DIVA: Dual-Space Intent-Aware Visual Attenuation for Vision-Language-Action Policies

Key points

• Vision-language-action (VLA) policies typically feed dense visual patch tokens into a language-action backbone, preserving scene context but offering no explicit mechanism to regulate how strongly different visual tokens influence policy computation.1
• DIVA combines high-level task intent with low-level visual evidence to estimate patch-wise relevance anchors, then applies them in two complementary spaces: it reweights projected visual tokens before backbone entry and persistently attenuates low-relevance visual states within the backbone.1
• DIVA preserves the full visual token sequence and requires no external grounding supervision.1
• Real-world experiments further show consistent gains under task-irrelevant visual perturbations, supporting the robustness of intent-aware visual attenuation beyond simulation.1

Sources (1)

• [1]DIVA: Dual-Space Intent-Aware Visual Attenuation for Vision-Language-Action Policies
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 09:37 PM

  We introduce DIVA, a Dual-Space Intent-Aware Visual Attenuation module with an anchor-then-attenuate design.

  Vision-language-action (VLA) policies typically feed dense visual patch tokens into a language-action backbone, preserving scene context but offering no explicit mechanism to regulate how strongly different visual tokens influence policy computation.

Extractive summary: sentences quoted from the sources.