# Boston Dynamics put Atlas on wheels to stress-test its motors and its learned controller

> Swapping feet for scooter rims let engineers push actuators to their limits. The simulation-trained policy transferred almost perfectly, and even landed a 360-degree jump first try.

- URL: https://devisdead.ai/news/boston-dynamics-put-atlas-on-wheels-to-stress-test-its-motors-and-its-learned-co
- Published: 2026-11-09T20:37:11Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: boston dynamics, atlas, robotics, reinforcement learning
- Originally reported by Unite.AI (https://www.unite.ai/boston-dynamics-turns-atlas-into-a-wheeled-actuator-testbed/).

Boston Dynamics swapped Atlas's feet for wheels, and it wasn't for a product. The wheeled humanoid is a research testbed for pushing motors to their limits and checking how well a learned whole-body controller survives the move from simulation to real hardware, according to [Unite.AI](https://www.unite.ai/boston-dynamics-turns-atlas-into-a-wheeled-actuator-testbed/), whose article was written by an AI research agent and reviewed by its editors.

## Why wheels

Accelerating, turning and stopping on wheels keeps the motors under heavy load across their full range of torque and speed. Years of deploying Spot and Stretch taught the company that real robots meet conditions a lab can't reproduce, so the goal was to expose how actuators and the learned controller behave near their limits. Atlas controls engineer Xiaoyu (Andy) Wu wrote up the work.

## Sim-to-real

The controller was trained across millions of simulations. The hard part was making the simulated motor behave like the real one; a policy trained on an idealized motor may rehearse moves the hardware can't make. The team matched low-level control settings, ran motor characterization experiments, tuned the simulated motor to Atlas's measurements, and checked each simplification so none gave the policy an unrealistic advantage.

The project combined actuator design, motor control and reinforcement learning, usually separate specialties. Wu, a classical motor-control engineer on his first big RL project, says adjusting rewards, training commands and posture handling produced smooth, stable rolling and an essentially perfect transfer to hardware.

## The demos

Some moves started from human motion-capture clips with lots of foot sliding, which the policy turned into steered rolling. A 360-degree jump tracked a recording of a human jumper and worked on the first try. The hardware change was tiny: a custom adapter that bolts a standard scooter wheel rim onto the ankle motor Atlas already had, modeled as an ankle that spins continuously.

## What carries over

Lessons are already being applied to the standard, two-legged Atlas, aiming for policies more robust to failures. Wu is also starting work on policies for failed actuators. For robotics teams, the takeaway is that accurate motor models matter as much as the learning algorithm.

**[Scoreboard](/scoreboard) update:** AI +1. A simulation-trained robot controller transfers to real hardware almost perfectly.

Stay undead.