Boston Dynamics has unveiled the humanoid robot “Atlas” with its feet replaced by wheels, demonstrating rapid driving and a 360-degree spinning jump. It applied reinforcement learning (RL)-based control technology to learn a new mode of locomotion in a research experiment designed to verify the performance of the robot’s actuators and AI control model.
Boston Dynamics showed wheeled Atlas driving quickly and changing direction in a real test space through an official video released on the 9th (local time). In addition to sprinting sideways and performing a 360-degree spinning jump, the robot demonstrated balancing on one wheel and getting up on its own from a prone position before driving away.
The key objective of this experiment was to determine whether a whole-body control model trained through reinforcement learning could operate stably using the new mode of locomotion. Wheel-driven movement continuously demands high torque and motor speed during starting, turning, and stopping, making it well suited for testing the performance limits of the actuators.
The hardware modification was relatively simple. Only one custom adapter was added to enable ordinary scooter wheels to be attached to the existing ankle actuators. On the software side, the wheels were modeled as ankles capable of continuous rotation, allowing the feet of the existing robot model to be replaced with wheels.
The researchers did not individually program the wheel-driving behaviors. Instead, they incorporated the characteristics of the wheels into the existing whole-body reinforcement learning model and adjusted the training conditions, enabling the AI to learn how to roll the wheels while maintaining balance and changing direction.
It was also important to reduce the differences between simulation and hardware so that the reinforcement learning model could function properly on the physical robot. If motor movements are modeled too ideally in a virtual environment, the AI may learn motions that the actual hardware cannot perform. The researchers measured the characteristics of the real motors and adjusted the simulation model and control parameters to reduce what is known as the “simulation-to-reality gap (Sim2Real Gap).”
Some movements were learned using motion capture data of human actions. By applying the sliding motion of human feet to wheel-driven locomotion, Atlas learned to steer its wheels while moving sideways in a sliding manner. The 360-degree spinning jump was learned by tracking motion capture data recorded from a human performer, and it succeeded on the first attempt on the actual robot.
Heard some people like wheels?
— Unitree (@UnitreeRobotics) April 23, 2026
Humanoid robots are the ideal form of general-purpose robots (perfect for general AI and human-derived data). They can work without wheels — but they can also have wheels if they want. Whatever works. pic.twitter.com/Ms24NTlIvY
Chinese company Unitree had previously introduced a wheeled humanoid robot. Last April, it attached wheels, skates, and other equipment to its humanoid robot “G1” and demonstrated high-speed driving, turning, balancing, and more.
While Unitree focused on implementing various mobility methods, Boston Dynamics focused on validating the performance of existing actuation devices and reinforcement learning control models, and applying them to the bipedal Atlas.
Boston Dynamics emphasized that this wheeled Atlas is a research test platform rather than a commercial product. The motor modeling and whole-body reinforcement learning technologies developed through the experiments are also being applied to the Atlas that uses conventional feet.
The goal is to improve stability in real-world environments and the ability to respond to failures, while expanding the robot's range of motion.
Reporter Park Chan cpark@aitimes.com
#AD176958096310.ad-template { margin:auto; position:relative; display:block; clear:both; z-index:1; padding-bottom:20px; } #AD176958096310.ad-template .col { float:left; text-align:center; width:100%; } #AD176958096310.ad-template .col .ad-view { position:relative; } #AD176958096310::after {content:"";display:block;clear:both;}
