---
type: "post"
id: "atlas-on-a-roll-c3da0b"
slug: "atlas-on-a-roll-c3da0b"
title: "Atlas On a Roll"
description: "Boston Dynamics used scooter wheels on Atlas to stress its actuators and study how a whole-body reinforcement-learning controller transfers from simulation to hardware."
retrieval_nugget: "Boston Dynamics used scooter wheels on Atlas to stress its actuators and study how a whole-body reinforcement-learning controller transfers from simulation to hardware. The experiment is not a product, but it shows why accurate motor models and low-level controls matter when teams want learned behavior to survive contact with a real robot."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["automation","ai"]
entities: ["Boston Dynamics","Atlas","Xiaoyu Wu","MuJoCo"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Boston Dynamics is testing Atlas with scooter wheels to study reinforcement-learning locomotion and actuator torque limits."
prior_refs: []
source_url: "https://bostondynamics.com/blog/atlas-on-a-roll"
source_urls: ["https://bostondynamics.com/blog/atlas-on-a-roll"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:atlas-on-a-roll-c3da0b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/atlas-on-a-roll-c3da0b/","markdown":"https://newruntime.com/posts/atlas-on-a-roll-c3da0b.md","json":"https://newruntime.com/posts/atlas-on-a-roll-c3da0b.json"}
markdown_url: "https://newruntime.com/posts/atlas-on-a-roll-c3da0b.md"
---

# Atlas On a Roll

## Retrieval answer

Boston Dynamics used scooter wheels on Atlas to stress its actuators and study how a whole-body reinforcement-learning controller transfers from simulation to hardware. The experiment is not a product, but it shows why accurate motor models and low-level controls matter when teams want learned behavior to survive contact with a real robot.

## Summary
Boston Dynamics used scooter wheels on Atlas to stress its actuators and study how a whole-body reinforcement-learning controller transfers from simulation to hardware. The experiment is not a product, but it shows why accurate motor models and low-level controls matter when teams want learned behavior to survive contact with a real robot.

## What happened
Boston Dynamics controls engineer Xiaoyu (Andy) Wu describes a research testbed in which the company replaced Atlas’s feet with off-the-shelf scooter rims. A single custom adapter mounted each rim to an existing ankle actuator; the company says this was the only new part in the build. In the robot model, each wheel was represented as an ankle able to spin continuously, allowing the team to start from its baseline robot model and substitute wheels for feet.

The objective was not simply to make a humanoid roll. Boston Dynamics wanted a demanding and continuous test of actuator performance: every launch, turn and stop made the motors work across a broad torque-speed range while carrying the robot’s load. The project also gave Wu, whose background is in classical motor control, a way to work across actuator design, low-level control and whole-body reinforcement learning.

The team trained in simulation before running the controller on Atlas. The company says reinforcement learning ran through millions of simulated robot experiences, but useful transfer depended on making the simulated motors resemble the real actuators. Engineers checked that low-level control parameters matched across software setups, ran characterization experiments and tuned the simulated motor model against measurements from the physical robot. They also had to decide which approximations were safe and which could be exploited by a learned policy.

Boston Dynamics adapted existing policy frameworks rather than building a new stack from scratch. It made targeted changes to reward composition, the range of training commands and how the policy held posture. The company reports that smooth rolling emerged from training and that the robot remained stable and easy to drive. In side-dash movements inspired by human motion-capture clips, the policy learned to steer and roll the wheels where a person would slide their feet. A 360-degree jump followed a recorded human trajectory and worked on its first try, according to Wu.

The company explicitly calls wheeled Atlas a research testbed, not a product. Wu says the work is informing the standard Atlas configuration, and that his next project concerns policy responses to actuator failures. Boston Dynamics frames the longer-term goal as improving reliability at performance limits and making learned control more robust to failure modes.

## What the archive adds
New Runtime’s earlier coverage of PhoneBot (October 18, 2026) described a low-cost humanoid using a commodity smartphone as a sensing and computing module ([PhoneBot](https://newruntime.com/posts/phonebot-a-low-cost-open-humanoid-robot-platform-e91d3b/)). The Atlas work is a new event, not a continuation: it focuses on actuator stress-testing and learned control transfer rather than low-cost robot architecture.

## Trend signal
This is consistent with the idea that production-agent performance depends on the system around a model, not model choice alone. For robotics, Boston Dynamics’ account makes the corresponding point about learned policies: actuator dynamics, low-level controls and the training harness have to work together. The connection to [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) is analogical, not direct evidence about software agents.

## Why it matters
A policy can appear successful in simulation while relying on motor behavior that the physical robot cannot reproduce. Boston Dynamics’ process makes the verification work visible: match control parameters, characterize the hardware, tune the simulated plant and examine which simplifications a policy might exploit. That is a useful engineering lesson for teams building embodied AI systems, where training throughput alone is a poor measure of readiness.

The experiment also shows the value of unusual testbeds. Wheels create sustained torque and speed demands that feet do not, exposing interactions among actuator limits and whole-body control. The reported transfer is encouraging, but it is not a quantified reliability result. Teams should treat it as a design example for evaluating sim-to-real gaps, not as proof that the method generalizes to other robots or tasks.

## Limits and uncertainty
Boston Dynamics does not publish quantitative actuator loads, torque-speed measurements, training duration, policy evaluation metrics or a comparative failure rate in this account. The description is the company’s own account, and independent testing of the transfer or robustness claims is not presented. It also does not establish how well the approach will handle actuator failures; Wu describes that as future work. The wheeled configuration remains a testbed, not a customer product.

## Skill to build now
For a simulation-trained robot policy, make a short hardware-to-simulation calibration checklist: verify low-level parameters across environments, characterize the real actuator under representative loads, tune the simulated motor model against those measurements, and test whether the policy exploits modeling shortcuts. Boston Dynamics’ account gives each of these steps as part of its own transfer process.

## Reader takeaway
The unusual part of Atlas on wheels is not the stunt; it is the calibration work beneath it. Boston Dynamics says learned whole-body control transferred because the simulated motors and physical actuators were brought into close agreement. For teams building robots, that makes hardware modeling and policy verification core parts of the learning system, not cleanup after training.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Automation](https://newruntime.com/topics/automation.md): Explore the automation topic hub.
- topic hub: [Ai](https://newruntime.com/topics/ai.md): Explore the ai topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)