Boston Dynamics Tests Reinforcement-Learning Control on Atlas Fitted With Wheels

In a blog post published November 9, 2026, Boston Dynamics said it replaced the feet of its bipedal Atlas robot with wheels and tested a trained controller on the real robot. The wheeled configuration is not a product but a research testbed for continuously operating actuators at high torque and speed, the company said. To attach the wheels, it used an adapter to fit commercially available scooter rims to the existing ankle actuators; the adapter was the only new part.

Boston Dynamics adjusted the simulation to match the real-world motor characteristics and tested the transfer of learned control to a physical robot. The example shows that aligning a learned model with actuator characteristics is a challenge when applying reinforcement learning to robots.

The technical focus is not the footage of Atlas moving on wheels itself, but the accuracy of the model when transferring reinforcement-learned control to real-world motors. Boston Dynamics described the wheeled configuration as a research testbed, not a product. The setup, which used commercially available scooter rims attached to existing ankle actuators, was not an experiment to design a new drive system and demonstrate its performance. The aim appears to have been to create conditions for continuously operating the motors at high torque and speed and examine interactions between learned control and the real actuators. Limiting new hardware to a single adapter and treating the wheels in software as ankles made this an example of reusing Atlas' existing mechanisms and control models to test different operating conditions. The company also compared control parameters in the simulation with the real robot and adjusted the model to match measured motor characteristics. This shows that having trained a system in simulation alone is not enough to conclude it will perform the same way on a physical robot. The blog says that, as an explanation of reinforcement learning, millions of simulations are run before deployment on a real robot, but it does not give the number of training runs in this experiment or quantitative performance data from the physical robot. Related coverage from Physical AI News includes Agility Robotics to go public on Nasdaq through SPAC merger and Report: Tesla expected to begin taking Optimus reservations in early December. Those reports concern commercialization, while the Atlas blog focuses on research into transferring control to a physical robot; they represent different stages of robotics businesses. The test does not indicate that Atlas is moving toward mass production or commercial deployment. The broader industry implication is that real-world performance depends not only on reinforcement-learning algorithms but also on alignment among actuator design, low-level control and whole-body control models. Boston Dynamics tested how these layers connect using the wheeled configuration and is applying the findings to the standard configuration. This is a development challenge that is difficult to capture by evaluating the robot hardware and learning system separately. However, the blog provides no quantitative data on torque, speed, payload, failure rates or the scale of improvements in the standard configuration. Key questions going forward include which performance metrics the models and controls adjusted in the wheel tests improved in the standard configuration, and how reliability in real-world environments was evaluated.

In software, each wheel was modeled as a continuously rotating "ankle," replacing the feet in the baseline robot model with wheels. To narrow the gap between reinforcement-learning simulations and the physical robot, the company compared low-level control parameters across software environments, characterized the physical robot and adjusted the simulated motor model. It also changed the reward structure, the range of commands used during training and how the robot maintained its posture. According to the blog, some of the control used human motion-capture data. Demonstrations included lateral movement, in which a foot-sliding motion was converted into wheel steering, and a 360-degree jump that followed a human jump recording. The company says it is also applying findings from the wheeled configuration to the standard configuration.