Towards Building a Foundation Model for Wireless Sensing: A Pilot Study with FMCW Radar Sensor

This paper presents a pilot study toward a foundation model for wireless sensing using FMCW radar. We propose a transformer-based architecture trained on data from Infineon’s BGT60TR13C mmWave sensor with self-supervised objectives designed to capture temporal–spatial signal characteristics without labels. Our framework introduces strategies for handling sparse channel representations and provides a unified normalization across diverse radar configurations, enabling task-agnostic representation learning. Experimental evaluation on presence detection shows that the learned embeddings generalize effectively and achieve competitive performance compared to pretrained models such as DINO and LWM. These results validate the feasibility of RF foundation models and highlight their potential to advance physical AI through adaptable and label-efficient wireless sensing systems.