Physical Intelligence's RLT uses RL tokens to make robot fine manipulation 3x faster
svlevine · x · 2026-11-05
Physical Intelligence (π), with Sergey Levine among the authors, introduced RLT (RL tokens), a method for fine-tuning VLA robots with online RL on precision tasks.
- Approach: add a special output token to the VLA as a compact interface to a lightweight RL policy, adapting behavior from just hours of real-world experience without fine-tuning the full model
- Results: across four challenging manipulation tasks, RLT speeds up the most precise stages by up to 3x, surpassing human teleoperation speed
- Related GSBR work: a new behavior regularizer combining support constraints with continuous penalties to balance exploration against staying near known-good data in offline-to-online RL
- Motivation: broad competence isn't enough — aligning a screwdriver with a tiny screw quickly requires RL refinement beyond demonstrations
More from Embodied
- Peking University and NVIDIA use a coding agent to replace teleoperation for robot data — DAGroup-PKU · 2026-11-05
- Corebrum OS adds built-in VNC screen share, demoed remote-viewing an NVIDIA Thor robot from a Mac — chrismatthieu · 2026-11-05
- NerF Author Ben Mildenhall Teases Mystery AI Gadget With Robot Icon and 8:00 Display — XRarchitect · 2026-11-05
- KAIST's H2RB transfers human demonstrations to robots, boosting object transfer from 8.0 to 51.8 — kaist-ai · 2026-11-05
- New AI glasses UI concept demo shows off futuristic interface design — AnhPhuNguyen1 · 2026-11-05
- Berkeley project uses pretrained base policy's support to guide exploration for fast online adaptation — abhishekunique7 · 2026-11-05