Boston Dynamics Puts Scooter Wheels on Atlas and Publishes No Speed Figure

Boston Dynamics has swapped the feet of its Atlas humanoid for wheels and says the result "isn't a product. It's a research testbed," according to The Robot Report, which published on November 9, 2026. The company's own post, "Atlas On a Roll", is written by Xiaoyu (Andy) Wu, a controls engineer on the Atlas team, and reports no speed, payload or runtime figure.

One adapter, one off-the-shelf scooter rim

The hardware change is small. A single custom adapter mounts an off-the-shelf scooter rim onto Atlas's existing ankle actuator, and Wu's post says it is the only new part. In software, each wheel is modelled as an "ankle" that can spin continuously, while the rest of the robot model is reused.

The control approach combines classical model-based motor control with whole-body reinforcement learning. The post says the policies were trained in simulation over millions of runs before reaching hardware, using Google DeepMind's open-source MuJoCo physics engine, and that the team adapted existing policy frameworks instead of starting over. Changes were made to the reward composition, the range of commands the policy trains on and how it holds posture. The simulated motor models were tuned against measurements from the real robot.

Behaviour	How it was produced
360-degree jump	Trajectory-tracking policy trained on a human motion-capture recording; worked on the first try, per Wu
Side dashes	Emerged from the policy rolling the wheels rather than sliding as feet would
Rolling and legged switching	Stated goal per The Robot Report: smooth changes between fast flat-ground rolling and legged motion over obstacles

What the post does not contain

There is no wheel diameter, top speed, battery life or failure rate in either source. The Robot Report states that no dimensions, weight or speed figures were given. Wu's post says the purpose was to push the actuators through their full torque-speed range during launches, turns and stops, and to measure how faithfully a learned whole-body controller works against real hardware.

The caveats are in the post itself. Wu writes that a normal robot with feet could have done comparable behaviour, and that simulation approximations risk a policy learning moves the hardware cannot perform. The next study is how the policy responds to failed actuators, which The Robot Report calls key to safety.

The Robot Report also quotes Jonathan Hurst, co-founder and Chief Robot Officer at Agility Robotics, who said Digit likely will not get wheels, though Agility is exploring other wheeled configurations. Digit is the machine behind Agility's $465 million public listing.

Why a wheeled humanoid matters to the commercial case

The number that matters here is not the jump but the energy and wear cost of walking. Legs spend motor torque on every step; a wheel on flat factory floor does not. A warehouse buyer comparing a legged humanoid with a wheeled mobile base is comparing a general-purpose body against a cheaper, faster one, and this experiment is Boston Dynamics testing how much of the legged robot's control stack carries over when the legs stop stepping. The post does not make that argument and gives no cost data, so it is an inference from the design, not a Boston Dynamics claim.

Boston Dynamics is preparing to ramp up Atlas testing at Hyundai facilities, according to The Robot Report, and is developing a new four-fingered gripper. Another machine that has to prove itself outside a demonstration is covered in our report on Axon's $1.2 billion Skydio distribution deal.

What would change this read: a published speed, runtime and torque-limit figure for the wheeled configuration, or a Boston Dynamics statement that wheeled variants will be offered to customers. Until then it is a research result with no deployment count.