Scholarlink

With the rapid advancement of Unmanned Aerial Vehicle (UAV) technology, multi-UAV cooperative formation tracking of dynamic targets with obstacle avoidance presents a critical research challenge. This paper proposes a novel meta-reinforcement learning (Meta-RL) algorithm named MA-PEARL, which integrates the core principles of Multi-Agent Soft Actor-Critic (MASAC) and Probabilistic Embeddings for Actor-critic Reinforcement Learning (PEARL). By incorporating a meta-task set, the multi-UAV cooperative formation system, governed by the MA-PEARL algorithm, achieves end-to-end decision-making, enabling effective obstacle avoidance and precise tracking of dynamic targets following diverse motion trajectories. This approach significantly enhances the system’s adaptability to heterogeneous tasks. Extensive simulations conducted within the ROS and Gazebo co-simulation environment validate the effectiveness and robustness of the proposed algorithm.

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

Meta-RL-Based Multi-UAV Cooperative Formation for Dynamic Target Tracking and Obstacle Avoidance

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bangsong Lei,
  • Baoguang Wang,
  • Zhaoxu Ren,
  • Yumei Yue

摘要

With the rapid advancement of Unmanned Aerial Vehicle (UAV) technology, multi-UAV cooperative formation tracking of dynamic targets with obstacle avoidance presents a critical research challenge. This paper proposes a novel meta-reinforcement learning (Meta-RL) algorithm named MA-PEARL, which integrates the core principles of Multi-Agent Soft Actor-Critic (MASAC) and Probabilistic Embeddings for Actor-critic Reinforcement Learning (PEARL). By incorporating a meta-task set, the multi-UAV cooperative formation system, governed by the MA-PEARL algorithm, achieves end-to-end decision-making, enabling effective obstacle avoidance and precise tracking of dynamic targets following diverse motion trajectories. This approach significantly enhances the system’s adaptability to heterogeneous tasks. Extensive simulations conducted within the ROS and Gazebo co-simulation environment validate the effectiveness and robustness of the proposed algorithm.