Towards Safer Mobile Agents: Scalable Generation and Evaluation of Diverse Scenarios for VLMs

TL;DR

This work addresses the scarcity of hazardous, temporally dynamic evaluation data for Vision-Language Models (VLMs) in mobility contexts. It introduces HazardForge, a scalable image-editing–based pipeline with layout-decision and validation modules to synthesize moving, intrusive, and distant hazards, and MovSafeBench, a 7,254-image MCQ benchmark across four scenarios and 13 object categories. Evaluation of seven VLMs reveals significant performance drops when handling anomalous objects and motion, with motion scenarios being the most challenging and no-edit baselines outperforming edited variants in some cases. The approach provides a practical, scalable framework for robust safety assessment of VLM-enabled mobile agents, highlighting concrete gaps and guiding future improvements.

Abstract

Vision Language Models (VLMs) are increasingly deployed in autonomous vehicles and mobile systems, making it crucial to evaluate their ability to support safer decision-making in complex environments. However, existing benchmarks inadequately cover diverse hazardous situations, especially anomalous scenarios with spatio-temporal dynamics. While image editing models are a promising means to synthesize such hazards, it remains challenging to generate well-formulated scenarios that include moving, intrusive, and distant objects frequently observed in the real world. To address this gap, we introduce \textbf{HazardForge}, a scalable pipeline that leverages image editing models to generate these scenarios with layout decision algorithms, and validation modules. Using HazardForge, we construct \textbf{MovSafeBench}, a multiple-choice question (MCQ) benchmark comprising 7,254 images and corresponding QA pairs across 13 object categories, covering both normal and anomalous objects. Experiments using MovSafeBench show that VLM performance degrades notably under conditions including anomalous objects, with the largest drop in scenarios requiring nuanced motion understanding.

Figures (7)

• Figure 1: Overview of MovSafeBench. Our benchmark assesses how reliably Vision-Language Models (VLMs) operate and how well they visually ground information when used in mobile agents, including vehicles and mobile robots.
• Figure 2: Overall workflow of our HazardForge. HazardForge generates the scene, including moving objects, intrusive objects, and remote objects.
• Figure 3: The detailed workflow of our intrusion image editing.
• Figure 4: The detailed workflow of our distance image editing.
• Figure 5: The detailed workflow of our quality check module.