Boston Dynamics Introduces Research Prototype of Atlas on Wheels

humanoid.guide introduced Boston Dynamics’ wheeled humanoid research prototype, “Atlas on Wheels,” on Nov. 11, 2026. In an official blog post on Nov. 9, Boston Dynamics said the robot, which has wheels in place of feet, is a research test bed for evaluating actuators and learned control, not a product. According to humanoid.guide, the robot is 190 centimeters tall, weighs 90 kilograms and has a top speed of 12 kilometers per hour.

Boston Dynamics describes wheeled Atlas as a research test bed, not a product, and is evaluating how simulation and learned control work together when matched to the characteristics of motors on a physical robot. This shows that model accuracy and control compatibility when transferring Atlas movements from simulation to the physical robot are among the company’s research challenges.

The prototype’s main purpose is not wheeled mobility itself, but a configuration that lets the company examine how physical actuators interact with learned control without substantially rebuilding the existing Atlas. Boston Dynamics said the only new part, apart from the adapter attaching commercially available rims to the ankle actuators, was the software treatment of the wheels as ankles that rotate continuously. Rather than building a dedicated chassis, the company used the existing robot’s mechanisms and model to create the operating conditions. That is the distinctive feature of this research approach. This announcement does not indicate commercialization or deployment for general tasks. The company explicitly calls the wheeled robot a research test bed. humanoid.guide has described capabilities such as sorting items, but the official blog’s focus is on running the motors continuously at high torque and speed and examining whether learned control operates within the physical robot’s performance limits. The mobility demonstration therefore cannot be taken as evidence of task performance or readiness for deployment. The technical focus is the transfer from simulation to the physical robot. If the motor characteristics in a training model differ from those of the real hardware, a policy could learn movements that work in simulation but cannot be executed on the physical robot. Boston Dynamics said it modeled the motors based on measurements from the physical hardware and also compared low-level control settings across software environments. The sequence of millions of simulations, measurements of the hardware’s characteristics, model adjustments and physical validation shows that performance transfer depends not only on training a control policy but also on what the simulation reproduces. Our related article, Boston Dynamics Announces Stretch Gen 2, covered a trailer-unloading robot focused on improving uptime and maintenance efficiency. While Stretch Gen 2 was a product announcement addressing challenges in field operations, wheeled Atlas remains at the research stage, examining the limits of control and actuators. Even for robots from the same company, efforts to improve maintainability in real-world operations and efforts to verify that learned control can transfer to physical hardware should be viewed as distinct development challenges. The next questions are how insights from the wheeled robot’s control and modeling have been applied to the standard Atlas configuration, and how much the company will disclose about validation conditions and reproducibility on the physical robot. The official blog says the insights are being brought back to the standard configuration, but it does not specify the scale of improvement, commercialization plans, operating time or payload conditions. humanoid.guide’s descriptions of speed and capabilities should also be distinguished from performance evaluations that include conditions for real-world operation.

According to Boston Dynamics, the only new part added in the modification was a custom adapter for attaching commercially available scooter rims to the existing ankle actuators. In the software, each wheel was treated as an “ankle” that rotates continuously, replacing the feet in the base robot model with wheels. The company said wheeled movement requires the motors to operate continuously across a wide range of torque and rotational speeds during starts, turns and stops, making it a way to test actuators near their limits. The company said it adjusted the simulated motor characteristics to match measurements from the physical robot for reinforcement learning, in which millions of robot movements are run in simulation before being applied to the real hardware. It also made changes to the reward structure, the range of commands used during training and posture maintenance. humanoid.guide reported that the robot can avoid obstacles, jump and land, and sort items.