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

SENTINEL is the first framework to provide multi-level safety evaluation of foundation model-based embodied agents — across semantic interpretation, plan generation, and physical execution — within a unified formal framework. Instead of heuristic rules or subjective FM judgments, it grounds practical safety requirements in temporal logic semantics that precisely specify state invariants, temporal dependencies, and timing constraints, and applies the pipeline to agents in VirtualHome and AI2-THOR against diverse safety requirements.

Authors: Simon Sinong Zhan, Philip Wang, Justin Liu, Yiyan Peng, Yiqi Lyu, Zinan Wang, Qineng Wang, Zhian Ruan, Xiangyu Shi, Xinyu Cao, Frank Yang, Zhenyang Ni, Kangrui Wang, Ruohan Zhang, Huajie Shao, Manling Li, Qi Zhu (*equal contribution)

@inproceedings{zhan2026sentinel, title={SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents}, author={Zhan, Simon Sinong and Wang, Philip and Liu, Justin and Peng, Yiyan and Lyu, Yiqi and Wang, Zinan and Wang, Qineng and Ruan, Zhian and Shi, Xiangyu and Cao, Xinyu and Yang, Frank and Ni, Zhenyang and Wang, Kangrui and Zhang, Ruohan and Shao, Huajie and Li, Manling and Zhu, Qi}, booktitle={Advances in Neural Information Processing Systems (NeurIPS)}, year={2026}, url={https://arxiv.org/abs/2510.12985} }