World model + APG: single-GPU 1.24h reproduction of SuperMimic humanoid control
ChongZzZhang · x · 2026-10-22
ChongZzZhang shares a new "world model + APG" paradigm enabling 1000x faster training of generalist whole-body humanoid trackers, claiming Claude Opus 5.5 high reproduced the SuperMimic paper (arXiv 2610.28480) via vibe coding within 1 hour.
The open-source repo supermimicvibed:
- Unofficial reproduction (no official code), built from the paper on Isaac Lab 3 (kit-less, MuJoCo-Warp)
- Targets the paper's Fig.4a: 40 G1-retargeted LAFAN1 clips (2.45h) on one GPU
- Chart shows normalized return 0.829 reached in 1.24h (paper: 0.82 in 1.2h on H100), with 40/40 clips passing evaluation after 1.47h of training
- Includes Dockerfile, runtrain.sh, and a MILESTONES.md log; total cost 1.5 GPU-hours
The author also muses that the community risks getting dumb and falling into "scaling slop," calling for more researchers like qiayuanliao.
Related event: ETH researcher vibe-codes SuperMimic reproduction in about an hour(2 posts)→
More from Embodied
- Samsung tests humanoid robot for 24/7 factory shifts, weighing CES 2027 debut — CyberRobooo · 2026-10-22
- A week with the dual-screen iPhone Duo: gestures, apps and the camera compromise — BenBajarin · 2026-10-22
- First public demo of automated T-shirt production line in North Carolina — MatthewChang · 2026-10-22
- DeepMind's Gemini Argon Compresses Turing Machines for Busy Beaver, Goldbach, Riemann in Lean 4 — robertarail · 2026-10-22
- Unitree unveils H2-X kung fu guardian robot with ~8h endurance that can lift a car — LinusEkenstam · 2026-10-22
- Sharpa's D01 humanoid: 21-DoF tactile hand, 0.2mm precision and a haptic data glove — arnie_hacker · 2026-10-22