Abstract
Zero-shot object-goal navigation enables an intelligent agent to explore and navigate to objects of unknown categories in an unfamiliar environment without specific target training. Existing methods using large language models (LLMs) treat them as flat reasoning tools, directly associating objects or regions but lack hierarchical spatial cognition for object localization. This leads to significant exploration blindness, insufficient semantic accuracy, and an inability to fully utilize LLMs' common-sense reasoning potential.
This paper introduces an LLM-driven Hierarchical Room-to-Object (HRO) framework for zero-shot object-goal navigation, guiding the agent in a coarse-to-fine manner. Experiments on the Gibson and HM3D datasets demonstrate that our HRO framework achieves superior success rates and generalization compared to existing LLM-based methods, highlighting LLMs' strong potential for zero-shot object-goal navigation.
Blogger's Review: The HRO framework significantly improves the efficiency and accuracy of zero-shot navigation by incorporating hierarchical cognition, showcasing the applicability of LLMs in complex environments. This research not only advances intelligent agent development but also lays the groundwork for future navigation systems.