Tsukada Lab. Computer Network Group

Trajectory Planning and Task Scheduling for UAVs in Dynamic Environments Using Feature-Driven Multi-Agent Reinforcement Learning

Quanxi Zhou, Wencan Mao, Yilei Liang, Yixuan Liu, Manabu Tsukada, Jon Crowcroft, John C. S. Lui, Yusheng Ji

IEEE Transactions on Mobile Computing, 2026年10月31日

3行要約

  • UAVによる物資配送の軌道と、地上端末に提供するエッジ計算の割り当てを同時に求める。
  • グラフまたは点群で表した環境特徴を、マルチエージェントの制御に取り込む。
  • 未学習のシミュレーション地図では、両方式ともMASACより完了時間、衝突率、計算遅延を改善した。

論文の図

図 1: 空中配送とエッジ計算の協調
図 1. 空中配送とエッジ計算の協調出典: Zhou et al., IEEE Transactions on Mobile Computing (2026)(著者原稿の図版) の図 1UAVが物資を配送しながら、基地局の支援を受けて地上端末や歩行者端末に計算を提供する構成を示す。

アブストラクト(原文)

Unmanned Aerial Vehicles (UAVs) can simultaneously perform material delivery over the air and provide MEC service to ground Internet of Things (IoT) devices. However, most existing research considers trajectory planning and task scheduling problems separately in homogeneous environments. In addition, the computational complexity and dynamically changing environmental factors, including building distributions, pedestrian trajectories, IoT device positions, and UAV start points and destinations, add to the complexity of the problems. To jointly optimize UAV trajectory planning and task scheduling with generalization ability, two novel Multi-Agent Reinforcement Learning (MARL) algorithms are proposed: Graph Neural Network-based Feature-Driven (GNNFD)-MARL and Point Cloud Network-based Feature-Driven (PCNFD)-MARL. GNNFD-MARL establishes a connectivity graph between UAVs, ground users, IoT devices, and base stations to guide model training. Meanwhile, PCNFD-MARL utilizes point cloud features derived from the varying distribution of users and IoT devices in different areas and times to assist joint decisions. Both approaches extract environmental features to generate self-adaptive policies suitable for diverse scenarios. Furthermore, the MARL frameworks based on actor-critic architecture are enhanced by improving both actor and critic networks. Experimental results demonstrate that both algorithms exhibit similar convergence properties and task performance, outperforming existing MARL methods. Furthermore, they exhibit distinct time efficiencies under different settings.

引用方法

Q. Zhou, W. Mao, Y. Liang, Y. Liu, M. Tsukada, J. Crowcroft, J. C. S. Lui, and Y. Ji, “Trajectory Planning and Task Scheduling for UAVs in Dynamic Environments Using Feature-Driven Multi-Agent Reinforcement Learning,” IEEE Transactions on Mobile Computing, 2026, doi: 10.1109/TMC.2026.3750340.

BibTeX

@article{Zhou2026e,
  title         = {Trajectory Planning and Task Scheduling for UAVs in Dynamic Environments Using Feature-Driven Multi-Agent Reinforcement Learning},
  author        = {Quanxi Zhou and Wencan Mao and Yilei Liang and Yixuan Liu and Manabu Tsukada and Jon Crowcroft and John C. S. Lui and Yusheng Ji},
  doi           = {10.1109/TMC.2026.3750340},
  year          = {2026},
  date          = {2026-10-31},
  urldate       = {2026-10-31},
  journal       = {IEEE Transactions on Mobile Computing},
  abstract      = {Unmanned Aerial Vehicles (UAVs) can simultaneously perform material delivery over the air and provide MEC service to ground Internet of Things (IoT) devices. However, most existing research considers trajectory planning and task scheduling problems separately in homogeneous environments. In addition, the computational complexity and dynamically changing environmental factors, including building distributions, pedestrian trajectories, IoT device positions, and UAV start points and destinations, add to the complexity of the problems. To jointly optimize UAV trajectory planning and task scheduling with generalization ability, two novel Multi-Agent Reinforcement Learning (MARL) algorithms are proposed: Graph Neural Network-based Feature-Driven (GNNFD)-MARL and Point Cloud Network-based Feature-Driven (PCNFD)-MARL. GNNFD-MARL establishes a connectivity graph between UAVs, ground users, IoT devices, and base stations to guide model training. Meanwhile, PCNFD-MARL utilizes point cloud features derived from the varying distribution of users and IoT devices in different areas and times to assist joint decisions. Both approaches extract environmental features to generate self-adaptive policies suitable for diverse scenarios. Furthermore, the MARL frameworks based on actor-critic architecture are enhanced by improving both actor and critic networks. Experimental results demonstrate that both algorithms exhibit similar convergence properties and task performance, outperforming existing MARL methods. Furthermore, they exhibit distinct time efficiencies under different settings.},
  keywords      = {Other},
  pubstate      = {published},
  tppubtype     = {article}
}