基于检索增强生成的任务规划方法综述OA

Retrieval-augmented generation(RAG)technology has become a key paradigm for improving the accuracy of large language model task response by dynamically integrating external knowledge to effectively alleviate the hallucination problem and knowledge timeliness of large language models.However,how to efficiently utilize the retrieval-augmented generation architecture to adapt to the planning requirements of complex tasks remains a core challenge.As a popular solution,retrieval augmentation generation is triggered according to different timings,which significantly enhances the planning robustness of large language models in knowledge-intensive scenarios and complex planning tasks.Firstly,the necessity of decision triggering in the retrieval-enhanced generation framework is explained.Next,focusing on the core issue of the timing for triggering RAG-enhanced task planning,this paper categorizes the relevant triggering mechanisms into three types:rule-oriented triggers,state-adaptive triggers,and interactive feedback triggers,and conducts a systematic review.Then,starting from the basic paradigm of retrieval-augmented planning,it introduces the general role of retrieval-generation architectures in task planning,abstracts existing work into two advanced paradigms:global node planning and hierarchical task planning,and further unifies the three paradigms in terms of time,space,and safety dimensions,forming a cross-paradigm evolutionary framework consisting of memory-augmented RAG,multi-agent collaborative RAG,and adversarially robust RAG.Finally,it summarizes the current challenges of applying retrieval-augmented generation technology to assist task planning in practical applications.