基于知识与大语言模型的高速环境自动驾驶决策研究

doi:10.16182/j.issn1004731x.joss.24-0065

摘要/Abstract

摘要：

针对当前自动驾驶系统缺乏持续学习和可解释性问题。提出具有认知、泛化和学习能力的决策模型，利用大语言模型（large language model， LLM）、注意力机制理解和解释驾驶场景，通过模拟人类驾驶行为和决策过程，实现对驾驶经验的积累和学习并根据驾驶经验不断提升决策能力。在仿真环境中闭环测试决策模型在高速场景的应用，仿真结果表明：知识驱动模型决策成功率比规则和数据学习方法提高了7%和4%，具备泛化和可解释能力，提高了自动驾驶系统的可信度和安全性。

关键词: 驾驶经验, 大语言模型, 注意力机制, 可解释性, 知识驱动

Abstract:

Aiming at the lack of continuous learning and interpretability of current autonomous driving system,a decision model with cognition, generalization and learning ability is proposed. The model utilizes large language model (LLM) and attention mechanisms to understand and explain driving scenes. the system can accumulate and learn from driving experiences, continuously improving its decision-making ability. In a simulation environment, the closed-loop test decision model is applied in high-speed scenarios.The simulation results show that the success rate of the knowledge-driven model is 7% and 4% higher than those of the rule-based and data-driven methods. Additionally, the model exhibits generalization and interpretability, thereby enhancing the reliability and safety of the automatic driving system.

Key words: driving experience, LLM, attention mechanism, interpretability, knowledge-driven

中图分类号:

TP391.9

王祥,谭国真 . 基于知识与大语言模型的高速环境自动驾驶决策研究[J]. 系统仿真学报, 2025, 37(5): 1246-1255.

Wang Xiang,Tan Guozhen . Research on Decision-making of Autonomous Driving in Highway Environment Based on Knowledge and Large Language Model[J]. Journal of System Simulation, 2025, 37(5): 1246-1255.

图/表 13

图1

图2

图3

图4

图5

表1

表2

表3

运动学参数设置

控制	参数设置	四车道	五车道
纵向	$a m a x$ /(km/h²)	20	21
	$D t$ /s	1.5	2
	$a s u b$ /(km/h²)	-17	-17
	$d 0$ /m	15	20
横向	$σ$	0.1	0.05
横向	$λ a$	0.2	0.2

表3

图6

表4

表5

表6

图7

参考文献 22

1	王若萱, 吴建平, 徐辉. 自动驾驶汽车感知系统仿真的研究及应用综述[J]. 系统仿真学报, 2022, 34(12): 2507-2521.
	Wang Ruoxuan, Wu Jianping, Xu Hui. Overview of Research and Application on Autonomous Vehicle Oriented Perception System Simulation[J]. Journal of System Simulation, 2022, 34(12): 2507-2521.
2	郭延永, 刘佩, 袁泉, 等. 网联自动驾驶车辆道路交通安全研究综述[J]. 交通运输工程学报, 2023, 23(5): 19-38.
	Guo Yanyong, Liu Pei, Yuan Quan, et al. Review on Research of Road Traffic Safety of Connected and Automated Vehicles[J]. Journal of Traffic and Transportation Engineering, 2023, 23(5): 19-38.
3	高镇海, 郑程元, 赵睿. 真实与虚拟场景下自动驾驶车辆的主动安全性验证与确认研究综述[J/OL]. 吉林大学学报(工学版). (2023-12-29) [2024-01-02]. .
	Gao Zhenhai, Zheng Chengyuan, Zhao Rui. Review of Active Safety Verification and Validation for Autonomous Vehicles in Real and Virtual Scenarios[J/OL]. Journal of Jilin University(Engineering and Technology Edition). (2023-12-29) [2024-01-02]. .
4	吕彦直, 魏超, 何元浩. 基于GCN和CIL的端到端自动驾驶换道方法[J]. 汽车工程, 2023, 45(12): 2310-2317.
	Yanzhi Lü, Wei Chao, He Yuanhao. An End-to-end Lane Change Method for Autonomous Driving Based on GCN and CIL[J]. Automotive Engineering, 2023, 45(12): 2310-2317.
5	李晓豪, 郑海斌, 王雪柯, 等. 基于改进交叉熵的模仿学习鲁棒性增强方法[J]. 控制与决策, 2024, 39(3): 768-776.
	Li Xiaohao, Zheng Haibin, Wang Xueke, et al. Imitation Learning Robustness Enhancement Based on Modified Cross Entropy[J]. Control and Decision, 2024, 39(3): 768-776.
6	李传耀, 张帆, 王涛, 等. 基于深度强化学习的道路交叉口生态驾驶策略研究[J]. 交通运输系统工程与信息, 2024, 24(1): 81-92.
	Li Chuanyao, Zhang Fan, Wang Tao, et al. Signalized Intersection Eco-driving Strategy Based on Deep Reinforcement Learning[J]. Journal of Transportation Systems Engineering and Information Technology, 2024, 24(1): 81-92.
7	熊璐, 贾通, 陈君毅, 等. 基于有限状态机的预期功能安全危害识别方法[J]. 同济大学学报(自然科学版), 2023, 51(4): 616-622.
	Xiong Lu, Jia Tong, Chen Junyi, et al. Hazard Identification Method for Safety of the Intended Functionality Based on Finite State Machine[J]. Journal of Tongji University(Natural Science), 2023, 51(4): 616-622.
8	傅明建, 郭福强. 基于深度强化学习的无信号灯路口决策研究[J]. 计算机工程, 2024, 50(5): 91-99.
	Fu Mingjian, Guo Fuqiang. Research on Decision-making at Intersection Without Traffic Lights Based on Deep Reinforcement Learning[J]. Computer Engineering, 2024, 50(5): 91-99.
9	Xiao Yi, Codevilla F, Gurram Akhil, et al. Multimodal End-to-end Autonomous Driving[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(1): 537-547.
10	高振海, 闫相同, 高菲. 基于逆向强化学习的纵向自动驾驶决策方法[J]. 汽车工程, 2022, 44(7): 969-975.
	Gao Zhenhai, Yan Xiangtong, Gao Fei. A Decision-making Method for Longitudinal Autonomous Driving Based on Inverse Reinforcement Learning[J]. Automotive Engineering, 2022, 44(7): 969-975.
11	林歆悠, 叶卓明, 周斌豪. 基于DQN强化学习的自动驾驶转向控制策略[J]. 机械工程学报, 2023, 59(16): 315-324.
	Lin Xinyou, Ye Zhuoming, Zhou Binhao. DQN Reinforcement Learning-based Steering Control Strategy for Autonomous Driving[J]. Journal of Mechanical Engineering, 2023, 59(16): 315-324.
12	Alayrac J B, Donahue J, Luc P, et al. Flamingo: A Visual Language Model for Few-shot Learning[C]//Advances in Neural Information Processing Systems. Red Hook: Curran Associates, Inc., 2022: 23716-23736.
13	中国科学技术信息研究所. 中国人工智能大模型地图研究报告[EB/OL]. (2023-05-28) [2023-10-06]. .
14	苏炯铭, 罗俊仁, 陈少飞. 智能博弈决策策略求解新视角实证分析[J]. 系统仿真学报, 2025, 37(2): 345-361.
	Su Jiongming, Luo Junren, Chen Shaofei. An Empirical Analysis of New Perspectives for Strategy Solving in Intelligent Game-theoretic Decision-making[J]. Journal of System Simulation, 2025, 37(2): 345-361.
15	Lin Yuchen, Huang Chengsong, Liu Qian, et al. On Grounded Planning for Embodied Tasks with Language Models[C]//Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. Palo Alto, CA, USA: AAAI Press, 2023: 13192-13200.
16	Yang A, Nagrani A, Seo P H, et al. Vid2Seq: Large-scale Pretraining of a Visual Language Model for Dense Video Captioning[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 10714-10726.
17	罗飞, 崔滨, 辛小江, 等. 大语言模型嵌入图书馆知识服务的风险范式与管控策略[J]. 图书与情报, 2023(3): 99-106.
	Luo Fei, Cui Bin, Xin Xiaojiang, et al. The Risk Paradigm and Management Strategies of Embedding Big Language Model into Library Knowledge Services[J]. Library & Information, 2023(3): 99-106.
18	Liu Chuanming, Fu Shuyu. Effective Protocols for kNN Search on Broadcast Multi-dimensional Index Trees[J]. Information Systems, 2008, 33(1): 18-35.
19	Nie Xiaotong, Liang Yupeng, Ohkura Kazuhiro. Autonomous Highway Driving Using Reinforcement Learning with Safety Check System Based on Time-to-collision[J]. Artificial Life and Robotics, 2023, 28(1): 158-165.
20	Kong J, Pfeiffer Mark, Schildbach G, et al. Kinematic and Dynamic Vehicle Models for Autonomous Driving Control Design[C]//2015 IEEE Intelligent Vehicles Symposium (IV). Piscataway: IEEE, 2015: 1094-1099.
21	李伟东, 马草原, 史浩, 等. 基于分层强化学习的自动驾驶决策控制算法[J/OL]. 吉林大学学报(工学版). (2023-12-19) [2024-01-10]. .
	Li Weidong, Ma Caoyuan, Shi Hao, et al. An Automatic Driving Decision Control Algorithm Based on Hierarchical Reinforcement Learning[J/OL]. Journal of Jilin University(Engineering and Technology Edition). (2023-12-19) [2024-01-10]. .
22	张新锋, 吴琳. 基于集成深度强化学习的自动驾驶车辆行为决策模型[J]. 汽车安全与节能学报, 2023, 14(4): 472-479.
	Zhang Xinfeng, Wu Lin. Behavior Decision-making Model for Autonomous Vehicles Based on an Ensemble Deep Reinforcement Learning[J]. Journal of Automotive Safety and Energy, 2023, 14(4): 472-479.

序号	名称	含义
1	Line-Left	转向左车道
2	Line-Right	转向右车道
3	Idle	保持当前状态
4	Faster	加速
5	Slower	减速

序号	参数	取值	范围
1	Top-k	3	[1,10]
2	Top-p	0.75	(0,1]
3	Temperature	1	(0,1]

方法	Line-Left	Line-Right	Idle	Faster	Slower
RDM	985	1 002	1 785	123	74
DRLM	785	685	2 033	141	35
KDM₂	877	943	1 943	104	27

车道数	方法	碰撞次数	成功率/%	平均速度/(km/h)
4车道	RDM	9	91	105.5
	DRLM	6	94	103.1
	KDM₁	12	88	104.6
	KDM₂	2	98	106.1
5车道	RDM	8	92	107.7
	DRLM	7	93	105.2
	KDM₂	3	97	108.8

方法	碰撞次数	成功率/%	平均速度/(km/h)
RDM	13	87	98.7
DRLM	11	89	96.4
KDM₂	4	96	100.2