Paper Details
Abstract
This paper presents CORA (COntextual Retrieval and Allocation), a unified framework that combines Retrieval-Augmented Generation (RAG) with multi-agent task allocation and planning systems to support aging-in-place in smart homes. The system utilizes a curated elderly care knowledge base to enable context-aware retrieval and grounded response generation. Simultaneously, distributed language agents manage task decomposition, allocation, and execution within an embodied environment. Evaluation in the AI2-THOR (AI2 - The House Of inteRactions) simulator demonstrates that our hybrid retrieval approach outperforms standalone methods (TF-IDF, BM25, semantic retrieval), achieving 0.613 Precision@3, while GPT-4o-based multi-agent coordination achieves up to 95% task completion rate, significantly outperforming other Large Language Models (LLMs) backbones. The results highlight the potential of integrating language-driven reasoning with embodied intelligence for scalable and adaptive eldercare systems.