Scholarlink

In tackling the challenge of multi-agent, multi-target encirclement by multiple UAVs within a confined space, this paper extends a deep reinforcement learning (DRL) framework under a centralized-training, decentralized-execution (CTDE) paradigm to propose a stage-wise reward and multi-task–enhanced variant of the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. First, we decompose the overall mission into sequential stages, assigning continuous rewards to each stage to guide the encircling UAVs through prerequisite subtasks before progressively driving them toward final target capture—thereby alleviating reward sparsity. Next, we introduce a greedy-algorithm–based target-assignment mechanism and develop an artificial-potential-field–inspired escape strategy for the target UAVs. Simulation results demonstrate that our method converges reliably, produces robust policies, and successfully executes coordinated multi-UAV encirclement of multiple targets in the presence of dynamic obstacles. Compared with standard MADDPG and MASAC implementations lacking stage-wise rewards and multi-task incentives, our approach achieves faster convergence, shorter capture times, and higher success rates.

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–Based Cooperative Encirclement Strategy for Multi-UAVs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wenrui Ma,
  • Wang He,
  • Wenwu Yu

摘要

In tackling the challenge of multi-agent, multi-target encirclement by multiple UAVs within a confined space, this paper extends a deep reinforcement learning (DRL) framework under a centralized-training, decentralized-execution (CTDE) paradigm to propose a stage-wise reward and multi-task–enhanced variant of the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. First, we decompose the overall mission into sequential stages, assigning continuous rewards to each stage to guide the encircling UAVs through prerequisite subtasks before progressively driving them toward final target capture—thereby alleviating reward sparsity. Next, we introduce a greedy-algorithm–based target-assignment mechanism and develop an artificial-potential-field–inspired escape strategy for the target UAVs. Simulation results demonstrate that our method converges reliably, produces robust policies, and successfully executes coordinated multi-UAV encirclement of multiple targets in the presence of dynamic obstacles. Compared with standard MADDPG and MASAC implementations lacking stage-wise rewards and multi-task incentives, our approach achieves faster convergence, shorter capture times, and higher success rates.