SuperMimic cuts humanoid full-body control training GPU-hours by over 1000x
kevin_zakka · x · 2026-10-22
The Berkeley team behind SuperMimic (retweeted by Pieter Abbeel) reports a massive reduction in training cost for humanoid full-body control:
- A single policy trained across 129,514 motions;
- Surpasses prior SOTA (128 GPUs for 7 days) using 1 GPU for 3 hours — over 1000× fewer GPU-hours;
- Two key ideas: a world model much faster than regular simulation, and backpropagation through time collecting far more signal than policy gradients.
Project page and demo video are public.
Related event: UC Berkeley's SuperMimic Cuts Humanoid Training GPU Time by Over 1000x(2 posts)→
More from Embodied
- Unitree's H2-X humanoid claims 8 hours of endurance—but that's just standing time — rohanpaul_ai · 2026-10-22
- Shenzhen AI hardware founder catalogs every AI hardware project actually worth seeing — Scobleizer · 2026-10-22
- HF tutorial: train Microduck in Isaac Lab 3.0 sim, deploy on real robot — Hugging Face Blog · 2026-10-22
- Robert Downey Jr. stars in promo for Snap's new futuristic smart glasses — CodeByPoonam · 2026-10-22
- First folding iPhone review: exciting for apps, but more like the original iPad — film_girl · 2026-10-22
- OpenTac tactile skin survives hammer, stomping, and office chair stress tests — _krishna_murthy · 2026-10-22