HYMN HYbrid Multi-technology Navigation

HYMN is a time-synchronised measurement dataset of UWB, BLE, WiFi FTM, 5G NR and GNSS, recorded against total-station ground truth in an industrial hall whose open gates create an indoor-outdoor transition. This page collects the dataset and the studies built on it.

Descriptor paper Dataset on Zenodo GitHub Benchmark code

  • hall grid
  • driveway, inside
  • driveway, outside
  • total station
HYMN campaign layout on an orthophoto Reference points and infrastructure of the HYMN campaign in the local frame of the dataset, drawn over the GeoSN orthophoto of the site. The hall extends to the right, the driveway crosses it near the left edge, and six reference points lie on the driveway outside the gates. A01B1 · driveway row inside the hall · x -10.83 m, y -3.07 m A01B2 · driveway row inside the hall · x -7.81 m, y -3.06 m A01B3 · driveway row inside the hall · x -4.84 m, y -3.08 m A01B4 · driveway row inside the hall · x -1.84 m, y -3.10 m A01B5 · driveway row inside the hall · x 1.18 m, y -3.08 m A01B6 · driveway row inside the hall · x 4.16 m, y -3.07 m A02B2 · hall grid · x -7.87 m, y -0.03 m A02B4 · hall grid · x -1.86 m, y -0.04 m A02B6 · hall grid · x 4.10 m, y -0.04 m A03B1 · hall grid · x -10.87 m, y 3.01 m A03B3 · hall grid · x -4.88 m, y 2.99 m A03B5 · hall grid · x 1.15 m, y 2.96 m A04B2 · hall grid · x -7.89 m, y 5.98 m A04B4 · hall grid · x -1.88 m, y 6.00 m A04B6 · hall grid · x 4.16 m, y 5.80 m A05B1 · hall grid · x -10.88 m, y 9.06 m A05B3 · hall grid · x -4.87 m, y 8.97 m A05B5 · hall grid · x 1.15 m, y 8.95 m A06B2 · hall grid · x -7.92 m, y 11.93 m A06B4 · hall grid · x -1.92 m, y 11.94 m A06B6 · hall grid · x 4.08 m, y 11.84 m A07B1 · hall grid · x -10.90 m, y 14.81 m A07B3 · hall grid · x -4.89 m, y 14.91 m A07B5 · hall grid · x 1.12 m, y 14.94 m A08B1 · hall grid · x -10.89 m, y 17.91 m A08B3 · hall grid · x -4.88 m, y 17.95 m A08B5 · hall grid · x 1.11 m, y 17.98 m A09B2 · hall grid · x -7.84 m, y 20.92 m A09B4 · hall grid · x -1.84 m, y 20.91 m A09B6 · hall grid · x 3.83 m, y 20.95 m A10B1 · hall grid · x -10.79 m, y 23.89 m A10B3 · hall grid · x -4.86 m, y 23.89 m A10B5 · hall grid · x 1.14 m, y 23.93 m A11B2 · hall grid · x -7.92 m, y 26.87 m A11B4 · hall grid · x -1.91 m, y 26.91 m A11B6 · hall grid · x 4.09 m, y 27.00 m A12B1 · hall grid · x -10.83 m, y 29.88 m A12B3 · hall grid · x -4.86 m, y 29.89 m A12B5 · hall grid · x 1.14 m, y 29.93 m A13B2 · hall grid · x -7.84 m, y 32.87 m A13B4 · hall grid · x -1.87 m, y 32.92 m A13B6 · hall grid · x 4.10 m, y 32.90 m T01 · outside the gates · x -20.14 m, y -3.38 m T02 · outside the gates · x -17.47 m, y -3.03 m T03 · outside the gates · x -14.81 m, y -3.05 m T04 · outside the gates · x 6.81 m, y -2.96 m T05 · outside the gates · x 9.37 m, y -2.78 m T06 · outside the gates · x 12.06 m, y -2.53 m UWB_01 · UWB anchor · height 3.43 m BLE_01 · BLE beacon · height 3.21 m WIFI_01 · WiFi access point · height 3.60 m UWB_02 · UWB anchor · height 3.13 m BLE_02 · BLE beacon · height 2.93 m WIFI_02 · WiFi access point · height 3.51 m UWB_03 · UWB anchor · height 4.11 m BLE_03 · BLE beacon · height 3.98 m WIFI_03 · WiFi access point · height 4.32 m UWB_04 · UWB anchor · height 3.29 m BLE_04 · BLE beacon · height 3.01 m WIFI_04 · WiFi access point · height 3.56 m UWB_05 · UWB anchor · height 3.34 m BLE_05 · BLE beacon · height 3.02 m WIFI_05 · WiFi access point · height 3.60 m UWB_06 · UWB anchor · height 3.29 m WIFI_06 · WiFi access point · height 3.62 m UWB_07 · UWB anchor · height 3.45 m UWB_08 · UWB anchor · height 3.65 m UWB_09 · UWB anchor · height 4.18 m UWB_10 · UWB anchor · height 3.53 m NR5G_11 · 5G NR radio unit · height 8.81 m NR5G_12 · 5G NR radio unit · height 8.89 m NR5G_13 · 5G NR radio unit · height 8.90 m Total station (Leica TS16) · origin of the local frame T01–T03 T04–T06 N 10 m
Reference points and infrastructure in the local frame of the dataset. Devices mounted at the same position are drawn side by side, and pointing at a marker shows its identifier. Coordinates: reference files. Background: orthophoto, Quelle: GeoSN, dl-de/by-2-0, resampled to the local frame.
Dusk outside the industrial hall. A large gate stands open and the lit hall interior with tripods and cables is visible through it. A white van is parked on the driveway in front of the building.
Open gate on the driveway, seen from outside the hall on the evening before the recording.
Wide-angle view of the empty hall interior under a steel roof. Painted point markings cover the raised concrete floor, a surveying tripod stands in the middle, and the lowered driveway with a yellow edge runs along the right side.
Hall interior with the raised floor and the lowered driveway on the right, photographed in March 2025, after the campaign.
Interior of the industrial hall seen along the lowered driveway. Radio infrastructure on tripods stands along the left wall and on the raised hall floor to the right, where a yellow surveying tripod carries the total station. A van stands on the driveway in front of the far gate.
Measurement setup along the driveway, with the total station on a yellow tripod. PLANS 2025, Fig. 3.
5 systems
UWB, BLE, WiFi FTM, 5G NR, GNSS
48 points
static, 3 min each
44 m × 18 m
industrial hall, Torgau, Germany
mm-level
ground truth, Leica TS16
23 Oct 2024
recording date
CC BY 4.0
data licence, code under MIT

Dataset

The measurements were recorded on 23 October 2024 in an industrial hall in Torgau, Germany. A driveway crosses the hall, and the gates on each side stayed open throughout the campaign, so the points on the driveway see mixed indoor and outdoor propagation. A measurement plate carrying all receivers was placed on each reference point and held static for three minutes while all systems recorded in parallel. A Leica TS16 total station surveyed every reference point and every infrastructure device at millimetre level.

Reference points. The 48 points form three groups: 36 grid points inside the hall (A02–A13), six points on the driveway inside the hall (row A01), and six points on the driveway outside the gates (T01–T06). The studies below label these groups differently, and each summary names its own zone scheme.

The measurement plate seen from above: a BLE tag with two antennas, a smartphone and a UWB tag in holders, a surveying prism in the centre, a round GNSS antenna and a round 5G NR antenna, all mounted on a black perforated plate.
Measurement plate with the mobile units from the table below: BLE tag, UWB tags, smartphone, GNSS antenna, 5G NR antenna and the prism for the total station. Each sensor has its own ground-truth position.
Infrastructure and mobile units per system in the HYMN campaign.
SystemInfrastructureMobile unit on the platePrinciple
UWB10 anchorsZigPos RTLS2 tagsZigPos RTLSTwo-way ranging (TWR)
BLE5 beaconsMetirionic DMK-215TagMetirionic DMK-215Phase-based ranging (PBR)
WiFi6 access pointsHPE Aruba 630 SeriesSmartphoneGoogle Pixel 6aFine Timing Measurement (FTM)
5G NR3 radio unitsIndustrial small cell systemUser equipmentwith 5G NR antennaFingerprinting (SNR)
GNSSSatellitesGPS, GLONASS, Galileo, BeiDouReceiverNovAtel PwrPak 7Pseudorange
Ground truthTotal stationLeica TS16PrismOptical survey

Contents

Processed data
One table per system in CSV, Parquet and pickle format, and a merged index that links the observations of all systems by timestamp and reference point.
Raw logs
The native output of each system, including GNSS observations in RINEX 3 with precise orbit and clock products.
Reference data
Coordinates of the infrastructure and the reference points in a local frame, in UTM33N (EPSG:25833) and in ECEF (EPSG:4978), and the start and end time of every static window.

Scope and limits

  • All recordings are static. The dataset contains no trajectories.
  • GNSS was logged while the plate occupied the 12 driveway points. The 36 hall-grid points have no GNSS observations.
  • The 5G NR data are SNR values and position outputs of the small cell system, not range measurements.
  • Each system logs at its own rate, and the data are not resampled to a common clock.

Quick start

Python 3.11 or 3.12. The example iterates over the merged index and computes WiFi ranging errors against ground truth.

git clone https://github.com/TUD-ITVS/HYMN-dataset
cd HYMN-dataset
python -m pip install -r examples/requirements.txt
python examples/example_iterator.py

Research built on HYMN

  1. Measurement campaign in Torgau
  2. PLANS 2025, Salt Lake City
  3. Dataset v1.0.0 on Zenodo
  4. Descriptor preprint, then IEEE Data Descriptions
  5. ICRA 2026 workshop, Vienna
  6. IPIN 2026, Rome

IPIN 2026Presented on 7 October 2026, Rome

From Least Squares to Deep Learning: Benchmarking Indoor Positioning on the HYMN Multi-Technology Dataset

Paul Schwarzbach, Muhammad Ammad

Question. How do range-based methods from least squares to a learned regressor compare on fused UWB, BLE and WiFi ranges, and does the regressor generalise to reference points it has not seen?

Fused UWB, BLE and WiFi FTM ranges at all 48 points. The paper labels the hall grid as indoor and the 12 driveway points as transition.

  • The Bayesian grid filter (BGF) reaches the lowest median error of 0.44 m, but a small fraction of epochs collapse near the grid bounds and stretch its P95 to 6.29 m.
  • A ResNet trained on random epoch-level splits sits close on the median (0.55 m) and has the shortest tail (P95 1.31 m).
  • Holding reference points out of training raises the ResNet's median to 3.22 m. Its random-split advantage reflects interpolation rather than generalisation to unseen locations.
Empirical CDF of horizontal positioning error for ILS, RLS, BGF, ResNet-SpatialHoldout and ResNet-RandomSplit on the fused input, from 0 to 10 metres. ResNet-RandomSplit and BGF rise steepest, ILS slowest.
ECDF of horizontal positioning error on the fused input. Dotted lines mark the median and P95. The abscissa is clipped at 10 m. IPIN 2026 paper, Fig. 3.
Horizontal positioning error on the fused input, as reported in the paper.
MethodMedian [m]P95 [m]
Iterative least squares (ILS)3.6218.66
Robust least squares, Huber (RLS)1.075.68
Bayesian grid filter (BGF)0.446.29
ResNet, random split0.551.31
ResNet, spatial hold-out3.227.57

ICRA 2026 Workshop1st Workshop on Robot Meets GNSS and Ranging for Seamless Autonomy · 5 June 2026, Vienna

Bridging the Indoor-Outdoor Gap: Cross-Technology Ranging for Seamless Robot Navigation

Paul Schwarzbach

Question. How do satellite and terrestrial ranging behave outdoors, in the transition zone and indoors, judged by measurement availability and by ranging residuals against ground truth?

Raw ranges of GNSS, UWB, WiFi FTM and BLE. The paper labels the six points outside the gates as outdoor, row A01 as transition and the hall grid as indoor.

  • UWB ranges stay tight inside the hall, with a P95 absolute residual of 0.99 m indoors against 3.70 m outdoors.
  • WiFi FTM and BLE ranges carry positive biases of several metres that change with the zone. For WiFi the bias grows from 1.17 m in the transition zone to 5.94 m outdoors.
  • Satellite visibility drops sharply along the driveway row inside the hall, so satellite and terrestrial ranging degrade together at the building boundary.
Four panels of empirical CDFs of absolute ranging residuals for GNSS, UWB, WiFi and BLE, with solid, dashed and dotted lines for the outdoor, transition and indoor zones. UWB residuals stay within a few metres, BLE outdoor residuals reach tens of metres.
Empirical CDFs of absolute ranging residuals per technology and zone. Linestyles encode zones and dots mark medians. Panels use independent abscissae. ICRA 2026 workshop paper, Fig. 5.

IEEE/ION PLANS 202528 April – 1 May 2025, Salt Lake City

An empirical assessment of indoor-outdoor localization based on signals of opportunity from multiple systems

Albrecht Michler, Paul Schwarzbach, Jonas Ninnemann, Muhammad Ammad, Oliver Michler

Question. Can an opportunistic fusion of GNSS and terrestrial ranging keep a position solution on a path that leads from outside the hall through the building and out again?

The 12 driveway points, visited in order from T01 through row A01 to T06. The paper classifies them as outdoor, transition and indoor.

  • A grid filter fuses GNSS, UWB, WiFi FTM and BLE measurements with separate measurement models for indoor and outdoor positions.
  • Positioning outside the hall is less accurate than in the indoor and transition zones. The paper traces this to the deliberately basic GNSS processing, which used only upper L-band code measurements without NLOS mitigation.
  • Indoor and transition accuracy follows from the accuracy and availability of the UWB ranges.
Left: empirical CDF of positioning error in logit scale for indoor, outdoor, transition and overall. Right: box plots of positioning error per zone. The outdoor box spans up to about 5 metres, the indoor and transition boxes stay near 1 metre.
Position errors per evaluation zone as an empirical CDF in logit scale (a) and as box plots (b). PLANS 2025 paper, Fig. 14, © 2025 IEEE.

Code & data

Archived artifacts with persistent identifiers.
ArtifactArchiveSourceLicence
HYMN datasetmeasurements, preprocessing, examples 10.5281/zenodo.17979434 TUD-ITVS/HYMN-dataset data CC BY 4.0, code MIT
IPIN 2026 benchmarkILS, RLS, BGF, ResNet pipeline 10.5281/zenodo.20058106 TUD-ITVS/hymn-localization-ipin2026 MIT

The dataset DOI resolves to the latest version on Zenodo.

Cite

If you use the dataset, please cite the descriptor.

@article{Ammad2026_HYMN,
  author  = {Ammad, Muhammad and Michler, Albrecht and Schwarzbach, Paul and
             Ninnemann, Jonas and U{\ss}ler, Hagen and Michler, Oliver},
  title   = {Descriptor: A Hybrid Indoor and Indoor--Outdoor Positioning
             Multitechnology Dataset ({HYMN})},
  journal = {IEEE Data Descriptions},
  volume  = {3},
  pages   = {379--387},
  year    = {2026},
  doi     = {10.1109/IEEEDATA.2026.3691044}
}
IPIN 2026 benchmark paper
@inproceedings{Schwarzbach2026_HYMN_benchmark,
  author        = {Schwarzbach, Paul and Ammad, Muhammad},
  title         = {From Least Squares to Deep Learning: Benchmarking Indoor
                   Positioning on the {HYMN} Multi-Technology Dataset},
  booktitle     = {16th International Conference on Indoor Positioning and
                   Indoor Navigation (IPIN)},
  address       = {Rome, Italy},
  year          = {2026},
  note          = {to appear},
  eprint        = {2609.25835},
  archivePrefix = {arXiv},
  doi           = {10.48550/arXiv.2609.25835}
}
IPIN 2026 benchmark code
@misc{Schwarzbach2026_HYMN_code,
  author       = {Schwarzbach, Paul and Ammad, Muhammad},
  title        = {{HYMN} Multi-Technology Indoor Positioning Evaluation},
  version      = {v1.1-ipin2026},
  howpublished = {Zenodo},
  year         = {2026},
  doi          = {10.5281/zenodo.20058106}
}
ICRA 2026 workshop paper
@misc{Schwarzbach2026_Bridging,
  author        = {Schwarzbach, Paul},
  title         = {Bridging the Indoor-Outdoor Gap: Cross-Technology Ranging
                   for Seamless Robot Navigation},
  howpublished  = {1st Workshop on Robot Meets GNSS and Ranging for Seamless
                   Autonomy, IEEE ICRA 2026, Vienna},
  year          = {2026},
  eprint        = {2604.25541},
  archivePrefix = {arXiv},
  doi           = {10.48550/arXiv.2604.25541}
}
PLANS 2025 paper
@inproceedings{Michler2025_PLANS,
  author    = {Michler, Albrecht and Schwarzbach, Paul and Ninnemann, Jonas and
               Ammad, Muhammad and Michler, Oliver},
  title     = {An empirical assessment of indoor-outdoor localization based on
               signals of opportunity from multiple systems},
  booktitle = {2025 IEEE/ION Position, Location and Navigation Symposium (PLANS)},
  pages     = {948--959},
  year      = {2025},
  doi       = {10.1109/PLANS61210.2025.11028349}
}

People & acknowledgement

Dataset authors

  • Muhammad Ammad iD
  • Albrecht Michler iD
  • Paul Schwarzbach iD
  • Jonas Ninnemann iD
  • Hagen Ußler iD
  • Oliver Michler iD

The work was carried out at the Chair of Transport Systems Information Technology, TUD Dresden University of Technology.

Acknowledgement

The measurement campaign, the dataset and the PLANS 2025 study were funded in part by the German Federal Ministry for Digital and Transport (BMDV) within the project IDEA (FKZ 19OI22020C). MRK Management Consultants GmbH provided access to the facility.