From Pixels to Actions: Advancing Native Vision‑Language‑Action Policies
reasoning
| Source: HF Papers | Original article
Researchers propose native vision-language-action policies that integrate perception and action without first compressing visual data, preserving fine‑grained cues such as geometry and boundaries.
A new research paper titled **“From Pixels to Actions – Towards Native Vision‑Language‑Action Policies”** proposes a fundamentally different architecture for vision‑language‑action (VLA) models. Rather than compressing raw images through a heavyweight visual encoder before any language or planning steps, the authors introduce **NEO‑VLA**, an encoder‑free policy built directly on the native NEO‑2B backbone. In this design, pixel and word representations are aligned in a shared semantic space from the outset, and the shallow layers of the backbone retain the fine‑grained visual cues—local geometry, object boundaries, depth and subtle state changes—that traditional pipelines tend to discard.
The shift matters because many manipulation tasks hinge on those precise visual signals. By keeping action‑relevant information close to the input, NEO‑VLA promises more accurate and efficient reasoning for robotic control, interactive agents and other embodied AI systems. The approach also streamlines the model stack, eliminating the need for separate vision and language modules and potentially reducing inference latency.
The paper, authored by researchers from S‑Lab (NTU), SenseTime Research and Ropedia, is slated for release soon. The community will be watching for benchmark results that compare NEO‑VLA against existing VLA systems, especially on tasks that demand high‑resolution spatial understanding. Follow‑up work is likely to explore scaling the native pre‑training regime, extending the method to heterogeneous datasets, and open‑sourcing the codebase. As we noted in our coverage of LongCat‑2.5’s native multimodality (2026‑11‑02), the move toward unified, encoder‑free models is gaining momentum, and NEO‑VLA could become a cornerstone of the next generation of embodied AI.
Sources
Back to AIPULSEN