hn.today

SuperMimic: Compute is no longer the bottleneck to robotics

supermimic.github.io7 points2 comments
Screenshot of SuperMimic: Compute is no longer the bottleneck to robotics

SuperMimic presents a method for training one humanoid controller to track a large library of whole-body motions, arguing that compute need not be the main barrier to capable robot control. Its central idea is to train a policy and a differentiable world model together: transitions are collected in a standard contact simulator, the model learns to predict them, and the policy is improved through imagined rollouts using gradients through the learned model - not through the simulator. The model is updated throughout training, and the full three-hour training time includes data collection, model fitting, and policy updates.

On a benchmark of 129,514 motions, SuperMimic exceeds the released SONIC checkpoint’s normalized return of 0.90 after three hours on one H100 GPU. SONIC reportedly required seven days on 128 L40 GPUs; the comparison represents more than 1,000 times fewer GPU-hours, though the systems used different hardware and training data. SuperMimic’s one-GPU policy still fails 25 benchmark motions, while a two-GPU run after 9.8 hours fails none; SONIC fails 179, including every flip and side roll. These motions are in the training set, so the results measure coverage rather than generalization. The site also demonstrates the same untuned policy on a physical Unitree G1. It attributes efficiency to a shared model for the motion library, faster imagined rollouts, and more effective first-order policy gradients.

Read on supermimic.github.io2 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Other

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.