Demo: A Material-Aware RF Digital Twin for mmWave Object Recognition with No Real-World Labels

Abstract

Millimeter-wave (mmWave) radar perception is bottlenecked by training data: every room reflects mmWave differently, so models do not transfer and labeling radar in each new space is impractical. We demonstrate a material-aware RF digital twin for object recognition with no real radar labels. From a 3D scan the twin infers per-surface electromagnetic materials and ray-traces frequency-modulated continuous-wave (FMCW) radar, and because the scene is known, each synthetic signature is automatically labeled and supervision is free. Our key finding: matching a simulated signal to real radar fails, but a scale-free geometry factor, shape read from how the reflection modulates over one rotation, transfers, because it divides out the absolute terms that do not survive the sim-to-real gap. In the live demo, attendees place household objects on a turntable before a real 77 GHz radar and the factor separates them by shape class, symmetric versus irregular, not by instance identity.