CES® 2027 NEWS
ROAD TO CES 2027Live now
ROAD TO CES · AN INTELITOYS EVENTJAN 6–9 · VEGAS
Home Robots

Boston Dynamics Bolts Wheels Onto Atlas to Stress-Test Its Whole-Body AI Controller

Announced the wheeled Atlas is a research testbed built to push the robot's actuators and reinforcement-learning stack to their limits.. Boston Dynamics replaced Atlas's feet with scooter wheels to find out how faithfully a learned whole-body controller works with real actuators at the edge of their performance envelopes.

InteliToys Editors ? Boston Dynamics Atlas wheeled research testbed
Atlas humanoid robot on wheels (image: Boston Dynamics)

Boston Dynamics has shown off an Atlas humanoid that rolls rather than walks. A post on the company's blog walks through the experiment, which the team framed as a way to find out whether a learned whole-body controller can hold up when the actuators underneath it are driven near their limits. The wheeled machine is not something Boston Dynamics plans to sell; it is a testbed, and the company is candid that the wheels were a means to an end rather than the goal.

The hardware change was minimal by design. One bespoke adapter lets a standard scooter rim bolt straight onto Atlas's ankle actuator, and Boston Dynamics says that piece is the only new hardware in the entire build. The company attributes the easy swap to choices made when Atlas was first engineered, when it was designed to be manufactured in volume and serviced in the field. In the post, the retrofit is compared to changing a bicycle wheel.

Reinforcement learning sits at the center of the experiment. The blog post is authored by Xiaoyu (Andy) Wu, a controls engineer on the Atlas team, who writes that the project marked his first serious engagement with RL. He describes the unglamorous groundwork: verifying that low-level control settings lined up across each software configuration, running tests to characterize the motors, and adjusting the simulated motor model until it reflected what the team had measured on the physical robot.

Ride the Las Vegas Monorail to CES - stops at the Las Vegas Convention Center

That fidelity matters because a policy trained in simulation can learn tricks the real hardware cannot reproduce. Boston Dynamics says the team altered the reward structure, broadened the commands the policy sees during training, and changed how the robot is asked to hold itself, after which smooth rolling appeared without being scripted. The robot remained stable and straightforward to drive, according to the company.

Wu explains that wheels create a demanding workload for the motors, forcing them to supply both high torque and high speed while supporting the robot's mass. Each start, turn and stop sweeps the actuators across their entire torque-speed range, which is precisely the stress the hardware was built for. The company says the team drew on recorded human motion for some of the maneuvers, favoring clips with heavy foot sliding. A 360-degree jump came from a trajectory-tracking policy following a human recording and succeeded immediately, while the side dashes taught the policy to steer the wheels through slides a person would perform with their feet.

Boston Dynamics says the lessons are already flowing back into the standard Atlas configuration. The company frames the work as part of a longer effort to keep actuator dynamics and whole-body learning tightly coupled, and it says the team has moved on to studying how learned policies might react when an actuator fails.

Source: Boston Dynamics.

InteliToys is editorial. We do not sell Boston Dynamics Atlas wheeled research testbed. Watch the source tape in the branded lightbox below.