FULL STORY
SuperMimic: Released, Then Replicated in an Hour
UC Berkeley released SuperMimic, a scalable humanoid whole-body control method cutting GPU training time by over 1000x. Within a day, an ETH Zurich researcher replicated it in about an hour using Claude Opus vibe coding.
2026-10-22 ~ 2026-10-22 · 2 episodes · 5 posts
Episode 1 · UC Berkeley's SuperMimic Cuts Humanoid Training GPU Time by Over 1000x (2026-10-22, 2 posts)
UC Berkeley's team released SuperMimic, which combines a physics simulator with a learnable world model to train humanoid whole-body control, cutting GPU time by over 1000x and enabling training on 129k motions in just 3 hours on a single GPU.
- SuperMimic trains humanoid control on 129k motions with 1 GPU in 3 hours, cutting GPU-hours 1000x — kevin_zakka · 2026-10-22
- SuperMimic cuts humanoid full-body control training GPU-hours by over 1000x — kevin_zakka · 2026-10-22
Episode 2 · ETH Researcher Reproduces SuperMimic Humanoid Control in One Hour with Claude (2026-10-22, 3 posts)
Chong Zhang of ETH Zurich used Claude to vibe-code a reproduction of SuperMimic for humanoid whole-body control in about an hour, open-sourcing it as zita-ch. Using a world model plus APG paradigm claimed to speed up training 1000x, it passed the LAFAN benchmark after roughly 1.24 GPU hours.
- World model + APG: single-GPU 1.24h reproduction of SuperMimic humanoid control — ChongZzZhang · 2026-10-22
- ETH researcher vibe-codes SuperMimic reproduction in 1 hour, LAFAN training in 1.5 GPUh — ChongZzZhang · 2026-10-22
- ETH researcher vibe-codes SuperMimic humanoid control repro in 1 hour — ChongZzZhang · 2026-10-22