汽车工程 ›› 2024, Vol. 46 ›› Issue (4): 564-576.doi: 10.19562/j.chinasae.qcgc.2024.04.002

• • 上一篇    

基于双估计强化学习结合前向预测控制的自动驾驶运动控制研究

杜国栋1,2,邹渊1(),张旭东1,孙文景1,孙巍1   

  1. 1.北京理工大学机械与车辆学院,北京 100081
    2.苏黎世联邦理工大学动态系统与控制系,苏黎世 8006
  • 收稿日期:2023-09-06 修回日期:2023-10-13 出版日期:2024-04-25 发布日期:2024-04-24
  • 通讯作者: 邹渊 E-mail:zouyuanbit@vip.163.com
  • 基金资助:
    国家重点研发计划项目(2021YFB2500900)

Research on Automatic Driving Motion Control Based on Double Estimator Reinforcement Learning Combined with Forward Predictive Control

Guodong Du1,2,Yuan Zou1(),Xudong Zhang1,Wenjing Sun1,Wei Sun1   

  1. 1.School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081
    2.Institute of Dynamic System and Control,ETH Zurich,Zurich 8006
  • Received:2023-09-06 Revised:2023-10-13 Online:2024-04-25 Published:2024-04-24
  • Contact: Yuan Zou E-mail:zouyuanbit@vip.163.com

摘要:

运动控制研究是实现自动驾驶目标的重要组成部分,针对传统强化学习算法在求解中因单步决策局限而导致控制序列次优的问题,提出了一种基于双估计强化学习算法及前向预测控制方法结合的运动控制框架(DEQL-FPC)。在该框架中引入双估计器以解决传统强化学习方法动作值过估计问题并提高训练优化的速度,设计前向预测多步决策方法替代传统强化学习的单步决策,以有效提高全局控制策略的性能。通过虚拟驾驶环境仿真,证明了该控制框架应用在自动驾驶汽车的路径跟踪以及安全避障的优越性,保证了运动控制中的精确性、安全性、快速性以及舒适性。

关键词: 自动驾驶汽车, 运动控制优化, 双估计强化学习算法, 前向预测控制方法

Abstract:

Motion control research is an important part to achieve the goal of autonomous driving. To solve the problem of suboptimal control sequence due to the limitation of single-step decision in traditional reinforcement learning algorithm, a motion control framework based on the combination of double estimator reinforcement learning algorithm and forward predictive control method (DEQL-FPC) is proposed. In this framework, double estimators are introduced to solve the problem of action overestimation of traditional reinforcement learning methods and improve the speed of optimization. The forward predictive multi-step decision making method is designed to replace the single step decision making of traditional reinforcement learning so as to effectively improve the performance of global control strategies. Through virtual driving environment simulation, the superiority of the control framework applied in path tracking and safe obstacle avoidance of autonomous vehicles is proved, and the accuracy, safety, rapidity and comfort of motion control are guaranteed.

Key words: autonomous vehicle, motion control optimization, double estimator reinforcement learning algorithm, forward predictive control method