Fusing Digital Twin and Real Fingerprints for Cell-Free mmWave Localization

Abstract

Ray-tracing digital twins (DTs) offer a scalable way to augment fingerprint databases for cell-free millimeter-wave (mmWave) positioning, yet the mismatch between simulated and physical propagation hinders naïve substitution. We propose DT-conditioning, a technique for cell-free mmWave fingerprint localization that treats the DT as a complementary side input fused with sparse real measurements instead of a standalone surrogate. An oracle Cramér-Rao bound (CRB) benchmark shows that DT side information strictly improves the positioning bound-most so along weak eigendirections of the real-only Fisher information matrix-yielding a 34.9% reduction on our testbed. We instantiate this view with a lightweight digital twin-conditioned (DTCond) neural architecture evaluated on a physical 28 GHz cell-free testbed with three distributed access points (APs). DTCond attains 0.257 m mean error on entirely unseen spatial regions in a leave-one-line-out protocol (48% gain over real-only baselines) and 0.186 m under a random split (41% gain), even though per-AP correlations fall as low as 0.17, whereas the evaluated domain-adaptation baselines perform worse than real-only methods.