Task Offloading in Multi-UAV Edge Computing Using an Improved Multi- Agent Proximal Policy Optimization Approach

Abstract With the rapid growth of UAV-enabled applications in inspection, logistics, and emergency communication, onboard computation and energy resources have become key bottlenecks for large-scale UAV swarms. Mobile edge computing (MEC) provides a practical way to relieve this pressure by offloading computation-intensive tasks to nearby edge and cloud servers. However, task offloading in multi-UAV environments remains challenging because of partial observability, time-varying wireless conditions, resource contention, and the need for coordinated decisions among multiple agents. To address these issues, this paper studies an energy-efficient task offloading problem for a heterogeneous UAV swarm and develops a three-layer computing architecture consisting of a UAV layer, an edge server layer, and a cloud layer. Based on this architecture, we propose an MAPPO-based framework, called SS-MAPPO, which combines a top-K sparse attention mechanism with a Gaussian-based latent policy representation. The problem is formulated as a decentralized partially observable Markov decision process (Dec-POMDP), where each UAV makes offloading decisions according to its local task attributes, queue states, and observable server conditions. Simulation results show that SS-MAPPO reduces energy consumption, accelerates convergence, and improves decision stability compared with the evaluated baselines under the tested settings. These results indicate that the proposed framework is a promising approach for energy-aware collaborative offloading in dynamic multi-UAV edge computing environments.