# OSMA-Bench评测揭示VLM/VLA家庭导航语义安全到运动控制映射失效

> 原文链接：https://robotics.hellobike.com/papers/2026-10-26-2610-34674-4e6d8be2
> 频道：papers | 更新日期：2026-10-26 | 标签：auto, arxiv, hf-daily-papers
> 核心摘要：研究发布OSMA-Bench基准与SSI指标，测试发现VLM及VLA策略虽能识别障碍物语义，却无法将其转化为匹配的避障速度与距离参数。增加显式参数推理步骤可改善VLM策略的运动适应性，暴露了语义理解与底层运动控制间的接地鸿沟。

# OSMA-Bench评测揭示VLM/VLA家庭导航语义安全到运动控制映射失效

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation](https://arxiv.org/abs/2610.34674)（HF Daily Papers · arXiv 2610.34674）

## 核心内容

针对家庭机器人导航中过度保守或忽视语义安全的问题，研究团队构建了OSMA-Bench基准。该基准包含空间匹配但障碍物类型不同的场景集，并提出语义敏感度指数（SSI）量化评估模型是否根据障碍物类别调整 clearance、速度及平滑度。评测结果显示，现有的VLM和VLA策略即便在显式安全提示下，仍无法稳定地将语义安全等级映射为连续运动参数，表明当前具身智能在“认知-行动”闭环中存在显著断层。

技术分析指出，现有策略虽具备正确的障碍物语义分类能力，但在生成动作时丢失了安全约束信息。对于采用“识别-编程”范式的VLM策略，研究验证了在语义理解与代码生成之间插入显式参数推理步骤，能有效提升运动参数对障碍物安全属性的响应精度。这一发现证实了单纯依赖端到端大模型难以保证物理安全，需在架构层面强化从高层语义到低层动力学参数的结构化转译机制。

## 关键看点

- **OSMA-Bench基准**：构建空间匹配场景集与SSI指标，量化测试导航策略对障碍物语义的安全响应一致性
- **语义运动脱节**：VLM与VLA策略能正确分类障碍物，但生成的速度与距离参数未体现安全优先级差异
- **架构优化验证**：在VLM识别与导航程序生成间增加显式参数推理步骤，有效改善了运动适配性能
- **安全接地缺口**：揭示了当前具身模型将语义谨慎转化为物理安全运动的系统性能力不足

## 工程与产业观察

该研究量化了VLM/VLA在家庭导航中“懂安全但做不对”的工程痛点，证明仅靠Prompt无法解决语义到运动的接地问题。开发者应在感知与规划间引入结构化参数推理模块，或将安全约束作为独立损失函数纳入VLA训练，而非盲目信任端到端模型的隐式对齐能力，这对提升服务机器人实际部署安全性具有关键参考价值。

## 英文摘要 (Abstract)

Safe household navigation requires robots to regulate clearance and speed according to what an obstacle is and how close it is, without unnecessary conservatism. Yet existing evaluations do not test whether semantic safety distinctions are reflected in continuous motion. To that end, we introduce OSMA-Bench, an Obstacle-Conditioned Safe Motion Adaptation Benchmark built on spatially matched scenario sets that differ only in obstacle type. We also propose the Semantic Sensitivity Index (SSI), which measures whether clearance and distance-dependent speed and smoothness follow the caution ordering of obstacles. Evaluating VLM-based and vision-language-action (VLA) policies on the benchmark, we find that neither consistently adapts its motion to obstacle safety, even with explicit safety prompts. Our analysis shows that the policies can correctly distinguish obstacles by safety relevance but fail to reflect this distinction in the motion parameters or actions they produce. For VLM-based policies, which recognize the obstacle and then generate a navigation program, adding an explicit parameter-inference step between these two stages improves adaptation. These findings expose a gap in semantic-to-motion grounding, highlighting the need to translate semantic caution into appropriately safe motion.

Hugging Face Daily Papers 热度：4 upvotes
作者：hyunwoong Bae, Jisu Han, Jiyeon Choi, Dohoon Kim, Taesup Moon
:::

## 原始参考来源 (Sources)

- [HF Daily Papers · arXiv 2610.34674：Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation](https://arxiv.org/abs/2610.34674)