RAIL-5G: A Challenging Real-World Dataset and Benchmark for AI-Based 5G Indoor Localization

Jonas Pirkl; Andreas Eidloth; Maximilian Kasparek; Tobias Feigl; Christopher Mutschler

eess.SPOct 18, 2026

RAIL-5G: A Challenging Real-World Dataset and Benchmark for AI-Based 5G Indoor Localization

Authors: Jonas Pirkl, Andreas Eidloth, Maximilian Kasparek, Tobias Feigl, Christopher Mutschler

Organizations: Fraunhofer Institute for Integrated Circuits (IIS), Division Positioning and Networks, 90411 Nürnberg, Germany · Department Computer Science & Artificial Intelligence, Technische Universität Nürnberg, 90461 Nürnberg, Germany · Programming Systems Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91052 Erlangen, Germany

Abstract

Reliable radio-based indoor localization is essential for industrial automation, yet existing methods are often evaluated in environments that do not fully reflect the severe multipath and non-line-of-sight (NLOS) conditions encountered in real deployments. This limits our understanding of how AI, i.e., modern geometric, unsupervised, and semi-supervised localization methods generalize to industrial 5G settings. We present RAIL-5G, a public real-world 5G NR channel impulse response (CIR) benchmark recorded in an operational railway maintenance facility. RAIL-5G covers an 811m $^2$ indoor area with eight distributed radio units, robot-mounted user equipment (UE) trajectories, high-accuracy iGPS reference positions, and severe metal-induced multipath caused by train compartments and maintenance infrastructure. We benchmark geometric time-difference of arrival (TDoA), supervised fingerprinting, unsupervised channel charting, and self-supervised wireless pretraining methods. Our results highlight the difficulty of RAIL-5G: on the full dataset, TDoA reaches 7.02m mean average error (MAE), the best unsupervised method reaches 2.46m, and supervised fingerprinting achieves 0.52m. Self-supervised pretraining reduces the labeling effort by up to one order of magnitude, but the remaining gap to supervised learning shows that robust AI-based localization under severe multipath remains an open challenge. The dataset and code are available at https://iis.fraunhofer.de/rail-5g.

Figures & tables

Fig. 1: RAIL-5G environment, viewed from RU 1 toward RU 4/5. Large metallic train compartments cause severe multipath propagation.

Fig. 2: Map of the recording area. The environment consists of an open area on the left and two narrow corridors between metallic train compartments on the right. The distributed ru are shown in dark blue.

Take	Description	Data Points
#1	Meander lower corridor	52,588
#2	Meander free space	51,552
#3	Meander upper corridor	54,963
#4	Serpentine pattern lower corridor	43,130
#5	Serpentine pattern free space	25,809
#6	Serpentine pattern upper corridor	31,682

TABLE I: Overview of the recorded takes.

Panel of Figure 4 — Fig. 3: Overview of full (grey) and subset (blue) data.

	Full						Lower Corridor						Open Area
Method	MAE $\downarrow$	CEP $\downarrow$	R95 $\downarrow$	KS $\downarrow$	TS $\uparrow$	CT $\uparrow$	MAE $\downarrow$	CEP $\downarrow$	R95 $\downarrow$	KS $\downarrow$	TS $\uparrow$	CT $\uparrow$	MAE $\downarrow$	CEP $\downarrow$	R95 $\downarrow$	KS $\downarrow$	TS $\uparrow$	CT $\uparrow$
TDoA	7.019	6.283	15.615	—	—	—	2.443	1.867	6.122	—	—	—	6.276	6.242	10.923	—	—	—
FP	0.520	0.389	1.312	—	—	—	0.590	0.497	1.342	—	—	—	0.739	0.607	1.808	—	—	—
PCA	3.870	3.759	7.172	0.292	0.928	0.960	2.750	2.289	5.567	0.313	0.919	0.956	2.203	2.014	4.362	0.339	0.872	0.920
Isomap*	4.455	4.143	9.095	0.405	0.962	0.956	1.896	1.709	3.958	0.212	0.951	0.949	2.016	1.867	4.006	0.321	0.912	0.909
t-SNE*	3.461	3.175	7.511	0.295	0.996	0.977	3.733	2.887	9.530	0.399	0.995	0.966	1.665	1.336	4.164	0.282	0.995	0.937

TABLE II: Localization results on 3 dataset splits (* are non-parametric methods and also fitted on the test set).

Fig. 4: Norm. CIRA metric change per real-world distance. Larger changes in the corridors indicate stronger spatial variability of the radio channel.

Panel of Figure 7 — Fig. 5: Channel charts colorized by reference position after optimal affine alignment.

Fig. 6: Localization performance after self-supervised pretraining and supervised fine-tuning with different fractions of labeled samples.

Wireless Localization

Wireless localization uses radio signals and sensor fusion to estimate the positions and trajectories of people, devices, and transmitters.

See all 65 papers

Radiance-Field Guided Pretraining: Scaling Localization Models with Unlabeled Wireless Signals
Dec 8, 2025Guosheng Wang, Shen Wang, Lei YangSelf-Supervised Pre-Training Wireless Localization
Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China
ChronoChart: Setup-Agnostic Channel Charting for Label-Free Indoor Localization
Oct 18, 2026Majid Sharghi Foroushani, Jonas Pirkl, Jonathan Ott +4Wireless Localization Temporal Representation Learning
Fraunhofer Institute for Integrated Circuits (IIS), Division Positioning and Networks, 90411 Nürnberg, Germany · Programming Systems Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91052 Erlangen, Germany · Machine Learning and Positioning Systems Lab, University of Technology Nuremberg (UTN), Nürnberg, Germany
Adaptive Learning Strategies for AoA-Based Outdoor Localization: A Comprehensive Framework
May 6, 2026Bac Trinh-Nguyen, Sara Berri, Sin G. Teo +2Wireless Localization Adaptive Learning
1ETIS UMR 8051, CY Cergy Paris University, ENSEA, CNRS, FR · Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), SG · 3Singapore Institute of Technology (SIT), SG +1

Dec 8, 2025J/K move · Enter open · S save

Radiance-Field Guided Pretraining: Scaling Localization Models with Unlabeled Wireless Signals

Guosheng Wang, Shen Wang, Lei Yang

Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China

Self-Supervised Pre-Training Wireless Localization

Open paper AskPDF arXiv

Radio frequency (RF)-based indoor localization offers significant promise for applications such as indoor navigation, augmented reality, and pervasive computing. While deep learning has greatly enhanced localization accuracy and robustness, existing localization models still face major challenges in cross-scene generalization due to their reliance on scene-specific labeled data. To address this, we introduce Radiance-Field Reinforced Pretraining (RFRP). This novel self-supervised pretraining framework couples a large localization model (LM) with a neural radio-frequency radiance field (RF-NeRF) in an asymmetrical autoencoder architecture. In this design, the LM encodes received RF spectra into latent, position-relevant representations, while the RF-NeRF decodes them to reconstruct the original spectra. This alignment between input and output enables effective representation learning using large-scale, unlabeled RF data, which can be collected continuously with minimal effort. To this end, we collected RF samples at 7,327,321 positions across 100 diverse scenes using four common wireless technologies--RFID, BLE, WiFi, and IIoT. Data from 75 scenes were used for training, and the remaining 25 for evaluation. Experimental results show that the RFRP-pretrained LM reduces localization error by over 40% compared to non-pretrained models and by 21% compared to those pretrained using supervised learning.

Abstract

Wireless Localization

Explore similar work

Radiance-Field Guided Pretraining: Scaling Localization Models with Unlabeled Wireless Signals

ChronoChart: Setup-Agnostic Channel Charting for Label-Free Indoor Localization

Adaptive Learning Strategies for AoA-Based Outdoor Localization: A Comprehensive Framework