· Research & technology· 日本語で読む
Boston Dynamics Tests Reinforcement-Learning Control With Wheeled Atlas
Boston Dynamics used Atlas fitted with wheels as a research test bed for reinforcement-learning control. The company described the work in an official blog post dated November 9, 2026, and Robotics and Automation News reported on it with video on November 22. The focus was not wheel locomotion itself, but adapting simulation-trained control to real actuators.
Key facts
- Boston Dynamics describes wheeled Atlas as a research test bed, not a product.
- A custom adapter was used to attach commercially available scooter rims to the existing ankle actuators; it was the only newly made part.
- The company chose wheeled locomotion as a condition that continuously demands high torque and fast rotation under real loads.
- The reinforcement-learning work used millions of simulations, and the simulated motor model was adjusted to match measurements from the real robot.
- The 360-degree jump was performed using a trajectory-tracking policy that followed a recording of human motion; the company says it succeeded on the first attempt.
What happened
On November 22, 2026, Robotics and Automation News showed footage of Boston Dynamics’ Atlas moving on wheels, skidding and performing a 360-degree jump, and reported on research into motion control using reinforcement learning. Boston Dynamics described the same work in an official blog post dated November 9, positioning wheeled Atlas as a research test bed, not a product.
Why it matters
Boston Dynamics says the challenge in the research is to incorporate motor characteristics measured on the real robot into a simulation model, then run the learned control on physical hardware. The work is relevant to developers examining how to align actuator modeling and whole-body control when transferring learned control to real robots.
Analysis
The focus of the work is less the installation of wheels on Atlas than the test of whether measured motor characteristics from the real robot can be incorporated into a training simulation and the whole-body control developed there transferred to hardware. Boston Dynamics says wheeled locomotion places sustained demands on high torque and fast rotation under real loads. In other words, the design tests the intersection of actuators, low-level motor control and whole-body reinforcement learning, rather than proposing a new mode of locomotion.
We introduced wheeled Atlas as a research test bed in an article dated November 9 Boston Dynamics introduces research prototype of Atlas fitted with wheels. Robotics and Automation News published its article on November 22, but the official blog also described the same work on November 9, so there is no basis to treat the later report as a new product announcement or specification change. The claim that only one new part—an adapter—was needed to fit the wheels demonstrates a design feature the company has cited for Atlas’ production and serviceability in a research modification; it is not a figure showing operational results for production units.
From a control-development perspective, the work illustrates that simulation accuracy depends not only on high-level learning, but also on the alignment of measured motor characteristics with low-level control parameters across software components. The company conducted millions of simulations, as well as characterizations of the real robot and adjustments to the motor model. Skidding and 360-degree jumps using recordings of human motion are described as examples of the learned policy adapting its movements to the wheels’ characteristics, but they do not establish that the same behavior will transfer to other tasks or environments.
The company says it is also applying the findings to Atlas in its standard configuration, indicating an intention to feed insights from the wheeled experiment back into the broader control design of Atlas. However, publicly available information does not establish the motor model’s error, the number of test repetitions, the durability or safety of the wheeled configuration, or the specific changes made to the standard configuration. Key questions are whether the learning results are reproducible and under what conditions and metrics control performance was evaluated on the real robot.
Details
Boston Dynamics removed Atlas’ feet and attached commercially available scooter rims to the existing ankle actuators. The company says the only newly made part was one custom adapter. It chose wheeled locomotion as a condition for testing the motors and their model, because it continuously demands high torque and fast rotation under real loads.
The company used millions of simulations to train the control system and adjusted the simulated motor model to match characteristics measured on the real robot. Some movements used human motion-capture recordings, and the 360-degree jump was performed using a trajectory-tracking policy that followed a recording of human motion. Boston Dynamics says it is also applying insights from this work to Atlas in its standard configuration.
Sources
Translated from the Japanese original: /news/boston-dynamics-sharin-tsuki-atlas-de-kyoukagakushuu-no-seig-0318fd. The Japanese version is authoritative if they differ.