Scholarlink

Autonomous navigation of Unmanned Aerial Vehicles (UAVs) in dynamic, noisy environments requires robust strategies to interpret temporal state information effectively. While Deep Q-Networks (DQN) enhanced with Long Short-Term Memory (LSTM) address temporal dependencies, they often face challenges in training stability and sample efficiency. This study proposes an enhanced LSTM-DQN framework that uniquely integrates Prioritized Experience Replay (PER) and Curriculum Learning (CL) to overcome these limitations. By combining LSTM’s temporal modeling with PER’s sample efficiency and CL’s progressive task structuring, our approach achieves superior stability and performance in a noisy 2D simulation environment. A custom reward function tailored for Gaussian noise (σ = 0.01) enables precise navigation, with the model achieving an 80% success rate (peaking at 86%), compared to a 30% success rate (peaking at 47%) for baseline LSTM-DQN models without PER and CL. We provide a comprehensive methodology, an expanded ablation study isolating the contributions of PER and CL, and qualitative trajectory analyses to demonstrate robustness. This framework establishes a scalable foundation for autonomous UAV navigation, with future directions targeting Complex environments and real-world validation.

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

A Synergistic Approach to UAV Navigation: Integrating LSTM-DQN with Prioritized Experience Replay and Curriculum Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ayad Saadi Ahmed,
  • Mohammed Obaid Mustafa

摘要

Autonomous navigation of Unmanned Aerial Vehicles (UAVs) in dynamic, noisy environments requires robust strategies to interpret temporal state information effectively. While Deep Q-Networks (DQN) enhanced with Long Short-Term Memory (LSTM) address temporal dependencies, they often face challenges in training stability and sample efficiency. This study proposes an enhanced LSTM-DQN framework that uniquely integrates Prioritized Experience Replay (PER) and Curriculum Learning (CL) to overcome these limitations. By combining LSTM’s temporal modeling with PER’s sample efficiency and CL’s progressive task structuring, our approach achieves superior stability and performance in a noisy 2D simulation environment. A custom reward function tailored for Gaussian noise (σ = 0.01) enables precise navigation, with the model achieving an 80% success rate (peaking at 86%), compared to a 30% success rate (peaking at 47%) for baseline LSTM-DQN models without PER and CL. We provide a comprehensive methodology, an expanded ablation study isolating the contributions of PER and CL, and qualitative trajectory analyses to demonstrate robustness. This framework establishes a scalable foundation for autonomous UAV navigation, with future directions targeting Complex environments and real-world validation.