汽车工程 ›› 2024, Vol. 46 ›› Issue (4): 596-604.doi: 10.19562/j.chinasae.qcgc.2024.04.005

• • 上一篇    

基于逆模型预测控制的拟人驾驶控制

刘辉1,张发旺1,聂士达1(),段京良2,郭丛帅1,郭凌雄1   

  1. 1.北京理工大学机械与车辆学院,北京 100081
    2.北京科技大学机械工程学院,北京 100083
  • 收稿日期:2023-09-04 修回日期:2023-10-30 出版日期:2024-04-25 发布日期:2024-04-24
  • 通讯作者: 聂士达 E-mail:nieshida@bit.edu.cn
  • 基金资助:
    国家自然科学基金(52002212);部级基金项目(2020-063)

Human-Like Driving Control Based on Inverse Model Predictive Control

Hui Liu1,Fawang Zhang1,Shida Nie1(),Jingliang Duan2,Congshuai Guo1,Lingxiong Guo1   

  1. 1.School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081
    2.School of Mechanical Engineering, University of Science and Technology Beijing, Beijing 100083
  • Received:2023-09-04 Revised:2023-10-30 Online:2024-04-25 Published:2024-04-24
  • Contact: Shida Nie E-mail:nieshida@bit.edu.cn

摘要:

本文提出一种基于逆模型预测控制的拟人驾驶控制方法,利用模型预测控制产生的实轴轨迹与真实轨迹的损失函数更新控制模块代价函数的权重系数实现拟人化驾驶控制。将拟人驾驶控制构建成一个双层优化问题,在下层利用模型预测控制求解一个典型的最优控制问题产生实轴驾驶轨迹,在上层最小化所产生的实轴轨迹和真实驾驶轨迹的误差更新下层代价函数的权重系数,基于极大值微分原理构造辅助系统求解实轴轨迹关于代价函数权重系数的梯度。实车采集真实驾驶轨迹并进行模仿测试与泛化验证,结果表明:本文所提出的方法相比于两类基于虚轴轨迹的逆最优控制方法,在3个工况下与真实驾驶轨迹最大误差分别平均降低了73.52%和65.03%,驾驶行为更加拟人化,且具备泛化性能。

关键词: 自动驾驶, 拟人化驾驶, 逆最优控制

Abstract:

In this paper, a human-like driving control based on inverse model predictive control is proposed, which realizes human-like driving by updating the weight coefficients of the cost function of the control module using the loss function of the real-time trajectory generated by the model predictive control and the driver's trajectory. The human-like driving control is constructed as a two-layer optimization problem. In the lower layer, real-time state trajectories are generated by solving a typical optimal control problem using model predictive control. The optimization objective function of the lower layer is then updated by minimizing the error between the generated real-time trajectories and those of human drivers in the upper layer. The auxiliary systems based on the differential Pontryagin's Maximum Principle are constructed to solve the gradient of the weight coefficients of the cost function for the real axis trajectory. The driver's driving data are collected from the real vehicle, imitated, and tested. The results show that the method proposed in this paper, compared with two types of inverse optimal control methods based on the virtual-time trajectory, reduces the maximum error with the real trajectory by 73.52% and 65.03% in the three test conditions, with the driving behavior more anthropomorphic and has the generalization performance.

Key words: self-driving, human-like driving, inverse optimal control