Efficient AI Inference Offloading for UAVs via Distributed Auctions in Dynamic Edge Environments

Unmanned aerial vehicles (UAVs) are increasingly utilized in applications such as traffic monitoring, environmental sensing and disaster response, all of which demand real-time Artificial Intelligence (AI) inference for timely and effective analysis and decision-making. However, limited onboard computation and energy constraints hinder their ability to perform complex, latency-sensitive tasks. While collaborative inference through UAV swarms or cloud offloading can alleviate local resource limitations, these methods often suffer from high latency, unstable communication or excessive energy consumption, making them unsuitable for dynamic and time-critical scenarios. Edge computing offers a promising alternative by bringing computation closer to UAVs, but existing approaches generally assume stable environments and struggle in highly dynamic multi-UAV systems with frequently changing topologies and resource availability. To tackle these challenges, we propose an adaptive AI inference offloading framework tailored for dynamic UAV environments. Our framework leverages a distributed auction-based mechanism that enables UAVs and edge servers to autonomously negotiate task offloading and resource allocation based on real-time mobility, link quality and resource availability. This decentralized approach ensures fair and efficient utilization of edge resources, reduces task delays and enhances system scalability and robustness. Extensive simulations demonstrate that our method significantly improves task success rate and reduces inference latency under dynamic conditions, outperforming existing baseline strategies.