Bridging the sim-to-real gap in RF localization with large-scale synthetic pretraining

• First systematic study quantifying the synthetic-to-real generalization gap in RF fingerprinting localization. • Novel Gaussian Process calibration method significantly improves alignment between simulated and real base station parameters. • Large-scale synthetic pretraining reduces real-world localization error by 50 • Demonstrated that simulation fidelity outweighs dataset size: calibrated synthetic data outperforms larger uncalibrated datasets. Radio frequency (RF) fingerprinting is a promising localization technique for GPS-denied environments, yet it tends to suffer from a fundamental limitation: Poor generalization to previously unmapped areas. Traditional methods such as k -nearest neighbors ( k -NN) perform well where data is available but may fail on unseen streets, limiting real-world deployment. Deep learning (DL) offers potential remedies by learning spatial-RF patterns that generalize, but requires far more training data than what simple real-world measurement campaigns can provide. In this paper, we investigate whether synthetic data can bridge this generalization gap. Using (i) a real-world dataset from Rome and (ii) NVIDIA’s open-source ray-tracing simulator Sionna, we generate synthetic datasets under varying realism and scale conditions. Specifically, we use Dataset A containing real-world measurements with real base stations (BS) and real signals, and create Dataset B using real BS locations but simulated signals, Dataset C with both simulated BS locations and signals, and Dataset B’ which represents an optimized version of Dataset B where BS parameters are calibrated via Gaussian Process to maximize signal correlation with Dataset A. Our evaluation reveals a pronounced sim-to-real gap: Models achieving 25m error on synthetic data degrade to 184m on real data. Nonetheless, pretraining on synthetic data reduces real-world localization error from 323m to 162m; a 50% improvement over real-only training. Notably, simulation fidelity proves more important than scale: A smaller calibrated dataset (53K samples) outperforms a larger uncalibrated one (274K samples). To further evaluate the generalization capabilities of the models, we conduct experiments on an unseen geographical region using a real-world dataset from Oslo. In the zero-shot setting, the models achieve a root mean square error (RMSE) of 132.2m on the entire dataset, and 61.5m on unseen streets after fine-tuning on Oslo data. While challenges remain before meeting more practical localization accuracy, this work provides a systematic study in the field of wireless communication of synthetic-to-real transfer in RF localization and highlights the value of simulation-aware pretraining for generalizing DL models to real-world scenarios.