Boston Dynamics Turns Atlas Into a Wheeled Actuator Testbed

Boston Dynamics detailed a wheeled Atlas experiment on November 9, 2026, after replacing the humanoid’s feet with wheels to probe actuator limits and the transfer of whole-body reinforcement learning from simulation to hardware.

The company describes the configuration as a research testbed, not a product. The experiment was designed to expose how real actuators and a learned whole-body controller behave together when the motors approach their operating limits. Accelerating, turning and stopping keep the motors under load while demanding high torque and high speed across their operating range.

Boston Dynamics said years of Spot and Stretch customer experience have shown that deployed robots encounter conditions a lab cannot fully reproduce. Atlas controls engineer Xiaoyu (Andy) Wu, who wrote the account, used the testbed to examine extreme actuation behavior and the robustness of his low-level control tuning.

Motor Models for Sim-to-Real Training

Training used millions of robot simulations before the policy reached hardware. The simulation-to-hardware problem, called sim-to-real, depended on making the modeled motor reproduce measured physical behavior closely enough for the learned policy to transfer. Policies trained against an idealized motor may rehearse movements the real robot cannot perform; hardware-matched models instead support behavior the robot can reproduce.

Preparation involved matching low-level control parameters across software setups, running motor-characterization experiments and tuning the simulated motor against measurements from Atlas. The team also evaluated simplifications individually, retaining harmless approximations while avoiding those that would give the policy an unrealistic advantage during training.

The project joined actuator design, low-level motor control and whole-body reinforcement learning (RL), which are usually separate specialties on Atlas. Wu’s background is in classical motor control, and this was his first substantial RL project, placing him at the boundaries among those layers.

The researchers adapted established policy frameworks, 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 characterized the resulting transfer to hardware as essentially perfect.

Rolling Demonstrations and a Minimal Retrofit

Some demonstrations began with human motion-capture clips that Wu selected for substantial foot sliding. The side-dash training converted those human foot slides into steered, rolling movements as the policy learned to use the wheels.

For the 360-degree jump, the controller 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. As the robot prepared for takeoff, its wheeled lower body rolled quickly into position before the upward push, showing a different use of traction from Atlas on feet.

The hardware behind those movements required little alteration. 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, allowing Wu to substitute wheels for feet without constructing a separate robot model.

Lessons for Standard Atlas

Boston Dynamics said lessons from the experiment are already being adapted to Atlas’s standard configuration. Closer integration between actuator dynamics and whole-body RL is 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.

Wu has also started work on policies for failed actuators, including whether reinforcement learning can generate emergent responses to those failures. The follow-on extends that reliability work to hardware whose behavior changes after a failure, testing whether learned control can respond when a real actuator departs from its expected behavior.