One-Shot Near-Field Localization with AI-Optimized Hybrid Beamformer Design

This paper introduces a learning-based approach to the near-field source localization problem adopting a hybrid analog-digital beamformer in an extremely large-scale multipleinput multiple-output (XL-MIMO) system. Hybrid analog-digital architectures gained significant attention in the literature due to the limited number of radio-frequency (RF) chains. However, the investigation of effective techniques tailored to partially connected hybrid beamformers for near-field localization is still missing in the literature. To this end, we leverage a Convolutional Neural Network (CNN)-based model to: (i) perform the analog beamformer design with proper training constraints; and (ii) estimate the single-user near-field position in a single snapshot. In the inference stage, the model is divided into two parts: the first accounts for the beamformer design, and the second acts as a localizing function. Simulation results demonstrate superior performance of the proposed method over existing solutions and robustness in multipath propagation conditions. In addition, our network is scalable and requires fewer RF chains than fullyconnected architectures.