机器人大语言模型导航研究进展综述（英）H-Pan等-2024-10-2

agents (computational software agents simulating hu-based on their focus on specific tasks.These effortsman behavior using LLMs)offers a new perspective are crucial for understanding the connection betweenfor improving human-machine interaction [14].LLMs and robot navigation.Thirdly.Section 4 pro-In summary,the application of LLMs in robot nav-vides a brief comparison and analysis of the relevantigation is a promising research direction.However,datasets.Fourthly,in Section 5,we introduce metricsthis field still faces numerous challenges,such as howfor evaluating dataset tasks and compare the perfor-to effectively encode environmental information into mance of relevant models.Subsequently,Section 6text,how to enable robots to understand and pro-discusses unresolved issues,potential research direc-cess complex environmental information.how to fa-tions,as well as future challenges and research trends.cilitate robots in making rational decisions,how toFinally,we conclude the paper in Section 7.improve human-robot interaction,and how to achieveautonomous decision-making and reasoning.To comprehensively understand LLMs-based nav-2Backgroundigation technology and advance further research inthis field,the paper summarizes the latest advance-In this section,we briefly review the progress inments in LLM-based navigation and discusses futureLLMs and robot navigation.research directions.It is also noted that a recent sur-vey [15,16 reviewed research related to LLM-based2.1 Large Language Modelsnavigation.In comparison to these works,this paperdiffers in the following aspects:LLMs represent a class of Transformer-based lan-guage models renowned for their expansive parame-This research study focuses on the exploration ofter count,often reaching hundreds of billions.TheseLLM-based navigation,which plays a pivotal role models undergo training using vast quantities of in-in advancing this technology.ternet data.endowing them with a broad array ofThis paper primarily examines the role of LLMslanguage capabilities,primarily manifested throughtext generation.Prominent examples of LLMs in-in various aspects of navigation:Perception,Plan-ning,Control,Interaction,and Coordination.clude GPT-3 [11],PaLM [17],LLaMA [18],and GPT-4 [19].One salient attribute exhibited by LLMs isLLM-based navigation methods are classified basedtheir emergent proficiencies,such as in-context learn-on the physical environments where robot naviga-ing (ICL)[11],instruction comprehension and chain-tion tasks are applied:indoor,on-road,and off-of-thought reasoning (CoT)[20]road environments.In contrast to traditional machine learning models.the prowess of sizable language models is predomi-This article presents a comprehensive survey of nantly evidenced by their deep bidirectional represenLLMs applied to navigation,encompassing theoreti-tations,potent context comprehension,and efficientcal foundations and practical applications.Figure 1 handling of complex tasks.Conventional machineoffers a descri