New Framework Boosts Zero-Shot Object Navigation for LLMs

Zero-shot object-goal navigation, where an AI agent must find an unknown object in an unfamiliar environment without prior training for that specific target, is a challenging task. While Large Language Models (LLMs) are often employed to leverage their vast prior knowledge, existing methods typically use them as flat reasoning tools, directly associating objects or regions. This approach often leads to inefficient exploration and insufficient accuracy in semantic understanding, failing to fully utilize the LLMs' common-sense reasoning capabilities. To overcome these limitations, a new framework called HRO (Hierarchical Room-to-Object) has been proposed. HRO guides the agent in a coarse-to-fine manner, mimicking human-like hierarchical spatial cognition by first understanding room semantics before localizing objects within those rooms. This structured approach significantly enhances the agent's ability to explore effectively and make more accurate semantic associations. Experiments conducted on the Gibson and HM3D datasets confirm that the HRO framework achieves superior success rates and generalization compared to other LLM-based methods, demonstrating the strong potential of LLMs when integrated with hierarchical spatial reasoning for zero-shot navigation.