Title: RoboStressBench: Evaluating VLM Resilience to Physical Visual Stress in Embodied Contexts

Abstract:

Vision-Language Models (VLMs) have demonstrated impressive capabilities in visual comprehension and are increasingly integrated into embodied AI systems, where dependable perception under authentic conditions is critical. Nevertheless, current benchmarks typically evaluate VLMs using pristine imagery or isolated distortions, neglecting the stresses generated by the physical formation of scenes. This approach suffers from two primary drawbacks: it addresses only a limited fraction of common visual stressors, and certain introduced perturbations seldom occur in realistic embodied settings. This discrepancy prompts a core inquiry: how can visual stress be defined rigorously to encompass the varied factors present in physical environments? To resolve this, we analyze visual perception through the lens of inverse graphics and present RoboStressBench, a benchmark designed to assess VLM robustness against physical visual stress within embodied scenes. Drawing inspiration from the physical rendering equation, RoboStressBench breaks down visual stress into four physically based dimensions: Material (M), Viewpoint (V), Lighting (L), and Geometry (G). This structure allows RoboStressBench to encompass a wide spectrum of visual stresses found in real-world settings, while facilitating a controlled examination of their impact on VLM abilities such as planning, reasoning, and visual recognition. Comprehensive tests of leading VLMs uncover specific failure modes linked to stress, revealing that distinct physical factors impair different embodied capabilities, a nuance often hidden by overall accuracy metrics. Additionally, we propose a stress-aware agentic solver that identifies visual stressors and triggers visual-editing skills prior to reasoning, thereby enhancing robustness in high-stress situations. In summary, RoboStressBench offers a rigorous evaluation framework for diagnosing and enhancing VLM perception under real-world physical stress, aiding the creation of more dependable embodied AI systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC