Boston Dynamics Swapped Atlas's Feet for Scooter Wheels

By James Calloway

On November 9, 2026, Boston Dynamics detailed a wheeled Atlas experiment in which the humanoid's feet were replaced with wheels — a research testbed, not a product, built to probe actuator limits and sim-to-real transfer. The retrofit behind it was strikingly small: a single custom adapter joining an off-the-shelf scooter rim to Atlas's existing ankle actuator. That contrast, ambitious whole-body learning mounted on minimal hardware changes, frames the whole project.

The stakes are real for the robotics industry. Deployed robots encounter conditions a lab cannot fully reproduce, Boston Dynamics said, citing years of customer experience with Spot (the quadruped) and Stretch (the warehouse case-handling robot). The wheeled configuration let the company test how real actuators and a learned controller behave when motors approach their operating limits — accelerating, turning and stopping keep motors under load while demanding high torque and high speed simultaneously. No marketing claim of a shipping product accompanies any of it; the company explicitly labels the machine a testbed.

Who ran it, and why that matters

Atlas controls engineer Xiaoyu (Andy) Wu wrote the account. His background is in classical motor control, and this was his first substantial reinforcement learning (RL) project — training a policy through trial and error in simulation rather than hand-coding control laws. That placed him at the boundary between three specialties usually kept separate on Atlas: actuator design, low-level motor control, and whole-body RL.

Wu used the testbed to examine extreme actuation behavior and stress-test the robustness of his low-level control tuning. That framing deserves attention. Much humanoid research publishes polished demonstrations without disclosing the tuning fragility underneath; this account centers on it.

What does 'sim-to-real' mean here?

Training consumed millions of robot simulations before the policy touched hardware. The core problem, known as sim-to-real transfer, is making a simulated motor reproduce measured physical behavior closely enough that a learned policy still works on the real robot.

The failure mode is concrete. Policies trained against an idealized motor may rehearse movements the real robot cannot physically perform. Hardware-matched models, by contrast, support only behavior the robot can reproduce.

Preparation involved three steps:

• Matching low-level control parameters across software setups
• Running motor-characterization experiments on the physical robot
• Tuning the simulated motor against measurements taken from Atlas

The team also evaluated modeling simplifications individually — keeping harmless approximations while rejecting any that would hand the policy an unrealistic advantage during training. That discipline is often where sim-to-real projects quietly fail, and Boston Dynamics is at least naming the practice.

What did the wheeled Atlas actually do?

The demonstrations, while eye-catching, carried the research payload.

For side-dash training, Wu selected human motion-capture clips that contained substantial foot sliding. The policy then converted those human foot slides into steered, rolling movements as it learned to exploit the wheels. The lesson transfers: motion data that would count as noise or error for a walking robot became useful steering signal here.

The 360-degree jump tracked a motion trajectory recorded from a human jumper. Wu characterized the move as relatively low-risk despite its energetic appearance and said it succeeded on the first attempt. First-attempt success on hardware is rare enough in robotics that the caveat matters — one move, one robot, one trial, as reported by the engineer who built it.

The most technically revealing detail came before takeoff. As the robot prepared to jump, its wheeled lower body rolled quickly into position ahead of the upward push — a use of traction unavailable to Atlas on feet.

How much hardware did it take?

Almost none, and that is the point. In the baseline software model, each wheel became an ankle capable of continuous rotation. That allowed Wu to substitute wheels for feet without constructing a separate robot model. The only added component was the custom ankle-to-scooter-rim adapter.

The researchers adapted established policy frameworks rather than inventing new ones, changing the reward composition, the range of training commands, and posture handling. Wu said those changes produced smooth rolling while Atlas remained stable and easy to drive, and he characterized the resulting transfer to hardware as essentially perfect. Readers should weigh that phrase against its scope: it describes this testbed, under these tested conditions, not Atlas generally.

What does this mean for the standard, bipedal Atlas?

Boston Dynamics said lessons from the experiment are already being adapted to Atlas's standard configuration. The goal is closer integration between actuator dynamics and whole-body RL, intended to improve policy robustness against failure modes and expand the robot's capabilities. The demonstrations therefore connect an unusual research configuration directly to the control system used by the conventional bipedal machine.

A second workstream has begun. Wu has started building policies for failed actuators, including whether RL can generate emergent responses to those failures. The follow-on extends the reliability work to hardware whose behavior changes after a failure — testing whether learned control can still respond when a real actuator departs from its expected behavior.

That question is the quiet headline. A humanoid that tolerates actuator degradation, rather than falling over when a motor misbehaves, is a precondition for any credible commercial deployment. Boston Dynamics has not claimed Atlas achieves this; it has only started asking.

The takeaway: a scooter wheel, a single adapter, and millions of simulations have given Boston Dynamics a cheap instrument for measuring where its reinforcement learning breaks — and the company is already porting the findings back to the robot on feet.