汽车工程 ›› 2023, Vol. 45 ›› Issue (9): 1499-1515.doi: 10.19562/j.chinasae.qcgc.2023.ep.006

所属专题: 智能网联汽车技术专题-规划&决策2023年

• •    下一篇

类脑学习型自动驾驶决控系统的关键技术

李升波(),占国建,蒋宇轩,兰志前,张宇航,邹文俊,陈晨,成波,李克强   

  1. 清华大学车辆与运载学院,汽车安全与节能国家重点实验室,北京 100084
  • 收稿日期:2023-02-13 修回日期:2023-03-16 出版日期:2023-09-25 发布日期:2023-09-23
  • 通讯作者: 李升波 E-mail:lishbo@tsinghua.edu.cn
  • 基金资助:
    十四五国家重点研发计划(2022YFB2502901);国家自然科学基金(U20A20334);清华大学自主科研计划资助

Key Technologies of Brain-Inspired Decision and Control Intelligence for Autonomous Driving Systems

Shengbo Eben Li(),Guojian Zhan,Yuxuan Jiang,Zhiqian Lan,Yuhang Zhang,Wenjun Zou,Chen Chen,Bo Cheng,Keqiang Li   

  1. School of Vehicle and Mobility,Tsinghua University,State Key Laboratory of Automotive Safety and Energy,Beijing 100084
  • Received:2023-02-13 Revised:2023-03-16 Online:2023-09-25 Published:2023-09-23
  • Contact: Shengbo Eben Li E-mail:lishbo@tsinghua.edu.cn

摘要:

作为高级别自动驾驶的下一代技术方向,类脑学习以深度神经网络为策略载体,以强化学习为训练手段,通过与环境的交互探索实现策略的自我进化,最终获得从环境状态到执行动作的最优映射。目前,类脑学习方法主要用于自动驾驶的决策与控制功能设计,它的关键技术包括:界定策略设计的系统框架、支持交互训练的仿真平台、决定策略输入的状态表征、定义策略目标的评价指标以及驱动策略更新的训练算法。本文重点梳理了自动驾驶决策控制的发展脉络,包括两类模块化架构(分层式和集成式)和3种技术方案(专家规则型、监督学习型和类脑学习型);概述了当前主流的自动驾驶仿真平台;分析了类脑决控的3类环境状态表征方法(目标式、特征式和组合式);同时介绍了自动驾驶汽车的五维度性能评价指标(安全性、合规性、舒适性、通畅性与经济性);然后详述了用于车云协同训练的典型强化学习算法及其应用现状;最后总结了类脑自动驾驶技术的问题挑战与发展趋势。

关键词: 智能网联汽车, 车路云协同, 类脑学习, 自主决策, 运动控制

Abstract:

As the technical trend of the next generation of high-level autonomous driving, brain-inspired learning is a class of methods that employ deep neural networks (DNN) as the strategy carrier and reinforcement learning (RL) as the training algorithm to realize strategy self evolution through continuous interaction with traffic environments, ultimately obtaining the optimal mapping from the environmental state to execution action. Currently, brain-inspired learning is mainly applied in decision-making and motion control modules of autonomous driving. Its key technologies include how to design its system framework to support interactive training, high-fidelity autonomous driving simulation platform, accurate and flexible representation of environment statues, multiple dimensional evaluation metrics, and effective training algorithm that drives policy updates. This paper systematically summarizes the history and future trends of decision-making and control functionalities in autonomous vehicles, including two main modular architectures (HDC, i.e., hierarchical decision & control and IDC, i.e., integrated decision & control) and three mainstream technical solutions (i.e., rule-based design, supervised learning, and brain-inspired learning). An overview of autonomous driving simulation platforms are briefly introduced, followed by three effective designing methods for representing traffic environment states (i.e., object-based design, feature-based design, and combined design). The paper also introduces multiple dimensional evaluation metrics for autonomous vehicles, which can describe self-driving performances including driving safety, regulatory compliance, driving comfort, travel efficiency, energy efficiency. Typical reinforcement learning algorithms, including their design principles, taxonomy, and algorithm performances, are introduced, as well as their application on brain-inspired autonomous driving systems in the systematic design of road-cloud cooperation.

Key words: intelligent and connected vehicle, vehicle-road-cloud cooperation, brain-inspired learning, decision-making, motion control