MAPS：通过模块级邻近调度保持视觉-语言表征以提升视觉-语言-动作泛化能力
MAPS: Preserving Vision-Language Representations via Module-Wise Proximity Scheduling for Better Vision-Language-Action Generalization
对做 VLA/机器人策略学习的人，这篇给出了一个无需额外参数、可直接插入现有 VLA 的微调调度策略：按模块（视觉编码器 vs 动作导向语言层）差异化地放松对预训练 VLM 的邻近约束，从而在适配下游操作任务的同时保住 OOD 泛化，可直接复用到 OpenVLA 类模型的微调流程中。
原文依据 MAPS linearly schedules this relaxation, enabling visual encoders to stay close to their pretrained priors while action-oriented language layers adapt more freely.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论视觉-语言-动作（VLA）模型继承了预训练视觉-语言模型（VLM）的强大先验，但朴素微调往往会破坏这些表征并损害泛化能力。现有补救措施——冻结模块或施加统一正则化——要么过度约束适配过程，要么忽视了 VLA 各组成部分的不同作用。我们提出 MAPS（模块级邻近调度），这是首个面向 VLA 的鲁棒微调框架。通过系统性分析，我们发现了应当以何种经验顺序放松邻近约束以平衡稳定性与灵活性的规律。MAPS 对该放松过程进行线性调度，使视觉编码器保持接近其预训练先验，而面向动作的语言层则能更自由地适配。MAPS 不引入额外参数或数据，可无缝集成到现有 VLA 中。在 MiniVLA-VQ、MiniVLA-OFT、OpenVLA-OFT 以及 SimplerEnv、CALVIN、LIBERO 等具有挑战性的基准上，以及在 Franka Emika Panda 平台上的真实世界评测中，MAPS 一致地提升了分布内与分布外性能（最高提升 30%）。我们的发现凸显出：以经验为指导地保持与预训练 VLM 的邻近性，是在 VLM 到 VLA 迁移中保持泛化能力的一条简单而有力的原则。
英文原文摘要 1291 字符
Vision-Language-Action (VLA) models inherit strong priors from pretrained Vision-Language Models (VLMs), but naive fine-tuning often disrupts these representations and harms generalization. Existing fixes -- freezing modules or applying uniform regularization -- either overconstrain adaptation or ignore the differing roles of VLA components. We present MAPS (Module-Wise Proximity Scheduling), the first robust fine-tuning framework for VLAs. Through systematic analysis, we uncover an empirical order in which proximity constraints should be relaxed to balance stability and flexibility. MAPS linearly schedules this relaxation, enabling visual encoders to stay close to their pretrained priors while action-oriented language layers adapt more freely. MAPS introduces no additional parameters or data, and can be seamlessly integrated into existing VLAs. Across MiniVLA-VQ, MiniVLA-OFT, OpenVLA-OFT, and challenging benchmarks such as SimplerEnv, CALVIN, LIBERO, as well as real-world evaluations on the Franka Emika Panda platform, MAPS consistently boosts both in-distribution and out-of-distribution performance (up to +30%). Our findings highlight empirically guided proximity to pretrained VLMs as a simple yet powerful principle for preserving generalization in VLM-to-VLA transfer.