HomeSafe-Bench: Evaluating VLMs on Unsafe Action Detection for Embodied Agents

arXiv 2603.11975·low signalXBlueskyLinkedInCopy link

First benchmark specifically targeting dynamic unsafe action detection for embodied agents in household settings, exposing that static-image safety evals fail to capture temporal risk patterns in real-world robot deployment. Tests VLMs on perception-latency and common-sense safety failures across household scenarios. As agentic systems extend into physical environments and autonomous execution, this benchmark gap is immediately relevant for teams shipping embodied or computer-use agents.