NEO-VLA: Native Vision-Language-Action Framework Outperforms Encoders
Original title: From Pixels to Actions - Towards Native Vision-Language-Action Policies
Native vision-language-action models achieve stronger robotic control by evolving visual representations jointly with language.
The paper presents NEO-VLA, which tokenizes images via lightweight patch convolutions and integrates them directly with language in the backbone for action policies. This avoids fixed visual encoders and allows probing at different depths. Experiments across manipulation tasks show native architectures outperform encoder-based methods, with action-critical signals peaking early in the hierarchy.
- Action policies benefit from evolving visual representations across depths
- Full fine-tuning reduces but does not eliminate performance gap
- Consistent superiority in diverse manipulation settings and shifts
- AI models
- NEO-VLA
From the original article
1 min read at huggingface.co
Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e. g.
By Weichen Fan, Haiwen Diao, Penghao Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.