Boston Dynamics put scooter wheels on Atlas — one adapter, sim training, and it drifts
heyshrutimishra · x · 2026-11-10
Boston Dynamics swapped Atlas's feet for scooter wheels and the humanoid taught itself to drive, drift and do 360 jumps.
- Hardware change was minimal: the only new part is a scooter rim bolted to the ankle actuator via a custom adapter — roughly as involved as swapping a bicycle wheel.
- Software route: each wheel is modeled as an ankle that can spin forever; the policy was trained in MuJoCo over millions of runs and transferred to the real robot with no hand programming.
- Side-dash moves intentionally used mocap clips with heavy foot sliding — the policy figured out to steer the wheels; the 360 jump came from a human mocap clip and worked on the first try.
- It's a research testbed, not a product, aimed at pushing actuators to full torque/speed and validating sim fidelity. Controls engineer Andy Wu, from a classical motor-control background, did this as his first RL project.
Related event: Boston Dynamics Puts Atlas on Wheels for Radical RL Stress Test(3 posts)→
More from Embodied
- Tesla's unsupervised Robotaxi hits 8 cities in a year; poster still bets on 50k by 2027 — mitchdeg · 2026-11-10
- AgenticGenTAMP: coding agents synthesize reusable TAMP policies across 28 simulated environments — tomssilver · 2026-11-10
- Skild AI open-sources Steel physics solver with NVIDIA: 13× faster than PhysX, ~3× less GPU memory — deepakpathak · 2026-11-10
- Georgia Tech Lunar Lab brings 3 robot papers to CoRL 2026: deformable dynamics, navigation, humanoids — fdellaert · 2026-11-10
- DeepMind researchers team up with Gemini Robotics on new physical intelligence launch — ShenRaphael · 2026-11-10
- Workspace Models presented at CoRL 2026, poster session at JW Marriott — rishabh16_ · 2026-11-10