面向ROS知识问答的层次化图检索增强生成方法OA
A Hierarchical Graph Retrieval-Augmented Generation Framework for ROS Knowledge Question Answering
针对机器人操作系统(ROS)图检索过程中易出现的上下文碎片化、宏微观结构脱节问题,本文提出一种层次化图检索增强生成框架(HiGraphRAG).该方法通过全局图谱重组修复物理连通断层,依托高斯混合模型构建自底向上的层次化摘要索引,并结合视角扩展机制实现多粒度结构化知识的精准召回.基于自制 ROS 数据集的实验结果表明,相较于主流 GraphRAG 范式,HiGraphRAG 的总体生成质量胜率达 56.4%,知识多样性胜率达 60.2%.该框架可有效弥补系统信息断层,显著提升大模型复杂推理的全面性与知识赋能能力.
To address the problems of context fragmentation and macro-micro structure disconnection in graph retrieval for the Robot Operating System(ROS),this paper proposes a Hierarchical Graph Retrieval-Augmented Generation framework(HiGraphRAG).The framework repairs physical connectivity faults through global graph reorganization,constructs a bottom-up hierarchical summary index based on Gaussian Mixture Models(GMM),and adopts a perspective expansion mechanism to realize the accurate recall of multi-granularity structured knowledge.Experimental results on a self-built ROS dataset show that compared with the mainstream GraphRAG paradigm,HiGraphRAG achieves an overall generation quality win rate of 56.4%and a knowledge diversity win rate of 60.2%.The framework effectively bridges system information gaps and significantly improves the comprehensiveness and knowledge empowerment of large language models in complex reasoning.
施亚龙;赵俊;侯劲
四川轻化工大学,四川 宜宾 644005四川轻化工大学,四川 宜宾 644005四川轻化工大学,四川 宜宾 644005
信息技术与安全科学
检索增强生成知识图谱机器人操作系统层次化索引层次化检索
Retrieval-Augmented GenerationKnowledge Graphrobot operating systemhierarchical indexhierarchical retrieval
《现代信息科技》 2026 (17)
58-65,8
评论