NEO-VLA: 2B native VLM on raw pixels hits 95.3 avg on LIBERO, beating Qwen3-VL and Gemma-3
liuziwei7 · x · 2026-11-06
NEO-VLA feeds raw pixels via a tiny conv patchifier into a 2B native VLM (NEO-2B, 40 layers) plus a flow-matching action expert with shared t-RoPE offsets (26D arm + 8D body actions). On LIBERO (30k-step checkpoint, 2000 episodes) it reaches 95.3 average success with full fine-tuning, beating 4B Gemma-3 (91.0) and 2B Qwen3-VL (90.8); frozen it already scores 90.9. Layer-wise results suggest the backbone's early layers act like a vision encoder — raw-pixel native input can beat VLM + vision encoder setups.
Related event: NEO-VLA: Native Vision-Language-Action Policy Without a Vision Encoder(3 posts)→
More from Embodied
- Harvard-led team's MFM memory system lets frozen-weight agents keep learning in Minecraft — du_yilun · 2026-11-06
- RAX lets coding agents unscrew a real jar with dexterous hands: 6 chained skills, 140° rotation — du_yilun · 2026-11-06
- HeLM: explicit language memory gives VLA robots +12.8 points on non-Markovian manipulation — SeoulNationalUniv · 2026-11-06
- 3 Years and Tens of Thousands of Lines of Code vs. One Prompt: AI Scene Generation's Leap — liliyu_lili · 2026-11-06
- UWaterloo 3D prints Invar36 flexures with directly mounted interferometer optics — jwt0625 · 2026-11-06
- LeRobot v0.7.0 ships as ecosystem passes 50,000 developers and 300+ organizations — LysandreJik · 2026-11-06