端到端自动驾驶的研究进展及挑战  

End-to-end Autonomous Driving: Advancements and Challenges

作　　者：褚端峰[1,2] 王如康 王竞一 花俏枝 陆丽萍[5] 吴超仲[1,2] CHU Duan-feng;WANG Ru-kang;WANG Jing-yi;HUA Qiao-zhi;LU Li-ping;WU Chao-zhong(Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan 430063,Hubei,China;Engineering Research Center for Transportation Information and Safety,Ministry of Education,Wuhan University of Technology,Wuhan 430063,Hubei,China;School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,Hubei,China;Computer School,Hubei University of Arts and Science,Xiangyang 441053,Hubei,China;School of Computer Science and Artificial Intelligence,Wuhan University of Technology,Wuhan 430070,Hubei,China)

机构地区：[1]武汉理工大学智能交通系统研究中心,湖北武汉430063 [2]武汉理工大学交通信息与安全教育部工程研究中心,湖北武汉430063 [3]武汉理工大学机电工程学院,湖北武汉430070 [4]湖北文理学院计算机工程学院,湖北襄阳441053 [5]武汉理工大学计算机与人工智能学院,湖北武汉430070

出　　处：《中国公路学报》2024年第10期209-232,共24页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52172393);湖北省杰出青年基金项目(2022CFA091);武汉市科技重大专项(2022013702025184)。

摘　　要：端到端自动驾驶方法无需人工预先定义规则和模块间显式接口,直接从原始传感器数据映射出轨迹点或控制信号,消除了传统模块化方法中的信息丢失、级联误差等固有缺陷,突破了规则驱动下的系统性能瓶颈。近年来,基于自监督学习的生成式人工智能展现出强大的“智能涌现”能力,显著推动了端到端方法的发展。然而,现有综述未能系统总结端到端自动驾驶的进展。为此,全面梳理了端到端自动驾驶研究进展、技术挑战和发展趋势。首先,归纳了端到端模型的输入和输出模态,并基于端到端自动驾驶的发展历程,对传统端到端方法、模块化端到端方法、生成式端到端方法的基本概念、研究现状、技术挑战进行了概述和对比分析;然后,总结了端到端模型的评估方法及其训练数据集;再则,探讨了当前端到端自动驾驶技术在泛化性、可解释性、因果性、安全性、舒适性等方面所面临的挑战;最后,展望了端到端自动驾驶的发展趋势,指出了场景数据生成为端到端模型训练提供支撑,知识驱动有助于提升模型的泛化能力,自监督学习有效提升模型的训练效率,个性化驾驶优化驾乘体验,世界模型则成为端到端自动驾驶进一步发展的关键方向。研究结果可为完善端到端自动驾驶的理论体系和提升系统性能提供重要的参考。End-to-end autonomous driving methodologies eliminate the need for manually defined rules and explicit module interfaces.Instead,these approaches directly map trajectory points or control signals from raw sensor data,thereby addressing the inherent shortcomings associated with traditional modular methods,such as information loss and cascading errors,and overcoming the performance limitations imposed by rule-driven frameworks.Recent advancements in self-supervised-learning-based generative artificial intelligence have exhibited substantial emergent intelligence capabilities,significantly promoting the evolution of end-to-end methodologies.However,the existing literature lacks a comprehensive synthesis of the advancements in generative end-to-end autonomous driving.Consequently,this paper systematically reviews the research progress,technical challenges,and developmental trends in end-to-end autonomous driving.Initially,the input and output modalities of the end-to-end models are delineated.Based on the historical progression of end-to-end autonomous driving,this paper provides an overview and comparative analysis of the foundational concepts,current research status,and technical challenges of traditional,modular,and generative end-to-end methods.Subsequently,the evaluation methodologies and training datasets utilized for end-to-end models are summarized.Furthermore,this paper explores the challenges currently faced by end-to-end autonomous driving technologies in relation to generalization,interpretability,causality,safety,and comfort.Finally,predictions are made for the future trends of end-to-end autonomous driving,emphasizing the fact that edge scenarios provide critical support for the training of end-to-end models,which can enhance the generalization capabilities.In addition,self-supervised learning can effectively improve training efficiency,personalized driving can optimize user experience,and world models represent a pivotal direction for the further advancement of end-to-end autonomous driving.The finding

关 键 词：汽车工程 自动驾驶 综述 端到端自动驾驶 生成式人工智能 可解释性 泛化性 

分 类 号：U469.79[机械工程—车辆工程]