Boston Dynamics bolts scooter wheels onto Atlas to test how well its simulator knows the motors

Boston Dynamics has put wheels on its Atlas humanoid and handed the driving to a learned controller. A single custom adapter mounts an off-the-shelf scooter rim onto Atlas's existing ankle actuator, and it is the only new part in the build, controls engineer Xiaoyu (Andy) Wu wrote in a company blog post on Nov. 9. In software, each wheel is modeled as an ankle that can spin continuously. "A wheeled Atlas isn't a product. It's a research testbed," the company says.

The wheels are the least important part of the experiment. Wu picked driving because every launch, turn and stop pushes the motors across their full torque-speed range while they carry real load, and the question was whether a policy trained in simulation still holds up when the real actuators run at the edge of their envelope. What Boston Dynamics is reporting is a motor model good enough to train against. That model belongs to the actuators it was measured on. Hyundai Mobis, which has won the order for Atlas's actuators, starts building prototypes at its Uiwang research center in November and expects to supply them in 2028, ZDNet Korea reported (our report).

Most of the work went into making the simulated motors match the real ones

Whole-body control trained with reinforcement learning runs through millions of simulations before it touches hardware, and the policy learns whatever the simulator lets it get away with. "If the simulation used for training assumes the motors are over-idealized, the policy can end up learning things the real hardware can't actually do," Wu writes. Closing that sim-to-real gap meant characterization experiments, checking that low-level control parameters matched across every software setup, and tuning the simulated motor model to what the team measured on the robot.

The learning side needed less. A few changes to the reward, the range of commands the policy trains on and how it holds its posture were enough for smooth rolling to emerge, the post says. Wu fed it human motion-capture clips with a lot of foot sliding. In side dashes, where a person would slide their feet, the policy learned to steer the wheels and roll through. A 360-degree jump, tracked from a recording of a person, worked on the first try.

The lessons go back to the legged robot

Boston Dynamics says the insights are "already being adapted back to our standard configuration," and Wu's next project is how the policy responds to failed actuators. Elsewhere the industry is split: many companies are developing wheeled humanoids while others stay committed to legs, The Robot Report notes, and Agility Robotics co-founder Jonathan Hurst said Digit likely won't get wheels, though the company is exploring other wheeled configurations.

The open question is how much of the motor model survives the move to Mobis's parts. Mobis says it will apply standardization and modularization to the parts and keep improving their performance, according to ZDNet Korea. If its actuators behave like today's at the limits, the model carries over; if they don't, this is the work that has to be done again. By Wu's account, the only new hardware was a custom adapter for an off-the-shelf scooter rim. The expensive part was knowing the motors.