NEO-VLA: encoder-free native VLA turns pixels into actions, beats encoder-based baselines
liuziwei7 · x · 2026-11-11
The fourth NEO-series work, NEO-VLA, proposes an encoder-free native Vision-Language-Action model going "from pixels to actions for System-1." Key claims: early visual features matter most for action; native representations transfer to action even when frozen; and native VLA outperforms encoder-based counterparts in robustness.
Related event: NEO-VLA: Encoder-Free Vision-Language-Action Model Released(2 posts)→
More from Embodied
- MBZUAI releases open world model PAN 2: LLM planning plus 29B diffusion hits 75.1% step accuracy — rohanpaul_ai · 2026-11-11
- ProtoMotions Deployment Stack: One Exported Graph, 1.3 ms C++ Loop for Real Robots — carlosdponx · 2026-11-11
- Moonshot's heymoonshot launches Luna, a voice-first AI wearable on Kickstarter — aftahi_ai · 2026-11-11
- Niantic Spatial adds object interaction in Gaussian Splats for embodied AI real2sim training — Scobleizer · 2026-11-11
- SpaceX unveils Starlink Mini 2 with built-in 71Wh battery, up to 4.5 hours runtime — Scobleizer · 2026-11-11
- VLK generates large-scale synthetic data for humanoid loco-manipulation without teleoperation — pabbeel · 2026-11-11