Stanford and Berkeley unveil Flow Reversal Steering, turning VLM hints into dexterous robot actions with up to 95-point success gains
svlevine · x · 2026-11-09
Researchers from Stanford and UC Berkeley (including Chelsea Finn and Sergey Levine) will present Flow Reversal Steering (FRS) at CoRL, tackling the gap between VLM semantic understanding and dexterous robot action.
- Core idea: take coarse but reasonable reference actions, pass them through a flow-matching generalist policy in reverse to find their latent noises, then denoise to map onto nearby high-quality action modes — converting semantic hints from humans or VLMs into precise manipulation.
- Distillation: an auxiliary policy trained via behavioral cloning to output those noises yields up to 95-point absolute task success rate boosts with under a minute of training.
- RL improvement: FRS bootstraps RL with semantic knowledge, improving tasks where standard RL fails.
Co-first authors Andy Tang and William Chen will present at poster 39 on Tuesday; code is coming soon.
More from Embodied
- Tesla Robotaxi expands to San Antonio starting tomorrow at 5am — mitchdeg · 2026-11-09
- Open-source robot manipulation stack Specialist: hardware, data, training, deployment all released — TinfoilTricorn · 2026-11-09
- RAI Institute's two-wheeled UMV robot does backflips with just 5 degrees of freedom — Scobleizer · 2026-11-09
- Team hacks robot arms to 85 points, wins MakerMods Physical AI Hackathon — IsaiahBallah · 2026-11-09
- Tesla Optimus engineer builds GROXIE, a countertop robot assistant powered by Grok — Scobleizer · 2026-11-09
- RLT paper wins CoRL oral: sample-efficient real-world RL to fine-tune VLA models — svlevine · 2026-11-09