On the Physical Plausibility and Distribution Alignment for Sim-to-Real RF Positioning

Jul 16, 2026

Upload by Violet

Reliable radio frequency (RF) positioning from cellular measurements is constrained by the high cost and limited coverage of real drive-test data, especially when models must operate on streets unseen during training. The work studies how simulation design choices affect sim-to-real transfer: base-station calibration, physical realism, synthetic-data scale, and RSSI distribution alignment. Using Sionna-based reconstruction of a Rome deployment, base stations are calibrated by adjusting location, height, azimuth, and transmit power. Unconstrained calibration better fits measured RSSI yet does not consistently improve accuracy. Synthetic pretraining helps on known streets; unseen-street gains require RSSI normalization, indicating distribution alignment dominates physical realism and scale.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

Fingerprint-based RF positioning challenges

Simulation for synthetic supervision

Base-station calibration regimes

Scale versus quality

Key takeaways

On the Physical Plausibility and Distribution Alignment for Sim-to-Real RF Positioning Ararat Saribekyan∗†§ , Armen Manukyan∗†, Hrant Khachatrian∗†, Theofanis P. Raptis‡ ∗ Yerevan State University, Yerevan, Armenia. †YerevaNN, Yerevan, Armenia. Email: {ararat, armen, [hrant](hrant}@yerevann.com)[}](hrant}@yerevann.com)[@yerevann.com](hrant}@yerevann.com) § American University of Armenia, Armenia. ‡Institute of Informatics and Telematics, National Research Council, Pisa, Italy. Email: theofanis.raptis@iit.cnr.it arXiv :2607 .04400v 1 [ cs .NI ] 5 Jul 2026 Abstract—Reliable radio frequency (RF) positioning from cellular measurements is limited by the high cost and limited coverage of real drive-test data, especially when models must work on streets not seen during training. Previous work showed that ray tracing simulations can provide useful synthetic data for pretraining deep positioning models. In this paper, we focus on the simulation side and study how base-station calibration, physical realism, synthetic-data scale, and RSSI distribution alignment affect transfer to real data. Using a Sionna reconstruction of a Rome deployment, we calibrate each base station by adjusting its location, height, azimuth, and transmit power. We compare physically plausible calibrations with unconstrained ones that allow unrealistic base-station placements. We also compare deployment-specific synthetic data with much larger city-scale datasets. Although unconstrained calibration matches measured RSSI better, it does not always improve positioning accuracy. All synthetic pretraining approaches improve performance on known streets, with the best result obtained using city-scale unconstrained data. However, larger synthetic datasets alone do not improve performance on unseen streets. The best results on held-out streets are achieved only after normalizing simulated RSSI values to better match the real distribution. Overall, the results suggest that distribution alignment is more important than physical realism or dataset size for sim-to-real RF positioning. Index Terms—RF positioning, sim-to-real transfer, synthetic data, Sionna RT, RSSI, base-station calibration I. INTRODUCTION RF positioning is attractive for outdoor localization because radio fingerprints can be available in many deployed cellular and wireless networks, including settings where GPS or vision may be unreliable or expensive [1] . In fingerprint-based positioning, a model receives signal measurements from nearby base stations and predicts the receiver position. The practical obstacle is that real drive-test measurements are sparse, costly to collect, and tied to a particular deployment. For example, in the Rome dataset studied in [2], [3], a real-only positioning model can still lead to errors even on the known-street split, and generalization to held-out streets is substantially harder. Synthetic radio propagation offers a natural way to increase supervision without collecting new drive-test data [4] . Prior work [5] on this Rome dataset introduced the A/B/B’/Cdataset hierarchy, the MapRadioFormer+ positioning backbone, Sionna RT simulation, Gaussian-process base-station calibration, and large-scale synthetic pretraining for sim-toreal RF positioning. We build on that setting by focusing on the simulation pipeline itself: How should the base stations be calibrated, what is the importance of physical plausibility, when does scale help, and how much does RSSI distribution alignment matter? In [5], the synthetic dataset optimized base-station position, height, and azimuth to improve rank correlation with measured RSSI. In this work, we extend that calibration by adding transmit power, which should not be constant across antennas, optimizing a log-domain RSSI RMSE objective. We compare two regimes: • Constrained: base-station locations, orientations, and powers are optimized under physical plausibility constraints, such as keeping base stations on buildings a

FAQs

为什么仅依赖真实驱动测试数据难以支持可靠的RF定位？

真实驱动测试数据采集成本高且覆盖有限，且通常绑定特定部署。当模型需要在训练期间未见过的街区/道路上工作时，泛化会显著更难。

论文比较了哪些基站校准方式，它们的主要区别是什么？

对比了“受约束校准”和“非约束校准”。受约束方式在物理可行性约束下优化基站位置、朝向与功率；非约束方式不加这些约束，允许出现物理上不现实但可能更贴合任务的“有效”参数。

哪些因素对sim-to-real RF定位的效果最关键？

尽管非约束校准能更好地拟合实测RSSI，但不一定带来更好的定位精度。结果表明，合成与真实RSSI分布的一致性（分布对齐）比物理合理性或单纯的合成数据规模更关键，尤其是在未见街区上提升时需要对模拟RSSI进行归一化以更好匹配真实分布。

Jul 16, 2026

Upload by Violet