A scooter rim bolted to a humanoid ankle is not a look anyone expects from Boston Dynamics. Yet that is exactly what appears in Atlas On a Roll, a video and blog post the company published on November 9, 2026, showing its electric Atlas robot rolling, spinning and drifting across a warehouse floor on wheels instead of feet.
The modification itself is almost comically simple. A single custom adapter mounts an off-the-shelf scooter rim onto Atlas's existing ankle actuator, and that adapter is the only new hardware in the entire build. In software, each wheel is treated as a continuously spinning ankle joint, letting the same whole-body controller Atlas uses for walking take over rolling locomotion with no separate driving module.
Why the Wheels Were Never the Point
Boston Dynamics controls engineer Andy Wu was blunt about the motivation: "The wheels weren't really the point of this. The point was pushing the actuation to its limits." The team wanted to know whether a controller trained almost entirely in simulation would still behave correctly when real motors were driven into continuous, high-torque, high-speed operation, a regime that is hard to study safely with legs alone.
Getting there required closing the usual sim-to-real gap by matching low-level motor control parameters across software setups and tuning simulated motor behavior against real measurements from the robot. Engineers then made targeted adjustments to reward composition, command ranges and posture-holding behavior so the policy, trained across millions of simulated runs, would transfer faithfully to hardware.
The payoff showed up in the demonstrated maneuvers: smooth rolling on flat terrain, steered side dashes, backward skating and drifting, and a 360-degree jump built from human motion-capture data that the robot nailed on its first attempt. Atlas also transitioned between fast rolling and articulated legged motion to cross obstacles, switching modes without a hardwired script for when to do which.
A Research Testbed, Not a Product Roadmap
Boston Dynamics was explicit that wheeled Atlas has no commercial future. It exists purely to stress-test the boundary between classical model-based motor control and reinforcement-learning policies, confirming that the hybrid approach holds up even when actuators are running at the edge of their performance envelope rather than in the comfortable middle.
That distinction matters for how rivals should read the news. Humanoid programs at Tesla, Figure and Unitree have leaned heavily on learned locomotion policies to cut development time, but validating those policies against hardware limits usually means risking expensive falls. Boston Dynamics effectively found a lower-stakes way to probe the same question, and the company says the insights are already being fed back into standard, legged Atlas configurations to expand its normal capability envelope and improve policy robustness.
No production plans, timeline or customer applications were attached to the wheel experiment, according to Boston Dynamics. The company frames it as one data point in an ongoing effort to make whole-body control transferable across very different hardware configurations, not a preview of a wheeled product line.