SafetyALFRED
收藏资源简介:
SafetyALFRED是由密歇根大学团队基于ALFRED基准扩展的多模态安全评估数据集,聚焦厨房场景下的六类现实危害(如设备误用、火灾隐患等)。该数据集包含1001条轨迹数据,通过AI2-THOR环境生成,整合了视觉帧序列、物体元数据及PDDL定义的安全修复动作。其创新性在于首次将静态危害识别与动态具身规划相结合,用于评估大语言模型在完成家务任务时识别和主动消除安全隐患的能力,填补了现有安全基准仅支持问答式评估的空白。
SafetyALFRED is a multimodal safety evaluation dataset extended from the ALFRED benchmark by a research team at the University of Michigan. It focuses on six types of real-world hazards in kitchen scenarios, such as equipment misuse and fire hazards. This dataset comprises 1001 trajectory samples generated within the AI2-THOR simulation environment, integrating visual frame sequences, object metadata, and PDDL-defined safety remediation actions. Its core innovation is the first integration of static hazard identification and dynamic embodied planning, which is designed to evaluate the ability of Large Language Models (LLMs) to identify and proactively eliminate safety hazards while completing household tasks, filling the critical gap that existing safety benchmarks only support question-answering-based evaluations.

- 1SafetyALFRED: Evaluating Safety-Conscious Planning of Multimodal Large Language Models密歇根大学; 博伊西州立大学 · 2026年



