Can AI perceive physical hazards and intervene?
Can AI Perceive Physical Danger and Intervene?
The researchers built an embodied AI physical safety benchmark based on real injury narratives and operational safety constraints, and used advanced generative models to turn transitions from safe to unsafe states into realistic images and videos, in order to probe multimodal safety understanding. The paper systematically analyzes the ability of mainstream foundation models to perceive risk, reason about safety, and trigger intervention, providing multifaceted insights into the deployment readiness of safety-critical agent applications. The authors also propose a post-training paradigm that teaches models to explicitly reason about the embodiment-specific safety constraints given in system instructions; the resulting models generate interpretable, transparent safety reasoning traces and achieve SOTA in constraint-satisfaction evaluations. The benchmark is released at https://asimov-benchmark.github.io/v2。