摘要
端到端自动驾驶方法作为卖现自动驾驶的方法之-',相对于“感知-决策-栓制”方法,有着思路简单、感知信息利用克分、工况过渡平滑的优点,目裔越来越多的学者对其开展研老。本丈青先讨论了卖现自动驾驶的两种方法;第二,呵顾了端到端旨动驾驶方法发展的易史幷将其分为3个阶段,即1989-2005年、2006-2015年和2016年至今;第三,总体概述了研老现状,卖现算法主要为谦度神经网络和深度增强学习等,并对基于不同算法的相关研老进行了详细论述.;最后,对端到端自动驾飒研老进行了总结。
As one of the methods to realize self-driving, Endto-End Self-driving has the advantages of simple logic,full use of perception information and smooth transitionof working conditions compared with Perception-Planning-Control method. Nowadays, more and moreresearches about it have been carried out. The articlefirstly discusses two methods to realize self-driving.Secondly, it reviews the history of the development ofEnd-to-End Self-Driving and divide it into three phases,namely 1989—2005 , 2006—2015 and 2016 to present.Thirdly, overview the research status, point out thatEnd-to-End Self-Driving is implemented mainly by DeepNeural Network and Deep Reinforcement Learning, anddiscuss the researches based on different algorithms.Finally, this paper summarize the development status ofEnd-to-End Self-Driving.
作者
王庭晗
罗禹贡
李克强
Wang Tinghan;Luo Yugong;Li Keqiang(State Key Lab of Automotive Safety and Energy,Tsinghua University,Beijing 10084)
出处
《中国基础科学》
2019年第S01期36-46,F0002,共12页
China Basic Science
基金
国家重点研发计划项目(2016YFB0100905)。
