基于深度强化学习与动态运动基元的自动驾驶类人轨迹规划

打开文本图片集

主题词：自动驾驶 轨迹规划动态运动基元深度强化学习安全避障约束中图分类号：U469.72 文献标志码：A DOI：10.19620/j.cnki.1000-3703.20250460

Human-Like Trajectory Planning for Autonomous Vehicles Based on Deep Reinforcement Learning and Dynamic Movement Primitives

Xiu Limei，Liu Yuansheng （CollegeofRobotics （College of Artificial Intelligence），Beijing Union University，Beijing 100101）

【Abstract】To address issues such as vehicle control jiterand trajectory discontinuitycausedby existing reinforcement learning methodsthatdirectlyoutputlow-levelcontrolcommands，a hierarchical control framework integratingDep Reinforcement Learning （DRL）andDynamic MovementPrimitives （DMP） is proposed.Theautonomous driving control taskis divided into2stages：high-levelsemanticdecision-makingandlow-level trajectorygeneration.Thehigh-levelDRLmodule outputs driving intentions and DMP parametersbasedonreal-timeenvironmentalobservations.Atthelowlevel，bymodeling humandemonstrationdata，DMPisusedtolearnlatentfeaturerepresentationsofdrivingskills.Additionall，adynamic obstacleavoidancecoupling termisconstructedbyincorporatingthepositionsofobstaclesandsuroundingvehicles，generating continuous，smoothcolision-avoidancetrajectories thatalignwith human driving habits.Tocomprehensivelyevaluatethe performanceof the proposed framework，highwaylane-changing scenarios are used asatestbed.Experimental results show significantadvantagesoverbaselinemethodsintermsofpolicyconvergence，trainingstability，lanechangesucessrate， trajectory smoothness，and steering continuity.

Key words： Autonomous driving， Trajectory planning，Dynamic motion primitives，Deep reinforcement learning，Safety-constrained obstacle avoidance

【引用格式】修丽梅，刘元盛.基于深度强化学习与动态运动基元的自动驾驶类人轨迹规划[J].汽车技术，2025（12）：10-18. XIULM，LIUYS.Human-Like TrajectoryPlanning for AutonomousVehiclesBasedon Deep ReinforcementLearning and Dynamic MovementPrimitives[J].Automobile Technology，2O25（12）：10-18.

1前言

经典自动驾驶系统通常采用分层设计，将环境感知、行为预测、决策优化、轨迹规划及运动控制等环节分模块串行化处理，依赖精确的物理模型实现各模块功能闭环[1-3]。（剩余16531字）