NEO-VLA ditches visual encoders for robot actions, hitting 90.9% on LIBERO frozen
机器之心 · wechat · 2026-11-15
Researchers from NTU S-Lab, SenseTime Research and Ropedia introduce NEO-VLA, a native vision-language-action model that removes the standalone visual encoder (CLIP/SigLIP/DINOv2) and lets the action expert read representations at multiple depths of a 2B-parameter native multimodal backbone.
Key findings from controlled experiments:
- With the backbone frozen, NEO-VLA reaches 90.9% on LIBERO, beating fully fine-tuned Qwen3-VL (79.5%) and Gemma-3 (84.5%), and 85.4% on LIBERO-Long.
- Ablations show shallow layers (0-10) are critical for action prediction: swapping their KV representations sharply increases translation error, while matched noise does not.
- Mixing multimodal understanding data into action post-training preserves language/vision skills and boosts RoboTwin2.0 Clean2Rand success to 69.4% (vs 10.7% with action data alone).
Pretraining aggregates nine public datasets totaling 2.23M trajectories (10k hours), unified into a 26-dim per-arm + 8-dim body action format with masking for heterogeneous robots. The full model hits 97.4% on LIBERO; on three real-robot fine-manipulation tasks (50 demos each) it matches or beats π0.5, and inference runs at 0.26s per 30-step action chunk on a single RTX 3090. Paper, code, model and data are open-sourced.
More from Embodied
- RaiSim vs Newton benchmark: CPU robot sim beats GPU by up to 95x at practical batch sizes — ChongZzZhang · 2026-11-15
- Researcher teaches a Unitree G1 humanoid to ride a mechanical bull after CoRL — RemiCadene · 2026-11-15
- PrimeSim: world model trained on 30k hours of robot demos runs near real-time on one A100 — jiqizhixin · 2026-11-15
- Viral claim: Claude runs a robot arm on a real factory task with zero training — but the model name doesn't exist — IgorCarron · 2026-11-15
- Zipline's Keller Cliffton on 3M drone deliveries and what breaks at 10x scale — a16z Podcast · 2026-11-15
- AI Tamagotchi: CMU student's E Ink gadgets turn objects into needy pets — Wired AI · 2026-11-15