• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 212-223.doi: 10.3901/JME.260617

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

基于深度强化学习的海上飞机自动加油对接方法研究

冷豪, 王巍, 蔡月日   

  1. 北京航空航天大学机器人研究所 北京 100191
  • 收稿日期:2025-12-23 修回日期:2026-05-14 发布日期:2026-11-20
  • 作者简介:冷豪,男,2002年出生。主要研究方向为强化学习下的轴孔装配问题,航空加油机器人设计。E-mail:sy2407702@buaa.edu.cn;王巍,男,1973年出生,博士,教授,博士、硕士研究生导师。主要研究方向为机器人学与自动化领域机器人设计。E-mail:wangweilab@buaa.edu.cn;蔡月日(通信作者),男,1983年出生,博士,副教授,硕士研究生导师。主要研究方向为机器人设计制造。E-mail:caiyueri@buaa.edu.cn
  • 基金资助:
    中央高校基本科研业务费资助项目(JK2024-50-02)。

Research on Automatic Refueling Docking Method of Offshore Aircraft Based on Deep Reinforcement Learning

LENG Hao, WANG Wei, CAI Yueri   

  1. Institute of Robotics, Beihang University, Beijing 100191
  • Received:2025-12-23 Revised:2026-05-14 Published:2026-11-20

摘要: 针对海上因海浪引发的舰船摇摆导致飞机受油口偏移,传统控制方法难以同时应对高定位误差和随机力扰动的问题,提出一种基于深度强化学习的海上飞机自动加油对接方法。本文通过建立海况环境下的舰船运动模型,明确因舰船摇摆导致的机械臂末端的位置与姿态误差边界,构建以多维感知信息为状态空间、以末端位姿增量为动作空间的强化学习框架,并结合模糊控制理论设计多维度奖励函数,以此引导系统实现柔顺对接。本文设计了仿真迁移策略,采用域随机化技术模拟物理参数与观测噪声,基于实测数据对油管扰动力进行精确建模,并完成力传感器重力补偿,增强虚实环境的一致性。鉴于海上环境的受油口偏移特性,使用基于模糊控制器的深度强化学习算法对策略进行训练,算法经仿真平台测试后迁移至实物样机完成验证。结果表明,所训练的策略性能良好,可以满足海上飞机加油场景的需求。

关键词: 自动加油, 深度强化学习, 模糊控制, 仿真迁移

Abstract: Aiming at the problem that the aircraft refueling port deviates due to ship sway caused by ocean waves, and traditional control methods are difficult to cope with high positioning errors and random force disturbances simultaneously, this paper proposes an automatic refueling docking method for offshore aircraft based on deep reinforcement learning. A ship motion model under sea state conditions is established to clarify the position and attitude error boundaries of the manipulator end-effector caused by the ship sway. A reinforcement learning framework is constructed, taking multi-dimensional perception information as the state space and the end-effector pose increment as the action space. Additionally, a multi-dimensional reward function is designed based on fuzzy control theory to guide the system in achieving compliant docking. A simulation-to-real transfer strategy is designed, utilizing domain randomization technology to simulate physical parameters and observation noise. Based on measured data, the disturbance force of the oil pipeline is accurately modeled, and the gravity compensation of the force sensor is completed, thereby enhancing the consistency between the simulation and real environments. In view of the deviation characteristics of the refueling port in the marine environment, a deep reinforcement learning algorithm based on a fuzzy controller is adopted to train the strategy. The algorithm is tested on a simulation platform and then transferred to a physical prototype for verification. The results show that the trained strategy has excellent performance and can meet the requirements of offshore aircraft refueling scenarios.

Key words: automatic refueling, deep reinforcement learning, fuzzy control, simulation-to-real transfer

中图分类号: