Scholarlink

<p>To meet the requirement of simultaneous arrival for multiple hypersonic glide vehicles (HGVs), we propose a time control entry guidance (TCEG) method leveraging deep reinforcement learning. First, the entry guidance problem is solved with a reinforcement learning framework based on a designed reference flight profile. By appropriately designing the observation space and training environment, the well-trained agent demonstrates robust guidance performance under varying widths of the heading error corridor. Then, a novel method for predicting the remaining flight time is established, which consists of two main components. The first component estimates the remaining flight time using an analytical formula, while the second component employs a deep neural network (DNN) to predict the residual error between the estimated and the true value. Subsequently, based on the predicted terminal time error, the threshold of the heading error and the observation vector are corrected in real time, thereby guiding the agent to dynamically adjust its output actions. This enables precise control of the terminal time. Since the generation of guidance commands only requires forward computations by the neural network, the proposed method exhibits excellent real-time performance. Finally, the effectiveness and robustness of the method are demonstrated through numerical simulations in various 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

Time control entry guidance method for hypersonic glide vehicles based on deep reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Zhenyu Liu,
  • Gang Lei,
  • Yong Xian,
  • Leliang Ren,
  • Shaopeng Li,
  • Daqiao Zhang

摘要

To meet the requirement of simultaneous arrival for multiple hypersonic glide vehicles (HGVs), we propose a time control entry guidance (TCEG) method leveraging deep reinforcement learning. First, the entry guidance problem is solved with a reinforcement learning framework based on a designed reference flight profile. By appropriately designing the observation space and training environment, the well-trained agent demonstrates robust guidance performance under varying widths of the heading error corridor. Then, a novel method for predicting the remaining flight time is established, which consists of two main components. The first component estimates the remaining flight time using an analytical formula, while the second component employs a deep neural network (DNN) to predict the residual error between the estimated and the true value. Subsequently, based on the predicted terminal time error, the threshold of the heading error and the observation vector are corrected in real time, thereby guiding the agent to dynamically adjust its output actions. This enables precise control of the terminal time. Since the generation of guidance commands only requires forward computations by the neural network, the proposed method exhibits excellent real-time performance. Finally, the effectiveness and robustness of the method are demonstrated through numerical simulations in various scenarios.