Adaptive Retrieval-Augmented Embodied Intelligence RAG Reasoning Framework

Large Language Models (LLMs) have demonstrated significant potential in task reasoning for embodied intelligence. However, when confronted with unfamiliar scenarios, these models frequently generate fabricated or erroneous reasoning, a phenomenon known as “hallucination”, which limits their applicability in real-world tasks. In response to this challenge, this paper proposes an embodied task reasoning method based on adaptive Retrieval-Augmented Generation (RAG) technology. Firstly, we store real-world environmental information in a vector knowledge graph, facilitating graph retrieval and reasoning generation. Secondly, when graph retrieval information is insufficient, we employ an adaptive web retriever constructed using LLMs and web search interfaces, which dynamically enhances reasoning generation through a “retrieve-reason-evaluate” loop framework. Finally, new long-term knowledge is extracted and stored in the knowledge graph via an active memory module, improving retrieval efficiency for subsequent similar tasks. Experimental results demonstrate that this method effectively reduces the hallucination phenomenon during the reasoning process, optimizes the task reasoning workflow, and provides a scalable framework for reliable knowledge acquisition and dynamic reasoning in embodied intelligence.