The rapid growth of heterogeneous Internet of Things (IoT) applications imposes stringent requirements on computation latency, service reliability, and energy efficiency. Unmanned aerial vehicles (UAVs) equipped with computing and communication capabilities can provide flexible coverage and computation support; however, dynamic network conditions, limited onboard resources, and heterogeneous service demands make computation offloading challenging. This paper proposes a service-aware multi-agent proximal policy optimization (SA-MAPPO) framework for computation offloading in UAV-assisted multi-access edge computing (MEC)-enabled IoT systems. We consider a hierarchical device–UAV–MEC–cloud architecture in which multiple UAVs collaboratively determine device association, service placement, task partitioning, and offloading decisions. The problem is formulated as a decentralized partially observable Markov decision process to account for dynamic task arrivals, service availability, communication conditions, computing capacity, task deadlines, and UAV energy consumption. Using centralized training and decentralized execution, SA-MAPPO enables cooperative UAV decision-making based on local observations while exploiting global information during training. Simulation results demonstrate that the proposed approach consistently outperforms the considered baseline methods in terms of task success ratio and energy efficiency under different IoT-device densities. These results confirm the effectiveness of the proposed scheme in supporting reliable and energy-efficient computation offloading for dynamic UAV-assisted MEC-enabled IoT systems. 

Computation offloading,mobile edge computing (MEC),MAPPO,MADRL,UAVs,IoT Systems