Boston Dynamics Swaps Atlas Feet for Wheels to Push Its Electric Motors to Their Limits

Boston Dynamics has published a blog post describing an internal experiment in which controls engineer Xiaoyu (Andy) Wu replaced the feet on an Atlas humanoid with scooter wheels and trained a whole-body reinforcement learning controller to operate the resulting machine — entirely in simulation, with no additional real-world fine-tuning after the policy was deployed to hardware.

The wheeled Atlas is not a product. Boston Dynamics is explicit about this. The machine was built to do something the standard bipedal robot cannot: keep Atlas's electric motors running at high torque and high speed simultaneously, for extended periods, to expose where the company's motor models and control tuning diverge from reality. The findings are already being applied to the production Atlas now being tested at Hyundai facilities.

One Custom Adapter, Off-the-Shelf Hardware

The hardware change was deliberately minimal. Wu attached an off-the-shelf scooter rim to each of Atlas's existing ankle actuators using a single custom adapter — the only new part in the build. Boston Dynamics described the modification as roughly as involved as changing a bicycle tire.

On the software side, each wheel was modeled as an ankle joint capable of continuous rotation, with the rest of the robot's existing model left unchanged. Changes to the reward composition, the range of training commands, and how the policy maintained posture were sufficient for smooth rolling behavior to emerge from training. No special wheel-specific controller was developed from scratch.

This minimalism was intentional. By keeping the hardware close to stock Atlas, the experiment isolates one variable: what happens to the learned controller when the end-effector properties change from foot to wheel, and when the contact dynamics shift accordingly.

Why Wheels Stress-Test Actuators in Ways Legs Cannot

Electric motors face a fundamental engineering tradeoff between torque and speed: a motor spinning fast delivers less torque than one turning slowly, and operating at the intersection of high torque and high speed is precisely where motors overheat, saturate, or diverge from the smooth behavior their models predict.

Bipedal walking keeps motors largely away from that intersection. Footstrike phases load the ankle joints heavily but briefly, and the swing phase offers recovery. Rolling locomotion has no such relief. Wheels require sustained simultaneous torque and speed — during acceleration, turning, and braking — placing Atlas's actuators under the kind of continuous combined loading that decades of Spot and Stretch deployments in customer facilities have shown that real environments eventually produce, even when lab conditions don't.

Wu characterized the resulting sim-to-real transfer as essentially perfect. The fidelity came from motor characterization work done on the hardware before training: the team ran measurements on the real actuators and tuned the simulated motor model to match them, so the policy trained against a simulation that reflected what the physical motors actually do at their limits, not an idealized proxy.

Training in Simulation, Jumping on the First Try

Training used MuJoCo, the physics engine widely used across academic and industrial robot learning research for its computational efficiency and accurate contact modeling. Some motion sources came from human motion-capture clips — specifically chosen to include heavy foot-sliding movements, which the team expected to produce interesting behavior when translated to wheels. A 360-degree jumping maneuver was derived from tracking a human jumper through that trajectory pipeline. It succeeded on the first attempt when transferred to the wheeled hardware.

The policy also developed behaviors that have no direct human analogue. During side dashes where a person would slide their feet, the wheeled Atlas learned to actively roll the wheels, using traction dynamically rather than passively. During the windup before the jump, the lower body rolled into position before the upward push — a movement coordination pattern that emerged from training rather than being scripted. These emergent behaviors are a signal that the whole-body RL approach is finding solutions the designers did not anticipate, which is both an engineering success and a reason to continue studying what the policy is actually doing.

A Crowded Field, But a Different Goal

The wheeled humanoid landscape has been filling rapidly. Hexagon's AEON, built entirely around wheel-feet from the start, has been deployed at BMW's Leipzig plant and in Schaeffler and Pilatus facilities, reaching rolling speeds of 2.4 meters per second on flat industrial floors. Unitree demonstrated its G1 humanoid on roller skates and ice skates in April 2026, and had earlier released the G1-D, a wheeled variant designed for data collection and AI training.

Boston Dynamics is doing something categorically different. Where Hexagon committed to wheels as the locomotion solution for factory and warehouse floors, and where Unitree is exploring wheeled mobility as a product direction, Boston Dynamics is using wheels as a diagnostic instrument. The goal is not to build a wheeled robot for deployment but to understand, under controlled stress, where the standard legged robot's control stack has remaining weaknesses before it encounters those conditions in a Hyundai assembly plant.

Fault Tolerance Is the Next Problem

Wu wrote in the post that he is now studying how the whole-body RL policy responds to failed actuators — specifically whether training can produce emergent recovery behaviors that allow the robot to continue operating when a motor fails mid-task. The question has clear production relevance: an Atlas running in a Hyundai facility cannot stop and await repair every time a motor behaves unexpectedly.

Boston Dynamics is currently testing production Atlas units at Hyundai facilities, following the January 2026 CES unveiling of the production-ready electric Atlas. The company installed Rohit Prasad — who led AI development at Amazon — as its CEO in October 2026, signaling a shift toward deploying AI-driven autonomy alongside the ongoing hardware program.

The wheeled experiment sits inside that transition: a company turning its research history into production reliability, using deliberately controlled stress to close the gap between what the simulator predicts and what a real factory floor eventually demands. Whether the fault-tolerance research succeeds will matter more for the actual Atlas deployment at Hyundai than any single locomotion demo.