WiC-RadGen: Dual-Band Wi-Fi Conditioned Radar Point Cloud Generation via Discrete VAEs and Masked Autoregressive Transformers

Wireless sensing technologies are crucial for a wide range of real-world applications, from indoor navigation and smart home automation to autonomous vehicle perception and industrial monitoring. While dedicated mmWave radar systems offer high-precision spatial measurements, they encounter deployment challenges due to their high cost and power consumption. In contrast, Wi-Fi-based sensing leverages existing communication infrastructure, offering cost-effective alternatives via channel state information (CSI) in 2.4 GHz systems and beamforming in mmWave Wi-Fi (60 GHz). However, traditional Wi-Fi sensing provides lower spatial fidelity than radar systems. This study introduces Wi-Fi Conditioned Radar point-cloud Generation (WiC-RadGen), which closes this performance gap by generating radar-quality point clouds from dual-band Wi-Fi signals. Our approach employs a two-stage architecture: initially, a discrete variational autoencoder with specialized loss functions (Chamfer distance, annealed KL divergence, and cluster loss) tokenizes sparse radar point clouds while preserving geometric coherence; subsequently, a conditional masked autoregressive transformer integrates 2.4 GHz CSI and 60 GHz beam SNR through cross-attention mechanisms for sequential token prediction. Technical innovations include an empty-token predictor with a differential Neyman–Pearson loss that efficiently identifies uninformative regions, focal loss to address class imbalance, and iterative unmasking for high-fidelity generation. Experimental validation demonstrates competitive localization accuracy approaching mmWave radar performance while retaining Wi-Fi’s deployment advantages. WiC-RadGen advances wireless sensing by enabling radar-precision applications through ubiquitous Wi-Fi infrastructure, reducing deployment costs without compromising spatial fidelity.