首页 >

视线角示范的高速滑翔飞行器深度强化学习一体化控制方法-自主智能无人系统专刊

谭鸣1,曹承钰1,赵逸云1,李繁飙2,桂卫华2,陈海鹏3   

  1. 1. 中南大学自动化学院
    2. 中南大学
    3. 中国运载火箭技术研究院
  • 收稿日期:2026-06-01 修回日期:2026-10-23 出版日期:2026-10-29 发布日期:2026-10-29
  • 通讯作者: 李繁飙
  • 基金资助:
    国家自然科学基金大飞机基础研究联合基金重点支持项目;国家自然科学基金委员会青年科学基金项目

Integrated Guidance and Control for High-Speed Gliding Flight Vehicles via Deep Reinforcement Learning with Line-of-Sight Angle Demonstration

  • Received:2026-06-01 Revised:2026-10-23 Online:2026-10-29 Published:2026-10-29
  • Supported by:
    Key Program of the Joint Fund for Basic Research on Large Aircraft of the National Natural Science Foundation of China;Young Scientists Fund of the National Natural Science Foundation of China

摘要: 针对高速滑翔飞行器末段在终端落角约束、模型不确定性和外部扰动条件下的一体化控制问题,提出一种视线角示范的深度强化学习一体化控制方法。首先,建立飞行器与目标相对运动及姿态动力学耦合模型并结合终端落角约束设计任务先验信息,构造融合飞行状态、视线角误差参考剖面跟踪误差及积分误差的状态表示,将末段一体化控制任务表述为增强马尔可夫决策过程下的策略优化问题。然后,设计基于反步法的专家控制器生成示范动作,并通过模仿学习方式为深度强化学习策略提供专家先验,从而提高训练初期的探索效率。最后,通过仿真验证表明,所提方法可在满足终端落角约束的同时实现由增强状态特征到控制指令的直接策略映射,并在模型不确定性和外部扰动条件下保持较好的鲁棒性。

关键词: 高速滑翔飞行器, 一体化控制, 终端落角约束, 深度强化学习, 模仿学习

Abstract: To address the integrated guidance and control problem of high-speed gliding vehicles in dive phase under terminal impact angle constraints, model uncertainties, and external disturbances, this paper proposes a line-of-sight angle demonstration-guided deep reinforcement learning-based integrated guidance and control method. First, a coupled model of vehicle-target relative motion and attitude dynamics is established. By incorporating task prior information associated with terminal impact angle constraints, an augmented state representation is constructed by integrating flight states, LOS angle error reference profiles, tracking errors, and integral errors. The terminal integrated guidance and control problem is then formulated as a policy optimization problem under an augmented Markov decision process. Subsequently, a backstepping-based expert controller is designed to generate demonstration actions, and an imitation learning mechanism is introduced to provide expert priors for the deep reinforcement learning policy, thereby improving the exploration efficiency during the early training stage. Finally, simulation results demonstrate that the proposed method can satisfy terminal impact angle constraints while achieving policy generation from augmented state features to control commands, and maintains satisfactory robustness under model uncertainties and external disturbances.

Key words: high-speed gliding flight vehicle, integrated guidance and control, terminal impact angle constraint, deep reinforcement learning, imitation learning

中图分类号: