基于大模型的自动驾驶决策与规划研究进展综述

打开文本图片集

主题词：自动驾驶大模型 决策和规划

中图分类号：U463 文献标志码：A DOI：10.19620/j.cnki.1000-3703.20250293

Large Language Models for Decision-Making and Planning in Autonomous Driving： A Survey

SongZiyul²，ChenJie³，JiangLinlin，Xia Zhaochenl²，DingHaitao1，2

（1.NationalKeyLaboratoryofAutomotiveChasis IntegrationandBionics，Jilin University，Changchunl3o25;2.College

of Automotive Enginering，Jilinuniversity，Changchun130o25;3.GlobalR&DCenter，ChinaFAWCorporationLimited， Changchun 130013）

【Abstract】Existing autonomous driving systems face numerous challenges in terms of interpretability，reward design， andgeneralization.Thispapersystematicallyreviewsthemulti-roleintegrationmechanismsandengineringimplementation pathsof large languagemodelsand visual language models inautonomous driving decision-making and planning.Firstly，a unified role-basedframework isproposed，dividing large modelsinto3functional modules：decisionand planningoutput， featureandactiongeneration，andreward functionmodeling.Onthisbasis，thispapercomprehensivelyreviewsthe representativeintegrationmethodswith reinforcement learning and imitationlearning，achievingaful-chainsummaryof empowerment mechanisms frominformation procesing to behavior output.The latestadvancementsof representative large modelsateach stageinbothdomesticandinternationalcontextsaresummarizedaswell.Furthermore，thispaperhighlights typicalsolutions topracticalengineeringchallnges，suchasinferencelatencyotimzation，automaticrewardsignalgeation end-to-end interpretabilityandcollaborativereasonng，whileexpandsthereviewtoincludeemerging topicssuchasreward model assistanceandmultimodal feature generation.Finally，futureresearch directionsareproposed inresponse to key chalenges including inferencelatency，model stabity，representationgaps，andtrainingandvalidation.Thisworkimsto provideasystematictheoreticalfoundationandapracticalengineeringroadmap forempoweringautonomousdrivingdecisionmaking and planning with foundation models.

Key words： Autonomous driving， Large Language Models （LLM），Decision-making and planning

【引用格式】宋子钰，陈捷，姜琳琳，等.基于大模型的自动驾驶决策与规划研究进展综述[J].汽车技术，2025（10）：21-31.SONG ZYCHENJ，JIANGLL，etal.LargeLanguage Models forDecision-MakingandPlannngin AutonomousDriving：ASurvey[J].Automobile Technology，2025（1O）： 21-31.

1前言

自动驾驶的研究范式正逐步从传统的模块化架构向端到端学习过渡[1-2]。（剩余23972字）