Nomura International (Hong Kong)'s latest report indicates that the SuperMimic paper from the University of California, Berkeley, shows that the computational power consumption for training humanoid robot full-body motion controllers can be compressed by over 1000 times. A task that previously required 128 GPUs running for a week can now be completed in just three hours with one H100 GPU. This technology has been accepted by CoRL 2026 and will be published on October 20.

Nomura analysts Frank Fan and Donnie Teng believe this is the most compelling evidence of the "cheapening" of robot body control training costs. They warn that as the barrier to entry for motion software significantly lowers, the industry's competitive focus will shift from algorithmic capabilities to hardware consistency. Companies that can maintain joint precision in mass production will gain a more lasting competitive advantage. SuperMimic achieves over 10 times faster training speed and approximately 64 times fewer training steps by establishing a robot "body model" and training the controller within this model, while maintaining stable model prediction errors. This controller has successfully performed complex actions on a Unitree Robotics G1 real robot.

Nomura emphasizes that SuperMimic primarily addresses advancements at the robot's "cerebellum" (motion control) level and does not involve cognitive decision-making or dexterous manipulation. The report estimates that actuator costs account for approximately 50% to 75% of the total component cost of humanoid robots. Robot manufacturers or key component suppliers with highly consistent control capabilities are expected to benefit. Nomura maintains a "Buy" rating on Unitree Robotics with a target price of RMB 370.