面向自动驾驶的大模型对齐技术:综述  

Large Model Alignment Technology for Autonomous Driving:A Review

作　　者：唐小林[1] 甘露[1] 李国法 李克强[2] 褚文博 Tang Xiaolin;Gan Lu;Li Guofa;Li Keqiang;Chu Wenbo(College of Mechanical and Vehicle Engineering,Chongqing University,Chongqing 400044;School of Vehicle and Mobility,Tsinghua University,Beijing 100084;China Intelligent and Connected Vehicles(Beijing)Research Institute Co.,Ltd.,Beijing 100176;Engineering Research Center of Mechanical Testing Tech.and Equip.Ministry of Education,Chongqing University of Technology,Chongqing 400054;Western China Science City Innovation Center of Intelligent and Connected Vehicles(Chongqing)Co.,Ltd.,Chongqing 401329)

机构地区：[1]重庆大学机械与运载工程学院,重庆400044 [2]清华大学车辆与运载学院,北京100084 [3]国汽(北京)智能网联汽车研究院有限公司,北京100176 [4]重庆理工大学机械检测技术与装备教育部工程研究中心,重庆400054 [5]西部科学城智能网联汽车创新中心(重庆)有限公司,重庆401329

出　　处：《汽车工程》2024年第11期1937-1951,共15页Automotive Engineering

基　　金：国家重点研发计划(2022YFB2503205);国家自然科学基金(52372377,52272421,52222215,52072051);重庆市自然科学基金(CSTB2023NSCOJOX0003,CSTB2023NSCQ-MSX0985);智能绿色车辆与交通全国重点实验室开放基金课题(KFZ2409)资助。

摘　　要：随着Transformer注意力机制的出现,以GPT为代表的通用基础大模型实现了智能的“涌现”,给自动驾驶迈向更高级别发展带来了曙光。受限于传统从头预训练方式需要大规模、高质量、多样性自动驾驶数据和高昂训练成本的困扰“,大模型+对齐技术”范式衍生。对齐技术作为通用基础大模型与自动驾驶之间的纽带,通过微调或提示工程等定制化方式,可高效、专业地解决自动驾驶领域内的工程性问题。对齐技术已是大模型在垂直领域发展的研究热点,但缺乏系统研究成果。基于此,本文首先对自动驾驶发展与大模型技术进行概述,从而衍生出对齐技术。然后,分别从微调和提示工程两个角度进行综述,系统化梳理并剖析各分类技术的结构或性能特点,同时给出实际的应用案例。最后,基于现有研究提出了对齐技术的研究挑战与发展趋势,为促进自动驾驶迈向更高级别发展提供参考。With the emergence of the Transformer attention mechanism,general-purpose large models rep-resented by GPT have achieved the"emergence"of intelligence,bringing a dawn to the advancement towards higher levels of autonomous driving.Limited by the traditional from-scratch pre-training approach,which requires large-scale,high-quality,diverse autonomous driving data and incurs high training cost,the"large model+alignment technology"paradigm has been derived.As a bridge between general-purpose large models and autonomous driving,alignment technology,through customization methods such as fine-tuning or prompt engineering,achieves efficient and professional solutions to engineering problems within the field of autonomous driving.Alignment technology has become a hot research topic in the development of large models in vertical fields,but it lacks systematic research re-sults.Based on this,this article firstly provides an overview of the development of autonomous driving and large model technology,thereby deriving alignment technology.Then,it reviews from the perspectives of fine-tuning and prompt engineering,systematically reviewing and analyzing the structure or performance characteristics of each clas-sification technology,while providing actual application cases.Finally,based on existing research,the research challenges and development trends of alignment technology are proposed,offering references for promoting the ad-vancement towards higher level of autonomous driving development.

关 键 词：自动驾驶 大模型 综述 对齐技术 微调 提示工程 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] U463.6[自动化与计算机技术—控制科学与工程]