Scholarlink

Fixed-wing unmanned aerial vehicles (UAVs) exhibit significant advantages in large-scale environmental sensor data collection. In forest or marine application scenarios, UAVs take off from a fixed location, traverse the sensors to collect data, and then return. This work investigates scenarios in which the UAV’s initial energy is insufficient to visit all nodes, aiming to optimize energy efficiency while simultaneously collecting data during the return flight. To address this problem, we introduce staged actor-critic reinforcement learning (S-ACL) to alleviate the difficulty of convergence and the suboptimal performance of sparse binary rewards. Based on it, we employ three reward functions within the twin delayed deep deterministic policy gradient (TD3) reinforcement learning framework and propose a stage-based safe action algorithm, staged safe-action TD3(SS-TD3). We establish an energy consumption model that accounts for acceleration and use a piecewise continuous-time flight model to enhance exploration efficiency. Experimental results demonstrate that, SS-TD3 achieves optimal energy efficiency comparing to baseline approaches.

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

Trajectory Design for Data Collection Under Insufficient UAV Energy: A Staged Actor-Critic Reinforcement Learning Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuejia Zhang,
  • Jing Mei,
  • Zhao Tong,
  • Can Wang,
  • Keqin Li

摘要

Fixed-wing unmanned aerial vehicles (UAVs) exhibit significant advantages in large-scale environmental sensor data collection. In forest or marine application scenarios, UAVs take off from a fixed location, traverse the sensors to collect data, and then return. This work investigates scenarios in which the UAV’s initial energy is insufficient to visit all nodes, aiming to optimize energy efficiency while simultaneously collecting data during the return flight. To address this problem, we introduce staged actor-critic reinforcement learning (S-ACL) to alleviate the difficulty of convergence and the suboptimal performance of sparse binary rewards. Based on it, we employ three reward functions within the twin delayed deep deterministic policy gradient (TD3) reinforcement learning framework and propose a stage-based safe action algorithm, staged safe-action TD3(SS-TD3). We establish an energy consumption model that accounts for acceleration and use a piecewise continuous-time flight model to enhance exploration efficiency. Experimental results demonstrate that, SS-TD3 achieves optimal energy efficiency comparing to baseline approaches.