Hyundai Motor Group's robotics subsidiary Boston Dynamics disclosed that its wheeled Atlas humanoid completed a 360-degree spinning jump on its first physical attempt after training entirely in simulation, with the company saying it precisely modeled motor operating characteristics in a virtual environment and transferred the learned motions to the real robot11,13. The company released the demonstration video on November 9 alongside a blog post by Atlas controls engineer Xiaoyu (Andy) Wu1,16.
The hardware modification was minimal: a single custom adapter mounts an off-the-shelf scooter rim onto each of Atlas's existing ankle actuators, and that adapter is the only new part in the build17. Wu compared the retrofit to swapping a bicycle wheel, attributing the ease to Atlas being designed for high-volume production and easy servicing. The scooter wheels attach to the robot's existing ankle drive units10.
Sim-to-real pipeline
In software, the team modeled each wheel as an ankle that can spin continuously and trained the same whole-body controller used on the walking version. The approach combined classical model-based motor control with whole-body reinforcement learning trained in simulation18. Boston Dynamics simulated millions of floor conditions and ran parameter checks and characterization tests to verify that the policy would not demand moves real actuators could not handle. The policy was then transferred to the physical robot with only minor changes to rewards, command ranges, and posture.
The results included rolling locomotion, rapid side dashes, and the 360-degree spinning jump. The video also showed Atlas balancing on a single wheel with the opposite leg raised, and pushing itself up from the ground before accelerating forward4. Wu said he deliberately sought motion-capture clips with heavy foot sliding because sliding is where wheels become interesting15. He added that he was surprised training figured out how to use the wheels to roll during the side dash: "I was pleasantly surprised that training did in fact figure out how to use the wheels to roll during the side dash," Wu said7.
The continuous load from supporting the robot's own weight while rolling pushed actuators to their limits, providing data on how well simulated motor behavior matches reality. A Boston Dynamics official said wheeled locomotion is suitable for testing motor performance limits because the motors must bear real loads during start, turn, and stop while consistently delivering high torque and speed12. Through the RL policy, Atlas processed real-time proprioceptive feedback and velocity commands, outputting dynamic pose and force adjustments to handle unmodeled disturbances.
Research testbed, not a product
Boston Dynamics stated that a wheeled Atlas is not a product but a research testbed. Wu's next research question concerns how the policy responds to failed actuators. Improvements from the wheeled experiments are being applied to the walking Atlas14. Hyundai Mobis will begin producing prototype actuators for Atlas this month, while Hyundai Motor Group is targeting deployment at its HMGMA plant in the United States in 2028 and an expanded role in assembly work by 2030. Hyundai's parent group recently deepened its humanoid-robot supply commitments alongside Samsung Electronics at a 170-executive summit scheduled for November 9[2].
ANALYSIS The experiment's value lies less in the wheeled form factor itself than in what it reveals about sim-to-real transfer fidelity: wheels impose continuous high-torque loads that expose gaps between simulated and physical motor models more aggressively than intermittent foot contacts do. That Boston Dynamics plans to carry findings back into Atlas's standard legged configuration positions the wheeled testbed as an accelerated validation loop for the walking robot's controller stack rather than a product fork.