NEO-VLA drops the vision encoder: native VLA pretraining pays off on RoboTwin
liuziwei7 · x · 2026-11-06
- LMMs-Lab (NTU S-Lab, SenseTime Research, Ropedia; authors include Dahua Lin and Ziwei Liu) releases NEO-VLA, a native vision-language-action policy, with paper, code, weights, and dataset all open.
- It removes the pretrained vision encoder: pixels go through a lightweight conv patch embedder into a native VL backbone, and a flow-matching action expert reads from chosen backbone layers.
- Finding 1: action-relevant cues (geometry, edges, depth) are organized across backbone depth. Finding 2: gains come from native pretraining, not encoder access — even giving the expert direct access to Qwen's vision tower doesn't replicate it. Finding 3: understanding data matters during post-training.
- Real-robot experiments use three tasks with 50 demonstrations each; strong results on LIBERO variants and RoboTwin C2R.
Related event: NEO-VLA: Native Vision-Language-Action Policy Without Vision Encoder(2 posts)→
More from Embodied
- Policy trained entirely in imagination matches PPO with 7.9× fewer simulator transitions — EpisodeYang · 2026-11-06
- Apple's mystery robot reportedly uses a Mac Mini as its planner brain — micoolcho · 2026-11-06
- Microdock v2 lets robot sit while docked, cutting servo power to zero — natesiggard · 2026-11-06
- Building an Asimov Robot in 22 Hours — and It Worked Out of the Box — whurley · 2026-11-06
- Gold Tesla Cybercab and Optimus shown at UP Summit — DMaguireARK · 2026-11-06
- NVIDIA launches SimReady Central with 3,500+ OpenUSD robot assets physically calibrated against real hardware — CyberRobooo · 2026-11-06