New study reveals VLMs confuse anomalies with real hazards in safety tasks
VLMs mistake unusual scenes for dangers—here's why that's a problem for robotics
A new paper by Indukuri et al., accepted to RO-MAN 2026, tackles a critical blind spot in Vision-Language Models (VLMs) used for safety-critical applications like disaster response and human-robot interaction. Current evaluations frame danger as a binary Safe/Unsafe decision, but the researchers argue this obscures whether a model detects a real physical hazard or simply reacts to an unusual scene element (an anomaly). They introduce an explicit distinction between hazard and anomaly and evaluate several state-of-the-art VLMs across two datasets with multiple prompting strategies.
The results reveal a troubling pattern: VLMs frequently misinterpret anomalousness as hazardousness, showing an over-reliance on contextual irregularity as a proxy for danger. By separating anomaly from hazard, the study provides a more informative assessment of VLM safety reasoning and exposes failure modes that binary judgments hide. The team also released their dataset on Roboflow to support further research. This work is a timely caution for deploying VLMs in emergency robotics, where confusing a harmless anomaly with a real threat could lead to disastrous decisions.
- VLMs over-rely on contextual irregularity, confusing anomaly with hazard instead of identifying true physical dangers.
- Binary Safe/Unsafe evaluations mask failure modes; separating hazard and anomaly reveals richer insights into model behavior.
- Public dataset released on Roboflow enables further research on VLM safety reasoning in emergency scenarios.
Why It Matters
Critical for deploying VLMs in emergency robotics where misjudging danger could have fatal consequences.