Edge Sensing for Efficient UAV Beam Selection: A Semantic-Driven Distributed Learning Framework

The evolution towards sixth-generation (6 G) communication networks hinges on the seamless fusion of communication and sensing, where edge sensing emerges as a critical paradigm for reliable uncrewed aerial vehicle (UAV) networks. This paper proposes a novel edge sensing framework for real-time, UAV beam selection that embodies this paradigm. Our approach leverages task-oriented intelligence at the network edge, where each base station (BS) acts as an intelligent edge sensor, and with use a co-located camera to perceive the UAV's visual context. A lightweight pre-trained network is deployed at each BS to extract the compact semantic features relevant to the beam selection task, thus avoiding the transmission of raw data or bulky models These features are then aggregated at a central server to train a global beam selection model. The proposed design transforms the complex beamforming problem into an efficient, sensing-driven classification task. Experimental results demonstrate that our framework achieves comparable accuracy compared to conventional federated learning benchmarks, while reducing the communication overhead and model size by 90 % and 94 %, respectively. The proposed framework can provide a concrete and efficient blue print for implementing task-oriented intelligent in future UAV networks.