Unsupervised Federated Learning With Harmonized Foundation Models for FMCW Radar-Based Hand Gesture Recognition

Frequency-modulated continuous wave (FMCW) radar-based hand gesture recognition (HGR) systems face deployment challenges due to variations in radar hardware, antenna layouts, and gesture classes, which cause distributional shifts across devices. These shifts limit the effectiveness of transfer learning (TL), which, while helpful for reusing knowledge, struggles with significant hardware and configuration changes and often requires substantial labeled data from the target domain. We present a radar-specific adaptation framework that enables cross-device gesture recognition with minimal labeled data. A key component is the harmonization module, which performs signal-level transformations to align the range and Doppler dimensions of radar data across differing configurations. In parallel, radar-specific data augmentation techniques simulate missing antenna channels and gesture variability to improve pre-training robustness. A transformer-based foundation model is pre-trained on harmonized and augmented data from a source radar and then fine-tuned using a small number of labeled samples from the target configuration. The adapted model is distilled into a lightweight architecture and deployed to clients sharing the same radar setup using an unsupervised federated learning (FL) pipeline. This enables on-device model refinement using only unlabeled data. Experiments on public datasets from Infineon and Texas Instruments radars show that our method achieves over 94% accuracy with just 20 labeled samples per class, outperforming baselines by more than 10%, and converging with four times fewer communication rounds during FL.