AI summaryⓘ
The authors created a new test called EgoSafetyBench to see how well vision-language models (VLMs) can act as safety guards for robots by watching video scenes from the robot’s perspective. Their test has two parts: one checks if the models can spot real dangers versus normal but unusual situations, and the other tests if they get confused by misleading signs or labels in the scene. They found that while the models often notice hazards in general, they miss some specific risky moments, especially when understanding context is needed. Also, misleading signs make models either miss dangers or raise too many false alarms. The authors highlight that just reacting a lot isn’t true understanding of safety risks.
Vision-language modelsEgocentric videoEmbodied agentsSafety benchmarksContextual hazardsIn-scene textContrastive evaluationRuntime safety guardFalse alarmsPhysical reasoning
Authors
Siddhant Panpatil, Arth Singh, Mijin Koo, Chaeyun Kim, Haon Park, Dasol Choi
Abstract
Vision-language models (VLMs) are now proposed as runtime safety guards for embodied agents in homes and factories. A deployable guard must catch genuinely unsafe situations while avoiding unnecessary intervention on routine but superficially alarming activity, a distinction that binary safety benchmarks obscure. We introduce EgoSafetyBench, an egocentric video benchmark of 1,200 robot-view scenarios annotated at half-second granularity, to evaluate VLMs as streaming guards across two tracks. The situational track (800 scenarios) spans four families, from routine and safe-but-suspicious scenes to obvious and contextual hazards. The visual-channel track (400 scenarios) targets in-scene text-a sign, sticker, or label visible in the scene-that can misrepresent the physical situation, pairing each misleading sign with a truthful version to test both whether a guard flags the text as misleading and whether the text corrupts its physical-safety judgment. Both tracks use contrastive ladders: near-identical scenarios differing only in a single visible deciding cue, so a correct call must hinge on that cue rather than the overall scene type. We evaluate ten open- and closed-source VLMs. We find that while guards reliably recognize videos containing hazards, they often miss specific hazardous moments, particularly contextual hazards. Furthermore, misleading in-scene signs degrade all tested guards: vulnerable models miss up to a third of hazards, while robust models over-intervene on safe content. Matched controls reveal that apparent safety robustness often reflects indiscriminate alarming rather than true physical reasoning.