it performed a 360-degree jump based on a motion-capture recording of a human — a maneuver that "just worked first try," according to the post.

Sliding and drifting were never programmed. The behaviors emerged from training. "Give the policy wheels, and it figures out how to use them," Boston Dynamics wrote.

Why wheels on a humanoid at all?

Because a wheel is a stress test for everything else. A spinning rim multiplies the forces coursing through the ankle actuator, the contact dynamics stray far from what a flat foot produces, and the controller has to keep balancing through disturbances its training only approximated. If a learned policy can handle that gap between simulation and reality, it can likely handle the messier failures that matter commercially.

The next target makes that explicit. Wu wants the policy to respond correctly when an actuator fails, which he sees as central to safety. A robot that degrades gracefully when a motor dies is a robot you can put near people. The team also plans to carry the findings back into Atlas's standard legged configuration, so the wheeled detour should pay dividends on feet.

What comes next?

The Robot Report reports that Boston Dynamics is preparing to expand Atlas testing at Hyundai facilities — a signal that the program is moving from lab demos toward deployment-scale evaluation under the ownership of its Korean parent, which bought the company in 2021.

For now, the company's own framing holds: no wheeled Atlas will ship. But the experiment shows a whole-body RL policy transferring from MuJoCo simulation to real actuators with zero hardware redesign, landing a 360-degree jump on the first try. As the team pushes the same policy toward actuator-failure recovery and ported improvements for the legged robot, the gap between research testbed and field-ready humanoid keeps narrowing.