[5G3-OS-37b] OS-37

2:30 PM - 2:45 PM JST(5:30 AM - 5:45 AM UTC)

[5G3-OS-37b-03] Prompt Compression for Dialog-Based Navigation Systems Using Need-Oriented Knowledge and Retrieval-Augmented Generation

〇Hiroaki Shimoma^1,2, Sudesna Chakraborty¹, Takeshi Morita^1,2, Aoi Oota², Masaki Asada², Shusaku Egami², Takanori Ugai^2,3, Masahiro Hamasaki² (1. Aoyama Gakuin University, 2. National Institute of Advanced Industrial Science and Technology,, 3. Fujitsu Limited)

Keywords:

Need-Oriented Environmental Knowkedge Base,Prompt Compression,Navigation,Large Language Model,Retrieval-Argumented Generation

In the field of Embodied AI, where agents interact with virtual or real environments, Large Language Models (LLMs) are increasingly used in dialog-based navigation systems. However, existing approaches embed all environmental knowledge directly into navigation prompts, causing the prompt length to grow as the number of objects increases, which leads to reduced inference accuracy and higher computational and API usage costs. To address this issue, we propose a prompt compression method that integrates Retrieval-Augmented Generation (RAG) with a need-oriented environmental knowledge base grounded in Murray’s theory of human needs. In the proposed system, each need is associated with relevant actions and environmental objects, and user needs are inferred through an RAG-based retrieval process. By externalizing environmental knowledge and selectively retrieving only task-relevant needs and objects, the method effectively limits prompt content. We evaluate the proposed approach using a dialog-based navigation dataset constructed from OpenEQA in the VirtualHome household simulator. Experimental results demonstrate that the proposed method significantly reduces prompt length while improving navigation accuracy.