Scholarlink

<p>Loitering Munition (LM) swarms pose a growing challenge to modern air-defense systems. A promising strategy is to utilize Loyal Wingmen (LWs) as a counter-UAV system (c-UAV) in a scenario defined as Cooperative Threat Engagement with Heterogeneous Drone Swarms (CTEDS). Unlike prior work, which optimizes policies based on simplified observations of positions and velocities, this study enhances situational awareness by integrating LiDAR perception, temporal history, and shared observations from allies. To handle this complexity, we propose a two-stage pipeline that combines supervised pre-training with fine-tuning using reinforcement learning (RL). Results show that the architecture extracts meaningful policies from high-dimensional and dynamic observations. This supports the feasibility of expanding perception as a foundation for scalable cooperative autonomy in CTEDS. The approach highlights the potential of graph attention networks for cooperative air combat and points to future research on hierarchical neural policies and multi-agent reinforcement learning with communication.</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

Deep Reinforcement Learning with Graph Attention for Cooperative Loyal Wingmen under Shared Spatio-Temporal Observations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月17日

查看全文
  • Davi Guanabara de Aragão,
  • Marcos R. O. A. Maximo,
  • José Fernando Basso Brancalion

摘要

Loitering Munition (LM) swarms pose a growing challenge to modern air-defense systems. A promising strategy is to utilize Loyal Wingmen (LWs) as a counter-UAV system (c-UAV) in a scenario defined as Cooperative Threat Engagement with Heterogeneous Drone Swarms (CTEDS). Unlike prior work, which optimizes policies based on simplified observations of positions and velocities, this study enhances situational awareness by integrating LiDAR perception, temporal history, and shared observations from allies. To handle this complexity, we propose a two-stage pipeline that combines supervised pre-training with fine-tuning using reinforcement learning (RL). Results show that the architecture extracts meaningful policies from high-dimensional and dynamic observations. This supports the feasibility of expanding perception as a foundation for scalable cooperative autonomy in CTEDS. The approach highlights the potential of graph attention networks for cooperative air combat and points to future research on hierarchical neural policies and multi-agent reinforcement learning with communication.