Scholarlink

<p>Unmanned aerial vehicles (UAVs) have recently been used for various purposes, necessitating autonomous flight and landing technologies. For stable operation in diverse environments, UAVs must be capable of landing not only on flat terrain but also on sloped or wave-disturbed surfaces. In this study, a mobile robot with a leveling function and reinforcement learning was used to enable a UAV to land stably and quickly, even on sloped terrain. The Deep Deterministic Policy Gradient (DDPG) algorithm was used as the reinforcement learning method, and a data sampling strategy was applied to separate the replay buffer based on specific criteria to enhance the learning performance. A 3D simulation environment with a physics engine was created to model the UAV and the target object, and a reference marker was attached to the target to enable UAV recognition using image recognition techniques. The data exchange process between the simulation model, reinforcement learning algorithm, and environment was established using Robot Operating System (ROS) to learn the autonomous landing of moving targets on various slopes. The results showed that this method achieved faster landing success than existing methods that use visual servoing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Autonomous Landing of Unmanned Aerial Vehicles on Sloped Terrain Using Deep Reinforcement Learning with Separated Replay Buffer

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Junsoo Baek,
  • Junyoung Kwak,
  • Hyunbin Park,
  • Baeksuk Chu

摘要

Unmanned aerial vehicles (UAVs) have recently been used for various purposes, necessitating autonomous flight and landing technologies. For stable operation in diverse environments, UAVs must be capable of landing not only on flat terrain but also on sloped or wave-disturbed surfaces. In this study, a mobile robot with a leveling function and reinforcement learning was used to enable a UAV to land stably and quickly, even on sloped terrain. The Deep Deterministic Policy Gradient (DDPG) algorithm was used as the reinforcement learning method, and a data sampling strategy was applied to separate the replay buffer based on specific criteria to enhance the learning performance. A 3D simulation environment with a physics engine was created to model the UAV and the target object, and a reference marker was attached to the target to enable UAV recognition using image recognition techniques. The data exchange process between the simulation model, reinforcement learning algorithm, and environment was established using Robot Operating System (ROS) to learn the autonomous landing of moving targets on various slopes. The results showed that this method achieved faster landing success than existing methods that use visual servoing.