From Pixels to Actions - Towards Native Vision-Language-Action Policies

Researchers propose NEO-VLA, an encoder-free vision-language-action framework that tokenizes images with lightweight patch convolutional layers and evolves them jointly with language throughout the backbone. The approach allows a flow-matching action expert to draw from instruction-conditioned visual representations at different depths, and native architectures reportedly outperform encoder-based counterparts under controlled comparisons. Layer-wise probing reportedly shows that action-critical visual signals often peak well before the final visual representation.