Do vision models fail when robots face real lighting and clutter?

Vision-language models power robot perception, but existing benchmarks test them on clean images, not the messy real world. RoboStressBench decomposes real-world visual chaos into four physical factors—materials, viewpoint, lighting, and geometry—then tests how each breaks state-of-the-art VLMs at recognition and planning. Different stressors wreck different capabilities, and the authors show a stress-aware agent that detects and edits problematic images before reasoning significantly improves performance in harsh conditions.