HAZARD Challenge: Embodied Decision Making in Dynamically Changing Environments

• DOI: 10.48550/arxiv.2401.12975
• Publication Date: 2024-01-01
• Authors: Qinhong Zhou; Sunli Chen; Yisong Wang; Haozhe Xu; Weihua Du; Hongxin Zhang; Yilun Du; Joshua B. Tenenbaum; Chuang Gan
• Keywords: Embodied agent; Benchmark (surveying); Autonomous agent

Recent advances in high-fidelity virtual environments serve as one of the major driving forces for building intelligent embodied agents to perceive, reason and interact with physical world. Typically, these remain unchanged unless them. However, real-world scenarios, might also face dynamically changing characterized by unexpected events need rapidly take action accordingly. To remedy this gap, we propose a new simulated benchmark, called HAZARD, specifically designed assess decision-making abilities dynamic situations. HAZARD consists three disaster including fire, flood, wind, supports utilization large language models (LLMs) assist common sense reasoning decision-making. This benchmark enables us evaluate autonomous agents' capabilities across various pipelines, reinforcement learning (RL), rule-based, search-based methods environments. As first step toward addressing challenge using models, further develop an LLM-based agent perform in-depth analysis its promise solving challenging tasks. is available at https://vis-www.cs.umass.edu/hazard/.