Boston Dynamics Swaps Atlas Feet for Wheels to Stress-Test Actuators

Boston Dynamics replaced its humanoid Atlas robot's feet with wheels on November 9, 2026, publishing results of an experiment designed to push the robot's actuators to their operating limits. The company calls the wheeled configuration a research testbed, not a product, and says lessons are already flowing back into Atlas's standard bipedal design.

The experiment set out to answer a narrow engineering question: how do real actuators and a learned whole-body controller behave together when motors approach their limits? Accelerating, turning and stopping keep the motors under sustained load while demanding both high torque and high speed. Boston Dynamics said years of customer experience with its Spot and Stretch robots have shown that deployed machines encounter conditions a lab cannot fully reproduce.

Who ran it, and why it matters

Atlas controls engineer Xiaoyu (Andy) Wu, who wrote the company's account, used the testbed to examine extreme actuation behavior and to stress-test the robustness of his low-level control tuning. The project joined three specialties usually kept separate on Atlas: actuator design, low-level motor control and whole-body reinforcement learning. Wu's background is in classical motor control, and this was his first substantial reinforcement learning project — placing him, and the experiment, at the boundaries between those layers.

What did the training involve?

The policy was trained in millions of robot simulations before it ever touched hardware. The core challenge is known in the field as sim-to-real transfer: the modeled motor had to reproduce measured physical behavior closely enough for the learned policy to survive contact with the real machine. Policies trained against an idealized motor may rehearse movements the real robot cannot perform; hardware-matched models instead support behavior the robot can actually reproduce.

Preparation involved three steps:

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

The team also evaluated simplifications one by one, keeping harmless approximations while discarding any that would hand the policy an unrealistic advantage during training. The takeaway: the fidelity of the motor model, not raw training volume, decides whether simulation skills survive on hardware.

How well did it transfer?

Wu adapted established policy frameworks, changing the reward composition, the range of training commands and posture handling. He 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." That is a strong claim from a first-time RL practitioner, and the company offers no third-party benchmark to confirm it — the demonstrations are the evidence.

Two demonstrations stand out. For side-dash training, Wu selected human motion-capture clips featuring substantial foot sliding; the policy converted those human foot slides into steered rolling movements as it learned to use the wheels. For the 360-degree jump, the controller tracked a trajectory recorded from a human jumper. Wu called the move relatively low-risk despite its energetic appearance and said it succeeded on the first attempt. Notably, the wheeled lower body rolled quickly into position before takeoff — a use of traction unavailable to Atlas on feet.

The hardware change itself was minimal. The only added component was a custom adapter joining an off-the-shelf scooter rim to Atlas's existing ankle actuator. In the baseline software model, each wheel became an ankle capable of continuous rotation, so Wu could substitute wheels for feet without building a separate robot model. In plain terms: a scooter part unlocked the whole experiment.

What happens next?

Boston Dynamics said the lessons are already being adapted to Atlas's standard configuration, with closer integration between actuator dynamics and whole-body reinforcement learning intended to improve policy robustness against failure modes and expand the robot's capabilities. Wu has also started work on policies for failed actuators, including whether reinforcement learning can generate emergent responses to those failures. The follow-on tests whether learned control can adapt when a real actuator departs from its expected behavior — the reliability question that motivated the wheeled detour in the first place.