EMBGuard: Constructing Hazard-Aware Guardrails for Safe Planning in Embodied Agents

MLLM-powered embodied agents deployed in real-world environments encounter physical hazards. However, existing approaches lack explicit mechanisms for identifying hazards and reasoning about action-conditioned risks, leading agents to either miss risky interactions or over-identify risks. To address this, we propose EMBGuard, the first MLLM-based safety guardrail for embodied agents designed to decouple physical risk reasoning from agent policy. By evaluating a (visual observation, action) pair, EMBGuard identifies hazardous configurations and provides natural language explanations of potential risks. Alongside EMBGuard, we contribute EMBHazard, a training dataset of 15.1K action-conditioned pairs, and EMBGuardTest, a benchmark of 329 manually curated real-world scenarios spanning seven physical risk categories. Through compositional variation of hazards and actions, we generate diverse risky and benign scenarios that agents may encounter during planning. Despite its compact size (2B, 4B), EMBGuard achieves performance competitive with proprietary MLLMs (e.g., GPT-5.1, Gemini-2.5-Pro) while significantly reducing the false-positive rates that hinder real-time deployment. We make the code, data, and models publicly available at https://github.com/dongwxxkchoi/EMBGuard.

AI assistants are starting to control robots and machines in our homes and workplaces. But unlike chatbots, these robots can cause real harm — knocking over a hot pot, gripping a sharp knife the wrong way, or pushing an object off a counter. Current systems either miss these dangers or become so cautious that they flag harmless actions as risky, making them impractical. We built EMBGuard, a separate safety checker that watches what a robot sees and what it is about to do, then decides whether that action is dangerous. Unlike previous approaches that bundle safety into the robot's decision-making, ours acts as an independent supervisor that explains why something is risky in plain language. We also created 15.1K training examples and a benchmark of 329 realistic scenarios covering seven types of physical hazards. Despite being small enough to run in real time, EMBGuard matches the safety judgment of much larger commercial AI systems while raising far fewer false alarms — a step toward robots that are both safe and practical.