SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of LLM-based Embodied Agents

Simon Sinong Zhan, Yao Liu, Philip Wang, Zinan Wang, Qineng Wang, Zhian Ruan, Xiangyu Shi, Xinyu Cao, Frank Yang, Kangrui Wang, Huajie Shao, Manling Li, Qi Zhu

Keywords: Safety in LLM-based Embodied Agent

Abstract: We present $\texttt{SENTINEL}$, a unified multi-level framework for evaluating the physical safety of LLM embodied agents using $\textit{formal safety semantics}$. In our approach, safety rules are grounded as temporal logic constraints, providing precise semantics for specifying state invariants, temporal dependencies, and timing requirements. These rules enable formal checking of embodied-agent behavior at multiple stages of decision-making. $\texttt{SENTINEL}$ is organized into a progressive evaluation pipeline: at the $\textit{semantic level}$, natural language safety requirements are interpreted as Temporal Logic (TL) specifications; at the $\textit{planning level}$, high-level action programs and subgoals are checked against these TL rules before execution; and at the $\textit{trajectory level}$, multiple simulated executions are merged into planning trees and verified against more physical-detailed Computation Tree Logic (CTL) specifications. This provides a reproducible protocol for jointly measuring task completion and safety compliance. By grounding safety in temporal logic and enabling formal evaluation across semantics, plans, and trajectories, $\texttt{SENTINEL}$ establishes a comprehensive pipeline for systematically assessing LLM-based embodied-agent safety, laying the foundation for agents that are not only capable but also reliably safe in realistic environments.

Primary Area: datasets and benchmarks

Submission Number: 3494