Scholarlink

<p>This study presents a data-driven control framework for regulating a rotary-wing unmanned aerial vehicle (UAV) using a neural ordinary differential equation (ODE) model. A neural ODE is trained to learn the UAV’s state regulation, with gradient optimization performed via automatic differentiation employing both forward and reverse modes. The neural network, featuring two fully connected hidden layers with Glorot initialization, is updated using the adaptive moment estimation method. On the other hand, the UAV is equipped with a position and attitude autopilot system. Control inputs are generated from the trained neural ODE using a 6-degree of freedom (6-DoF) dynamic model and rotor angular velocity relations. For practical implementation, the required inputs for the DC motors are computed through a Simscape-based system model. The proposed framework’s robustness is validated through numerical tests under varying initial conditions. Furthermore, an actual flight test is conducted, demonstrating the practical effectiveness and real-world applicability of the neural ODE-based regulator for rotary-wing UAV control.</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

Data-Driven Regulator of Rotary-Wing UAV Utilizing Neural Ordinary Differential Equations

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Jongho Park,
  • Mingu Kim

摘要

This study presents a data-driven control framework for regulating a rotary-wing unmanned aerial vehicle (UAV) using a neural ordinary differential equation (ODE) model. A neural ODE is trained to learn the UAV’s state regulation, with gradient optimization performed via automatic differentiation employing both forward and reverse modes. The neural network, featuring two fully connected hidden layers with Glorot initialization, is updated using the adaptive moment estimation method. On the other hand, the UAV is equipped with a position and attitude autopilot system. Control inputs are generated from the trained neural ODE using a 6-degree of freedom (6-DoF) dynamic model and rotor angular velocity relations. For practical implementation, the required inputs for the DC motors are computed through a Simscape-based system model. The proposed framework’s robustness is validated through numerical tests under varying initial conditions. Furthermore, an actual flight test is conducted, demonstrating the practical effectiveness and real-world applicability of the neural ODE-based regulator for rotary-wing UAV control.