Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Research & technology· 日本語で読む

Boston Dynamics Unveils Research Prototype of Wheeled Atlas, Demonstrates Drifting and 360-Degree Jump

Boston Dynamics unveiled a research prototype of Atlas fitted with wheels. Using reinforcement learning, it achieved drifting and a 360-degree rotational jump, while the company tested alignment between the real robot’s motors and its simulation. The company is also applying insights from the work to Atlas in its standard configuration.

Key facts

  • Boston Dynamics described wheeled Atlas as a research test bed, not a product.
  • The wheels were attached using commercially available scooter rims and existing ankle actuators; the only newly fabricated part was a single adapter.
  • Reinforcement learning used a model matched to the physical robot’s motor characteristics, with millions of simulations run before operating the robot.
  • Boston Dynamics said the 360-degree rotational jump based on human motion data succeeded on the first attempt with the physical Atlas.
  • Boston Dynamics said it is also applying insights from the experiments to Atlas in its standard configuration.

What happened

On November 9, 2026, Boston Dynamics introduced a research prototype of Atlas with wheels in place of feet. It demonstrated drifting and turning on wheels, as well as a 360-degree rotational jump based on human motion data. Gizmodo Japan reported on the experimental footage and development on November 16.

Why it matters

Boston Dynamics said wheeled movement continuously demands high torque and rotational speed from the motors, making it a way to test control near the limits of the real hardware. Incorporating measured motor characteristics into the training simulation is relevant to developing controls that can make simulated movements work on a physical robot.

Analysis

The focus of this prototype is not simply attaching wheels to Atlas, but examining how well whole-body control learned in simulation can be reproduced near the limits of the real robot’s actuators. Boston Dynamics said starting, turning and stopping on wheels continuously demand high torque and rotational speed from the motors. In other words, the wheels are being used less as a proposed mode of transport than as a demanding test condition for motors, low-level control and whole-body reinforcement learning. A related article from this newspaper, “Boston Dynamics Introduces Research Prototype of Wheeled Atlas”, also describes the robot as a research test bed. Key technical points in the latest disclosure include a setup in which a commercially available rim is attached to an existing ankle actuator with a single adapter, and the application of insights to Atlas in its standard configuration. By contrast, “Boston Dynamics Announces Stretch Gen 2” concerns a truck-unloading robot focused on improving uptime and maintenance efficiency. The product’s operational challenges and research into Atlas’s limits should not be treated as equivalent deployment records. Technically, if a training simulation over-idealizes the motors, it may produce controls that cannot be executed on the real robot. Using a motor model calibrated to measured values connects training conditions with hardware behavior and appears to be a step toward transferring control from simulation to reality. The drifting and jumps based on human motion data are best understood not as direct reproductions of those motions, but as examples of control learned to accommodate the characteristics of wheels. However, the published explanation does not specify the scope of control changes applied to Atlas in its standard configuration or provide quantitative results for performance or durability in the experiments. The next questions are which movements or failure conditions are covered by the application to the standard configuration, and how controls developed on a research robot will contribute to product reliability evaluations.

Details

Boston Dynamics said the robot is a research test bed, not a product. The only newly fabricated part was a single adapter to attach commercially available scooter rims to the existing ankle actuators. In software, the wheels were modeled as continuously rotating “ankles.” The company used reinforcement learning for motion control and ran millions of simulations before operating the physical robot. It adjusted the motors’ characteristics in the simulation based on measurements and also used human motion data for training. The company said the 360-degree rotational jump succeeded on the first attempt with the physical robot. Insights from the work are also being applied to Atlas in its standard configuration.

Sources

Translated from the Japanese original: /news/boston-dynamics-sharin-tsuki-atlas-no-kenkyuu-shisakuki-wo-k-c1ce7f. The Japanese version is authoritative if they differ.