CFG lifts single-step discrete action success from 27.1% to 49.2% in robot policy training
YouJiacheng · x · 2026-10-21
In a thread on fixing factorization error in robot policy training, Zeji Yi explains classifier-free guidance (CFG): during training the observation is sometimes replaced with a blank input so the head also learns an overall guess; at test time CFG concentrates each token's belief on what the observation favors over that guess. Single-step discrete success jumps from 27.1% to 49.2% on RoboCasa with nearly unchanged action error (16.4→16.3mm) — CFG fixes factorization error, not precision.
You Jiacheng adds a clarification: this CFG (contrastive decoding) is not plain distribution sharpening — discrete regression already decodes via argmax; the sharpening is relative to the unconditional prior, akin to prior correction to uniform in imbalanced classification.
More from Embodied
- SF hard-tech crowd hosts real humanoid robot fight night — edgarpavlovsky · 2026-10-21
- REK robot fighting league footage goes viral: Terminator-style mechs battle in the cage — edgarpavlovsky · 2026-10-21
- Hot take: home automation wins by standardizing spaces, not building robots — alysha_lobo · 2026-10-21
- Menlo Research open-sources Cyclotron, an RL training framework for its Asimov humanoid robot — freelerobot · 2026-10-21
- NVIDIA demos interactive AI surgical simulator running on a single RTX PRO 6000 GPU — Scobleizer · 2026-10-21
- Cactus runs local speech-to-actions on a $10 ESP32 with sub-30MB models — alexcovo_eth · 2026-10-21