Geometric Feature-Aware Trajectory and Offloading Joint Optimization for Multi-UAV Networks

Abstract

This paper investigates joint trajectory planning and computation offloading in multi-UAV assisted mobile edge computing (MEC) networks within complex 3D urban environments. Existing deep multi-agent reinforcement learning (MARL) methods often oversimplify obstacles as geometric no-fly zones, ignoring fine-grained channel fading effects. To address this, we propose a novel geometric feature-aware trajectory optimization framework, which explicitly extracts key obstacle features to quantify non-line-of-sight (NLoS) risks and develops a feature and geometric attention-based multi-agent deep deterministic policy gradient (FGA-MADDPG) algorithm. By incorporating a geometric-feature attention mechanism, FGA-MADDPG effectively prioritizes critical environmental feature to mitigate state space explosion. Extensive simulations demonstrate that our approach outperforms baselines by learning proactive detour strategies to maintain LoS links. Notably, robustness analysis confirms that FGA-MADDPG maintains optimal communication rates even under increasing environmental complexity.