Berkeley's SuperMimic compresses humanoid robot motion control training compute by over 1,000x

A major technological breakthrough in the field of robot motion control is reshaping the competitive landscape and value distribution of the humanoid robot industry.

According to Zhuifeng Trading Desk, Nomura International (Hong Kong) stated in its latest report that a paper from the University of California, Berkeley titled SuperMimic shows that the training compute consumption for a humanoid robot's whole-body motion controller can be compressed by more than 1,000 times—a task that originally required 128 GPUs running for a week can now be completed with a single H100 GPU running for three hours. The paper has been accepted by CoRL 2026, a top conference in the robot learning field, and was published on arXiv on October 20.

Nomura analysts Frank Fan and Donnie Teng pointed out in the report that this is the strongest evidence to date that the cost of robot body control training is becoming "cheap." They also warned that as the barrier to motion software drops significantly, the industry's competitive focus will shift from algorithm capability to hardware consistency, and companies able to maintain joint precision in mass production will gain more lasting competitive advantages.

Sharp Drop in Compute Demand: How SuperMimic Achieves a Thousand-Fold Efficiency Leap

The core innovation of SuperMimic lies in changing where the robot "practices." Traditional methods rely on physical simulators for repeated trial and error, and with each additional motion, the required compute grows linearly. The previously best-performing public system, SONIC, took about 21,000 GPU hours to train.

SuperMimic's approach is entirely different: it first builds a "body model" for the robot itself, learning the physical laws of that robot's motion, and then trains the controller inside this model. The Nomura report, citing data from the paper, says this method brings three efficiency gains: first, training inside the learned model is more than 10 times faster than in a physical simulator, reaching 62 times at the largest batch size; second, the model is differentiable, allowing the controller to directly learn the adjustment direction for each motion rather than relying on random trial-and-error inference, which alone cuts the required training steps by about 64 times; third, as motion data expands from 1 hour to 264 hours, the model's prediction error remains stable, and compute demand grows extremely slowly.

In benchmark tests, SuperMimic failed only 25 times across 129,514 motions, compared with 179 failures for SONIC; after 9.8 hours of training on two GPUs, SuperMimic achieved zero failures. The controller then completed high-difficulty motions such as side rolls and 360-degree flips on a real Unitree G1 robot.

The Nomura report also notes that the above thousand-fold comparison is not a strictly like-for-like comparison—SuperMimic used about 2.3 times less motion data and adopted a newer-generation GPU model, but the authors believe these factors are insufficient to explain the thousand-fold-scale gap.

Technical Boundary: What SuperMimic Solves Is the "Cerebellum," Not the "Cerebrum"

Nomura analysts emphasized that understanding SuperMimic's scope and limitations is equally critical for assessing its industry impact.

The study covers only the robot's body motion on flat ground, does not involve object grasping and manipulation, the model training data comes entirely from simulators, and the motions used in the benchmark overlap with the training motions. Nomura therefore characterizes SuperMimic as progress at the level of the robot's "cerebellum" (motion control), rather than a breakthrough at the level of the "cerebrum" (cognitive decision-making) or the "hands" (dexterous manipulation).

This distinction has practical significance for investors judging the speed of technology diffusion. The decline in motion control costs does not directly equate to an all-round leap in the overall capabilities of humanoid robots, but it does significantly lower the entry barrier for competitors at the motion software level.

Competitive Landscape Reshaped: Hardware Consistency May Become the New Moat

The Nomura report argues that the most profound impact of SuperMimic lies in changing the value distribution logic of the humanoid robot industry.

Citing Nomura industry survey data, the report notes that about 300 Chinese companies can now provide a certain key motion control function, and Unitree's 1-2 year lead in hardware and motion control is narrowing. SuperMimic further lowers this barrier, making work that originally required a GPU cluster achievable with a single GPU.

Against this backdrop, Nomura believes the center of competitive advantage will shift from motion software to hardware consistency. The report specifically notes that SuperMimic's controller, in simulation training, requires robot joint calibration error of only ±0.01 radians. Nomura estimates that if the kinematic parameters of more than 20 joints are inconsistent, the same set of reinforcement learning code may run normally on only a few machines. This means that companies able to achieve high joint consistency in mass production will gain stronger differentiation advantages after the technology becomes widespread.

Nomura further estimates that actuator costs account for about 50% to 75% of the total cost of humanoid robot components. The report therefore judges that robot OEMs or key component suppliers with highly consistent control capabilities may benefit more than motion software specialists. Nomura maintains its "Buy" rating on Unitree with a target price of RMB 370, corresponding to 25 times the expected 2027 price-to-sales ratio, citing the company's expected 122% compound annual sales growth from 2026 to 2028 and its mass production capability.

~~~~~~~~~~~~~~~~~~~~~~~~

The above excellent content comes from Zhuifeng Trading Desk.

For more detailed interpretation, including real-time commentary and frontline research, please join [Zhuifeng Trading Desk▪Annual Membership] Risk Disclosure and Disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.