Scholarlink

<p>This paper proposes a hybrid-action recurrent multi-agent proximal policy optimization (HA-RMAPPO) for cooperative defense against agile first-person-view intruders using passive radio frequency angle-of-arrival (AoA) sensing only. A centralized particle filter (CPF) with range-dependent bearing noise provides belief states and confidence ellipses. Under centralized training and decentralized execution, HA-RMAPPO uses a hybrid action: a discrete AoA-based sensing/interception assignment head and continuous planar velocity commands. A probabilistic threat score is defined by the earliest intersection between the confidence ellipse and the protected zone; this score triggers switching from uncertainty reduction to rapid interception. A two-phase curriculum is adopted: phase I trains basic coordination with ground-truth intruder states, and phase II switches to CPF belief states with threat-triggered control under partial observability. As an end-to-end sensing-to-interception pipeline that couples passive AoA perception with multi-agent control, the proposed system yields higher defense success with fewer breaches than independent proximal policy optimization and multi-agent proximal policy optimization in bearing-only scenarios.</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

Hybrid reinforcement learning control for UAV territory guarding with cooperative angle-of-arrival localization

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Wenjie Zhuang,
  • Longhao Qian,
  • Hugh Hongtao Liu

摘要

This paper proposes a hybrid-action recurrent multi-agent proximal policy optimization (HA-RMAPPO) for cooperative defense against agile first-person-view intruders using passive radio frequency angle-of-arrival (AoA) sensing only. A centralized particle filter (CPF) with range-dependent bearing noise provides belief states and confidence ellipses. Under centralized training and decentralized execution, HA-RMAPPO uses a hybrid action: a discrete AoA-based sensing/interception assignment head and continuous planar velocity commands. A probabilistic threat score is defined by the earliest intersection between the confidence ellipse and the protected zone; this score triggers switching from uncertainty reduction to rapid interception. A two-phase curriculum is adopted: phase I trains basic coordination with ground-truth intruder states, and phase II switches to CPF belief states with threat-triggered control under partial observability. As an end-to-end sensing-to-interception pipeline that couples passive AoA perception with multi-agent control, the proposed system yields higher defense success with fewer breaches than independent proximal policy optimization and multi-agent proximal policy optimization in bearing-only scenarios.