Poster: Generalizing mmWave Sensing with Simulation Synthesis and Generative Models

Wireless sensing through emerging millimeter-wave (mmWave) technology has gained traction recently, integrating extensively into human life and transitioning from research prototypes to standardization and commercialization, due to the high angular and range resolution provided by large antenna arrays and extensive sampling bandwidth. Current applications showcase vision-like capabilities such as gesture and posture tracking, person reidentification, and point cloud generation, often powered by deep learning models, moving beyond traditional uses for distance and speed detection. However, generalization remains a key challenge since RF signals, with short wavelengths and coherent multipath reflections, prove extremely sensitive to minor target changes that often disrupt deep learning models applied to unfamiliar data, and while collecting extensive datasets could mitigate this issue, it introduces separate challenges regarding specialized and expensive equipment requirements that limit everyday feasibility along with data specificity to particular radar configurations that complicates transferability between contexts, further exacerbating the generalization issue. Employing simulators has succeeded in computer vision and graphics but remains less explored for RF data synthesis which, to attain detailed scene information like object geometry, motion, material properties, and environmental context, involves costly equipment and labor-intensive efforts while lacking precision that widens the gap between simulation and reality.