Boston Dynamics fits Atlas with wheels to test whole-body control using RL and classical control

Boston Dynamics has released a video showing its Atlas humanoid moving on wheels instead of feet. According to The Robot Report, the experiment combines classical model-based motor control with whole-body reinforcement learning (RL) to maintain dynamic stability on real, unstructured terrain. The company itself stresses that a wheeled Atlas is not a product but a research testbed.

The video features Andy Wu, an Atlas control engineer at Boston Dynamics, who describes recent experiments in which he attached wheels to the robot's legs. His goal was to push actuation to its limits and better understand adaptability to unstructured environments. Wu discusses the limitations of purely model-based control when real-world conditions are unpredictable and hard to model or simulate.

Using a whole-body RL policy, Atlas processes proprioceptive feedback and velocity commands in real time. The system also generates dynamic posture and force adjustments to compensate for unmodeled disturbances and less-than-ideal ground contact.

Besides learning to move on wheels, Atlas performs spins and other aggressive maneuvers, in a style the article likens to a skater's. This shows it can switch smoothly between rolling quickly over flat surfaces and articulated legged movement over obstacles or uneven ground.

Wu trained the system in simulation. According to him, the robot learned all of its behaviors in simulation and executed them in the real world with surprising accuracy.

Wu said converting the locomotion mode was "wonderfully simple." A single custom adapter mounts an off-the-shelf scooter wheel onto Atlas's existing ankle actuator, and it is the only new part in the entire setup. Because Atlas was designed from the outset for volume production and easy repair, the conversion was comparable to changing a bicycle wheel. In software, each wheel is modeled as an "ankle" that can rotate continuously, so it was enough to take the robot's base model and replace the feet with wheels.

The next question Wu wants to tackle is how the policy responds to failed actuators, which is key for safety and for understanding how the system would react to a fault. The team's main takeaway is an understanding of the synergy between control methods. According to the video, combining classical control (baseline stability) with whole-body RL (reactive adaptation) delivers greater real-world agility than either approach on its own.

The experiment adds to the wheels-versus-legs debate in humanoid robotics. Many companies are developing wheeled humanoids, while others are betting on legs to navigate environments designed for people. The article notes that Agility Robotics has said it is also exploring wheeled robots as part of a broader family of multipurpose platforms. Its co-founder and chief robot officer, Jonathan Hurst, said Digit is unlikely to get wheels, although the company is studying other wheeled configurations.

The piece raises the question of how much of the human form a humanoid needs to be commercially useful. The experiments come as Boston Dynamics prepares to expand Atlas trials at Hyundai facilities and continues to develop its capabilities, including a new four-fingered hand.

Reader comments repeat the usual arguments: wheels seem more efficient on flat floors, though their stability under payload is questioned, and some advocate hybrid or modular designs, with wheels where they suffice and legs where there are stairs or uneven terrain. These are reader opinions, not data from the experiment.

The material is a press article about a video, with no performance metrics, speeds or quantitative results, so it is not possible to assess the real extent of the improvements beyond what the company and its engineer claim.