Robot motion training takes far less computing power. Why does Nomura place greater value on hardware consistency?

Training compute fell more than a thousandfold. What exactly improved?

SuperMimic substantially lowered the training threshold for whole-body motion control in humanoid robots: the paper’s authors say that after three hours of training on a single H100 GPU, it outperformed SONIC’s published model. The latter used approximately 21 thousand GPU hours for training, making the reduction in GPU hours more than a thousandfold.
Motion-tracking performance improved alongside the reduction in compute. In a benchmark of 129,514 motions, SuperMimic failed 25 times, while SONIC failed 179 times; the former achieved zero failures after training on two GPUs for 9.8 hours. The controller subsequently performed a sideways roll and a 360-degree flip on a real Unitree G1 robot.

Traditional methods have robots repeatedly learn by trial and error in a physics simulator; the larger the motion library, the greater the computing requirements. SuperMimic first learns how a single body moves, then lets the controller practice within that body model; different motions can reuse the same set of rules governing the body.
The savings come from both training speed and learning efficiency: at common batch sizes, practice within the model is more than 10 times faster than in the simulator. The model can also directly provide information on the direction in which a motion should be adjusted; this mechanism alone reduces the number of practice steps required by approximately 64 times.
When motion data increased from 1 hour to 264 hours, the body model’s prediction error remained stable, so the computing requirements from additional motions grew more slowly. On this basis, Nomura believes the key to the progress is changing where practice takes place, rather than adding data.

The “more than a thousandfold” comparison is not entirely like-for-like: SuperMimic used less motion data, with SONIC’s dataset about 2.3 times as large, and the former used a newer GPU model. The paper’s authors believe these differences are insufficient to explain the thousandfold gap.
The results are limited to body movements on flat ground and do not cover object manipulation or complex terrain; the body model learned only from simulation data, and the motions in the benchmark were also training motions. Nomura therefore characterizes it as progress in the “cerebellum” of motion control, with the boundaries of its capabilities still confined to body control.

Nomura’s industry survey shows that around 300 Chinese companies can already deliver a key motion-control function, and Unitree’s existing one-to-two-year lead in hardware and motion control may be narrowing. As training becomes simpler, the bank believes Unitree’s competitiveness will depend more on cost and scale.
Whether software can be reused across large numbers of robots depends on whether their bodies are sufficiently consistent. Joint calibration varied by only ±0.01 radians in the paper’s simulations; Nomura estimates that if the motion characteristics of more than 20 joints are inconsistent, the same reinforcement-learning code may work on only a few machines.
Nomura estimates that actuators account for approximately 50%–75% of humanoid robot component costs. The bank favors robot manufacturers and key suppliers capable of controlling consistency, and believes Unitree has an advantage over model companies that build their own hardware, because mass-production experience takes years to accumulate. The link to potential benefits is manufacturing stability, rather than simply possessing motion algorithms.