Berkeley researchers sample humanoid robot policies directly from weight space with diffusion
berkeley_ai · x · 2026-10-16
UC Berkeley researchers (Huixuan Cheng, Bike Zhang, Shankar Sastry, Koushil Sreenath) ask whether robot control networks can be generated like images — and the answer is yes.
- Method: a latent diffusion framework pairs a linear autoencoder that compresses actor network parameters with a Diffusion Transformer modeling the latent distribution, trained on thousands of RL policy networks for the Unitree G1.
- Results: with no RL run, no reward, and no fine-tuning, sampled networks deploy on a real Unitree G1 and walk, dance, and spin-kick.
- Finding: first-layer weights consistently attend most to hips and waist regardless of motion — an embodiment-first kinematic prior that emerges under both RL and imitation learning, suggesting it is not training-paradigm-specific.
More from Embodied
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Tesla's hover car isn't a flying car — compressed air can't sustain flight — tinyfool · 2026-10-16
- Scobleizer relays claim that Optimus demo will be even more mind-blowing than the road-flipping Roadster — Scobleizer · 2026-10-16
- Indie dev's AI PASSPORT wearable firmware hits 1,200+ downloads in days — ezshine · 2026-10-16
- 1X Opens California Factory for Mass Production of NEO Humanoid Robots — emmanuelvivier · 2026-10-16
- Waymo Fully Opens Robotaxi Service in San Antonio and Two Other Cities — emmanuelvivier · 2026-10-16