论文/arXiv 2610.09144·22 天前

DIVA：用于视觉-语言-动作策略的双空间意图感知视觉衰减

DIVA通过意图感知的双空间视觉衰减调控视觉token影响，并在LIBERO和真实世界实验中验证提升。

AI 整理

• 仿真与训练
• 具身模型

原文链接：DIVA: Dual-Space Intent-Aware Visual Attenuation for Vision-Language-Action Policies（arXiv 2610.09144 · 发布日期：2026-10-06）

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

DIVA通过意图感知的双空间视觉衰减调控视觉token影响，并在LIBERO和真实世界实验中验证提升。

核心要点

• VLA策略通常输入密集视觉patch token，但缺少显式机制调节不同视觉token对策略计算的影响。
• DIVA采用anchor-then-attenuate设计，结合高层任务意图与低层视觉证据估计patch级相关性锚点。
• 该模块在骨干入口前重加权投影视觉token，并在骨干内持续衰减低相关视觉状态，同时保留完整视觉序列。
• 在LIBERO上，DIVA将OpenVLA-OFT平均成功率从96.6%提升到98.0%，零样本LIBERO-Plus从69.6提升到72.6。

验证范围与局限：验证范围包括LIBERO仿真评测，以及存在任务无关视觉扰动的真实世界实验。

英文摘要 (Abstract)

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. We introduce DIVA, a Dual-Space Intent-Aware Visual Attenuation module with an anchor-then-attenuate design. 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. DIVA preserves the full visual token sequence and requires no external grounding supervision. On LIBERO, DIVA improves OpenVLA-OFT from 96.6% to 98.0% average success and raises its zero-shot LIBERO-Plus score from 69.6 to 72.6. Real-world experiments further show consistent gains under task-irrelevant visual perturbations, supporting the robustness of intent-aware visual attenuation beyond simulation.

作者：Kaixi Feng, Guoheng Sun, Ziyao Wang, Yexiao He, Zheyu Shen, Ang Li

来源

本页事实对应的公开页面，以原文为准。

1. arXiv 2610.09144：DIVA: Dual-Space Intent-Aware Visual Attenuation for Vision-Language-Action Policies访问 2026-10-21