NEO-VLA paper: native visual beats encoder features for robot action, data cleaned from 2.2M episodes
liuziwei7 · x · 2026-11-11
NEO-VLA's paper, GitHub and project page are public. The HF paper page argues a common VLA assumption is wrong: action control benefits from early visual computation, not semantic-distilled encoder features. Native representations transfer to action even frozen; exposing intermediate features in encoder-based VLMs yields no comparable gains. The 40-layer NEO-2B backbone drives a Franka Panda arm (26D arm + 8D body) with an action expert and 30-step action blocks, plus a 5-check data-cleaning pass over 2,231,586 episodes.
Related event: NEO-VLA: Encoder-Free Vision-Language-Action Model Released(2 posts)→
More from Embodied
- "No humanoid robot is being built by humanoids" — Matthew Chang's paradox — MatthewChang · 2026-11-11
- Unitree's hardware-only bet debated: control R&D, wait for software to catch up — teortaxesTex · 2026-11-11
- OpenTac: JHU/MIT open-source flex-PCB tactile skins that sensorize any robot surface — _krishna_murthy · 2026-11-11
- Chinese team open-sources a 15B-parameter humanoid robot foundation model — lmoroney · 2026-11-11
- Dyna Robotics' Taku performs long-horizon whole-body manipulation zero-shot, no on-site training — JasonMa2020 · 2026-11-11
- Humanoid robot hits roller skates in Matt Wolfe's latest AI news roundup — Matt Wolfe · 2026-11-11