Abbeel lab's SuperMimic replaces physics sim with differentiable world model, ~1000x faster sim-to-real
HirokatuKataoka · x · 2026-11-12
Pieter Abbeel's UC Berkeley group presented SuperMimic in an oral at CoRL2026: replacing the physics simulator in humanoid whole-body control RL with a neural network-based world model, speeding up sim-to-real by up to 1000x while also improving policy quality. With everything differentiable (objective→reward→world model→policy), training uses ordinary gradient methods. It succeeds at acrobatic moves like backflips that NVIDIA's SONIC failed, and the team revealed it's already showing strong results on manipulation and multi-fingered hands, beating PPO baselines by a wide margin. Code was released alongside the talk; Nikkei Robotics covered it in a 10-page feature.
Related event: Berkeley Open-Sources SuperMimic, Speeding Sim-to-Real by Nearly 1000x(2 posts)→
More from Embodied
- Deployment is where physical AI grows up — and its most important data source, per Tedrake's CoRL talk — m_wulfmeier · 2026-11-12
- Runway teases triple bet: real-time avatars, interactive worlds, robot action models — c_valenzuelab · 2026-11-12
- WICL: robots learn manipulation in-context, 83.4% success on all 50 RoboTwin tasks — _akhaliq · 2026-11-12
- After CoRL, robotics feels far from solved: Gemini 4 Carbon evaluated — DJiafei · 2026-11-12
- Wired tries the OpenAI-powered Microduck, a tiny robot duck that chases cats — Wired AI · 2026-11-12
- RoboCap-10k released: 10,000 hours of open egocentric robot data with metric actions — databoydg · 2026-11-12