Boston Dynamics puts Atlas on wheels to test its motor limits

The more consequential work involved matching simulation to hardware. If simulated motors are too idealized, a learned controller—or policy—can exploit capabilities that physical motors lack. Wu describes checking low-level control parameters across software configurations, measuring hardware behavior and adjusting the simulated motor model to match those measurements.

The team adapted established reinforcement-learning frameworks, changing the rewards used to guide learning, command ranges and posture. Wu says smooth rolling emerged from those adjustments. The company also credits Google DeepMind’s MuJoCo team for the open-source physics engine supporting the work. The experiment therefore combines existing hardware, simulation infrastructure and learning software rather than introducing an entirely new control system.

Physical and simulated versions of wheeled Atlas in Boston Dynamics’ company-reported control experiment. Credit: Boston Dynamics. Source: Boston Dynamics.

What the jumps demonstrate

Some movements started with a library of human motion-capture recordings. Wu selected clips involving sliding feet. In side dashes, the policy learned to steer the wheels and roll through the motion instead.

The 360-degree jump likewise followed a recorded human trajectory. Wu says it worked on the first attempt and was among the lower-risk moves, despite its energetic appearance. During the preparation for the jump, the lower body rolled into position before pushing upward. The demonstration shows adaptation to a different way of contacting the floor, with a human movement providing the target.

In the accompanying video, Wu describes the transfer from simulation to hardware as “basically perfectly.” That is his characterization: the announcement supplies no numerical transfer-error measure, repeated-trial success rate or endurance result. He also acknowledges that he could have tested a more conventional behavior on Atlas with feet.

Wheeled humanoids have precedents and commercial alternatives. A report on the experiment describes Unitree Robotics’ April demonstration of its G1 humanoid rolling on replacement wheels. It also identifies Swedish firm Hexagon’s AEON as an industrial humanoid designed around wheeled locomotion, with BMW bringing it into its Leipzig plant and a reported 1,000-unit commitment from Schaeffler.

Those are different kinds of evidence: a demonstration, an industrial introduction and a unit commitment do not establish comparable performance. They do show why putting wheels on a humanoid cannot itself establish novelty or commercial advantage. Boston Dynamics says its wheeled Atlas will remain a research tool.

Hyundai’s stake extends beyond a demonstration

Hyundai’s interest combines ownership, robot development and its own manufacturing operations. In a July 16 statement, the group said its shareholders were pursuing the acquisition of SoftBank’s entire stake in Boston Dynamics after SoftBank exercised a contractual right to require a purchase. The statement described governance reviews, approvals and settlement steps still pending; it did not announce a completed transaction or a purchase price.

Hyundai described a strategy combining Boston Dynamics’ robotics expertise with the group’s manufacturing capabilities and global value chain. It also pointed to Atlas demonstrations including transporting a 23-kilogram compact refrigerator. That is evidence of a demonstrated lifting task, not a continuous-duty payload rating or a measure of production throughput.

The group’s stated deployment plan starts with parts sequencing at Hyundai Motor Group Metaplant America in 2028. It envisages extending Atlas to component assembly by 2030, subject to technology validation, operational readiness and business requirements. These are planned applications and dates, rather than results achieved by the wheel experiment.

Hyundai’s company-reported Atlas deployment targets, subject to validation, operational readiness and business requirements. Source: Hyundai Motor Group. Source: Hyundai Motor Group.

The next evidence is recovery and sustained operation

Boston Dynamics says the wheel-test insights are already being adapted to standard Atlas. Wu says his next project is improving how the learned controller responds to failed actuators, including whether reinforcement learning can produce useful recovery behavior.

The next question is how reliably those changes work across repeated trials and sustained operation. The announcement gives no failure-frequency comparison, operating duration or cost per completed task. Those measures would help establish whether better motor modeling and failure responses deliver the dependable operation Hyundai’s factory plans require.