LabGuard Ensures Safety for Embodied AI Agents in Laboratories

Researchers have developed LabGuard, a comprehensive safety suite designed to ensure the secure operation of embodied AI agents in dynamic laboratory environments. The core challenge addressed is the conversion of human-readable safety rules, protocols, and manuals into machine-executable runtime constraints. LabGuard comprises three main components: LabGuard-IR, a typed executable representation for rules; LabGuard-Bench, a benchmark with 812 annotated laboratory rules; and LabGuard-Grounder, which maps natural language rules into the LabGuard-IR. This system then compiles these IR instances into runtime monitors that are applied at the agent's controller boundary. Experimental results demonstrate LabGuard's effectiveness, showing it generalizes to new rule sources and reduces unsafe events from 39.5% to 23.8%. When integrated with existing agent systems, it maintains task success while keeping interventions minimal, proving its capability to enhance safety without hindering performance.