汽车工程 ›› 2025, Vol. 47 ›› Issue (5): 797-808.doi: 10.19562/j.chinasae.qcgc.2025.05.001

• •    下一篇

面向狭窄环境的安全泊车路径规划算法研究

管家意1,李斌1,周傲1,赵治国1,林巧2,陈广1()   

  1. 1.同济大学,上海 201804
    2.易控智驾科技有限公司,北京 100083
  • 收稿日期:2025-01-14 修回日期:2025-03-04 出版日期:2025-05-25 发布日期:2025-05-20
  • 通讯作者: 陈广 E-mail:guangchen@tongji.edu.cn
  • 基金资助:
    国家重点研发计划项目(2024YFE0211000);国家自然科学基金面上项目(62372329);上海市科技创新行动计划社会发展科技攻关项目(23DZ1203400);同济大学-Qomolo商用车自动驾驶联合实验室和小米青年学者基金资助

Study on Safe Parking Path Planning Algorithm for Narrow Environment

Jiayi Guan1,Bin Li1,Ao Zhou1,Zhiguo Zhao1,Qiao Lin2,Guang Chen1()   

  1. 1.Tongji University,Shanghai 201804
    2.EACON Technology Co. ,Ltd. ,Beijing 100083
  • Received:2025-01-14 Revised:2025-03-04 Online:2025-05-25 Published:2025-05-20
  • Contact: Guang Chen E-mail:guangchen@tongji.edu.cn

摘要:

针对自动泊车系统中路径规划的安全性、实时性和可行性问题,本文提出一种基于混合动作空间约束强化学习的泊车路径规划算法。具体地,该算法利用混合动作空间强化学习框架将离散动作和连续参数相结合实现参数化轨迹规划,提高了规划路径的可执行性;在此基础上设计一种混合动作空间的约束强化学习算法实现安全策略优化,确保了泊车路径的安全性。此外,在模型训练过程中引入课程学习机制逐步引导策略探索,增强了模型训练稳定性和收敛速度。最后,在垂直车位和平行车位进行广泛的对比和消融实验,实验结果表明所提出的泊车路径规划算法在成功率、安全性和实时性等指标上均表现出色,且综合性能明显优于现有基线算法。

关键词: 自动泊车, 混合动作空间强化学习, 路径规划, 安全约束

Abstract:

For safe and feasible path-planning in real time of autonomous parking system, a parking path planning algorithm based on constrained reinforcement learning with a hybrid action space is proposed in this paper. Specifically, the proposed algorithm employs a hybrid action space reinforcement learning framework that integrates discrete actions with continuous parameters to achieve parameterized trajectory planning, thereby enhancing the executability of planned paths. On this basis, a constrained reinforcement learning algorithm within the hybrid action space is designed to optimize safe policy execution, ensuring the safety of parking paths. Moreover, a curriculum learning mechanism is introduced during model training to guide exploration progressively, improving training stability and convergence speed. Finally, extensive comparative and ablation experiments are conducted on both perpendicular and parallel parking scenarios. The experimental results show that the proposed parking path planning algorithm outperforms existing state-of-the-art methods in terms of success rate, safety, and real-time performance, exhibiting superior overall effectiveness.

Key words: autonomous parking, hybrid-action reinforcement learning, motion-planning, constraint optimization