Edge Intelligence for IoT Terminal Coordination: A Low-Power SoC-Embedded GNN Framework for UAV Swarm Task Allocation in Smart City Environments

Unmanned aerial vehicle (UAV) swarms serve as critical Internet of Things (IoT) terminal devices in smart city applications, including urban traffic management, logistics, and disaster response. However, dynamic task allocation under energy constraints and environmental uncertainties remains a key challenge for edge intelligence. This article proposes a novel SoC-embedded self-learning heterogeneous graph neural network (SHGNN) framework to address these issues. On the hardware front, a low-power development board centered on the Rockchip RK3588 SoC is developed. The RK3588 adopts an 8-nm process technology and integrates a neural processing unit, enabling adaptive power management and high computational efficiency for multitask IoT scenarios. On the software side, the SHGNN algorithm models UAVs, tasks, and environmental factors as a heterogeneous graph, incorporating a node-learning module and self-learning attention mechanism to dynamically adjust allocations in real time, even under unpredictable conditions such as traffic congestion or emergent threats. Validated through simulations inspired by real-world datasets, the framework significantly outperforms traditional methods, reducing task completion time (TCT) by 22%, achieving 94.1% allocation accuracy (AA), and improving energy efficiency (EE) by 30% while maintaining ultralow power consumption (5 W). This work demonstrates a practical hardware-software co-design for scalable and adaptive IoT edge systems, offering robust support for UAV swarm operations in dynamic environments.