Boston Dynamics puts Atlas on wheels to study model-based control plus RL
A Boston Dynamics video shows the Atlas humanoid with scooter wheels on its ankles, trained in simulation with a whole-body RL policy layered on classical control. The company calls it a research testbed, not a product.
Boston Dynamics released a video on 9 November 2026 in which its Atlas humanoid rolls on wheels fitted in place of its feet. The Robot Report’s account does not give Atlas’s degrees of freedom, payload or reach. The work is by Andy Wu, a controls engineer on the Atlas team. The company describes the wheeled robot as a research testbed and says it is not a product.
The hardware change is small, according to Wu. One custom adapter mounts a bought-in scooter rim on the ankle actuator Atlas already has, and that adapter is the only new part. In the robot model, each wheel is treated as an ankle joint that can spin without limit, so the feet were swapped for wheels in the baseline model and little else changed.
The controller combines classical model-based control for baseline stability with a whole-body reinforcement learning (RL) policy. The policy takes proprioceptive feedback and velocity commands and adjusts pose and contact forces to cope with disturbances the model does not capture. Wu trained all behaviours in simulation and ran them on the real robot. The video shows fast rolling on flat ground, legged moves over obstacles and pirouettes. Wu’s next question is how the policy responds when an actuator fails.
Why it matters
The useful detail is the split of work between controllers: a model-based layer that keeps the robot upright and a learned layer that absorbs what the model gets wrong. The Robot Report gives the team’s conclusion as better agility from the combination than from either method alone, but the evidence offered is a video, not published measurements. Treating a wheel as a continuously rotating joint is a cheap modelling choice to copy when you prototype wheel-leg hybrids in simulation. Note that this is a demo of a research configuration, not a shipping option.