Can SuperMimic Dramatically Cut the Cost of Robot Training?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

WBC is the underlying core of the human simulation movement. Currently, mainstream programs rely heavily on non-model reinforcement learning, and to cover more extensive human movements, group-level computing is needed. Recent work by UC Berkeley's SuperMimic has sparked a heated debate over how to reduce the training of 10,000-level human movement tracking strategies from multi-card clusters for days to a few hours. This work has prompted the industry to rethink the idea of "whether mass-scale whole-body control training is no longer necessary".

Directory

01. What does the profitability of Anthropic mean?
What is Supermimic? Why is GPU clustering no longer a bottleneck for WBC's scale expansion?...

WBC training doesn't necessarily require a lot of computing?

1 Researchers at UC Berkeley recently published the work "SuperMimic", which proposes a full-body control architecture based on a single micro-world model, an effective alternative to existing pathways for computationally-intensive control technologies, and demonstrates significant differences in training efficiency, attracting considerable attention in the AI community.

1 "SuperMimic" was originally a third-year PhD student at UC Berkeley, where its mentors Koushil Sreenath and Pieter Abbeel co-authored the paper. The paper has now been accepted as Oral by CoRL 2026.

In the field of physically intelligent and humanoid robots, the current popular whole-body control systems are mainly two technological routes, both of which have the richest set of motion data and the stronger performance of the controller. However, as the range of motion coverage expands, the computing costs of both systems will increase dramatically.

1 Whole-body control (WBC) is the underlying technology for achieving a robot's high agility and universality, covering movements such as walking, dancing and gesture recovery through a full-body controller to ensure that robots can operate in human-designed environments.

2 In existing popular programs, a focus class combines multiple expert strategies for single-stage movements into a unified controller; one class directly trains single general strategies on large sports arenas.

3 Since both programs rely on non-model reinforcement learning, cannot explicitly simulate the dynamics of the human-robot rigidity that is shared behind different human movements, and can only rely on a large number of simulation trials to estimate the effect of the movements, the calculation costs will increase dramatically as the size of the gym grows.

3 The Sonic, the team's best action tracker, requires 128 NVIDIA L40 GPUs for up to seven days of training on a cluster to allow the robot to master 129,000 different human actions, consuming approximately 21,000 GPU hours.[1-2]

Supermimic has been hotly debated that the work will take the training cost of large-scale human motion tracking strategies from 128 GPUs running for days to compression to a single GPU running for about 3 hours. The researchers said that "calculation will no longer be a core barrier to the expansion of underlying motion control, and the Scaling Law for motion tracking is now available".[1-3]

1 The SuperMimic team released benchmark comparison data with SONIC. In tests containing 129,514 episodes of human action (about 264 hours), SuperMimic ran on a single NVIDIA H100 GPU with a 3-hour convergence rate that surpassed SONIC's public checkpoint and performed better at the failure rate. After 9.8 hours of dual-card operation, its test movements in the training set reached zero failure.

2 Following the publication of the paper, researchers such as Assistant Professor at the Robotics Institute, LeCAR Lab's Centre Stone Conyers, and Genesis AI's founding research scientist, Yunlong Song, expressed their approval of the work on social media and shared ideas for expanding and focusing research on the topic.[1-4] [1-5]

How did Supermimic achieve the victory of a single microscopic world model and the gradient of the solution?

Researchers at Supermimic say that despite the differences in human movements, they all share the same laws of dynamics. Learning these shared laws through a micro-world model provides a rapid imaging and analysis gradient for the strategy, which the team believes is fundamental to achieving a thousand-fold increase in efficiency.