SafeMind: Benchmarking and Mitigating Safety Risks in Embodied LLM Agents

TL;DR

SafeMind targets safety in embodied LLM agents by formalizing a four‑stage reasoning pipeline and three orthogonal constraint types (Factual, Causal, Temporal). It introduces SafeMindBench, a multimodal benchmark of 5,558 instruction–image samples across Instr‑Risk, Env‑Risk, Order‑Fix, and Req‑Align tasks to systematically evaluate hazards and task completion. To mitigate risks, SafeMindAgent extends the Planner–Executor architecture with three cascaded safety modules and an external Safety Constraint Knowledge Base, coupled with retrieval and reflection–correction mechanisms. Empirical results show substantial safety improvements over strong baselines while preserving competitive task performance, highlighting the approach’s potential for safer real‑world deployment of embodied LLM agents. The work provides a unified risk taxonomy, a rigorous evaluation suite, and a practical mitigation strategy that can generalize to broader embodied‑AI safety challenges.

Abstract

Embodied agents powered by large language models (LLMs) inherit advanced planning capabilities; however, their direct interaction with the physical world exposes them to safety vulnerabilities. In this work, we identify four key reasoning stages where hazards may arise: Task Understanding, Environment Perception, High-Level Plan Generation, and Low-Level Action Generation. We further formalize three orthogonal safety constraint types (Factual, Causal, and Temporal) to systematically characterize potential safety violations. Building on this risk model, we present SafeMindBench, a multimodal benchmark with 5,558 samples spanning four task categories (Instr-Risk, Env-Risk, Order-Fix, Req-Align) across high-risk scenarios such as sabotage, harm, privacy, and illegal behavior. Extensive experiments on SafeMindBench reveal that leading LLMs (e.g., GPT-4o) and widely used embodied agents remain susceptible to safety-critical failures. To address this challenge, we introduce SafeMindAgent, a modular Planner-Executor architecture integrated with three cascaded safety modules, which incorporate safety constraints into the reasoning process. Results show that SafeMindAgent significantly improves safety rate over strong baselines while maintaining comparable task completion. Together, SafeMindBench and SafeMindAgent provide both a rigorous evaluation suite and a practical solution that advance the systematic study and mitigation of safety risks in embodied LLM agents.

Figures (5)

• Figure 1: Overview of the risk model and task categories in SafeMindBench.
• Figure 2: Data generation pipeline of SafeMindBench.
• Figure 3: (a) Composition of SafeMindBench; (b) Decision-making process of SafeMindAgent.
• Figure 4: (a) Safety performance of agents across four risk tasks; (b) Safety analysis of Req-Align task across constraint categories; (c) Ablation study on the safety modules of SafeMindAgent.
• Figure 5: Two-Dimensional Distribution of Safety Constraints.