Over-the-Air Edge Inference for Low-Altitude Airspace: Generative AI-Aided Multi-Task Batching and Beamforming Design

The exploitation of low-altitude (LA) airspace is advancing globally, boosting the sensing demands for heterogenous flying aircraft. To fulfill an accurate and intelligent sensing, the future 6G base station (BS) requires to aggregate features of multiple sensors’ views, then perform edge inference via loading the artificial intelligence (AI) model. However, this process confronts communication and computation bottlenecks owing to high-dimensional feature uploading as well as frequent memory access. To overcome these bottlenecks, we propose a multi-task over-the-air edge inference system for LA airspace, where feature aggregation is efficiently achieved employing over-the-air computation, and multiple inference tasks arriving at the BS are processed in batches to reduce memory access. Under this arrangement, we formulate a joint batching and beamforming design problem to maximize the number of completed tasks, constrained by completion latency and inference accuracy requirements. To address this intractable problem, we first examine the case with synchronous task arrivals and single batch. A spatial correlation-aware beamforming design approach is proposed to effectively suppress feature aggregation error and ensure inference accuracy. Next, we delve into the general asynchronous task arrival case. An AI-generated online policy is developed, which innovatively utilizes diffusion model to output batching decisions, thereby adapting to the dynamic and uncertain nature of task arrivals. Simulation results obtained on real-world dataset corroborate the importance of capturing the spatial correlation among sensors. In addition, the proposed approach realizes outstanding performance compared to benchmark batching, beamforming, and learning methods, and the completed task amount is close to offline optimization with prior task information.