PaperlayerView exact revision on arXiv

Source-linked AI summary

A Ray-Traced Digital Twin for 5G NR Positioning with OpenAirInterface

Nima Chaharbaghi, Florian Kaltenberger, Bartosz Podrygajlo, Rakesh Mundlamuri, Roberto Verdone

arXiv:2610.13138v1cs.ITmath.IT

TL;DR

Physical 5G positioning measurements are tied to a single propagation environment and antenna layout, leaving alternative deployment questions to new field campaigns. This paper builds a ray-traced twin that drives unmodified OAI UL-TDoA while preserving per-antenna absolute delays. It reports metre-level horizontal accuracy and finds that bandwidth does not resolve geometry-limited vertical error, whereas a height-diverse layout reduces RMS height error by a factor of four.

  • Problem

    Physical-testbed positioning measurements are tied to one environment and antenna layout, and existing ray-traced OAI twins target single-antenna links rather than positioning.

    Source: §I · p. 1

  • Method

    A ray-traced digital twin models distributed gNB antennas as independent radio devices to drive the unmodified OAI UL-TDoA pipeline with per-antenna absolute delays.

    Source: C. Distributed-antenna channel model · p. 3, §VIII · p. 8

  • Results

    Median horizontal error is 0.71–1.20 m; wider bandwidth truncates the horizontal-error tail without improving typical positions, while vertical accuracy remains geometry-limited.

    Source: C. Effect of channel bandwidth · p. 6, §VIII · p. 8

  • Takeaways & Limitations

    A height-diverse layout cuts RMS height error by a factor of four, and the height limit is predictable from antenna coordinates alone.

    Source: §VIII · p. 8

    The deterministic emulator evaluates only sixteen non-random positions, so repeated sweeps do not sample a random process.

    Source: §VII · p. 7

Abstract

from arXiv · show

Evaluating 5G positioning on a physical testbed ties every measurement to one propagation environment and one antenna layout, so any question about an alternative deployment requires a new field campaign. We present a ray-traced digital twin of the GEO-5G testbed at EURECOM that drives an unmodified OpenAirInterface (OAI) 5G stack with per-antenna channels computed by Sionna-RT over a 3D model of the site. The distributed gNB antennas are modelled as independent radio devices rather than as elements of a compact array, which preserves the per-antenna absolute delays on which time-difference-of-arrival (TDoA) positioning depends. With them preserved, the twin drives OAI's SRS-based UL-TDoA pipeline to metre-level accuracy consistent with that reported for the physical deployment. We then use the twin to answer a question the physical testbed cannot. Widening the carrier truncates the tail of the horizontal error distribution without improving a typical position, and leaves the vertical error untouched, because the latter is limited by the array geometry rather than by the measurement. That limit is predictable from the antenna coordinates alone, and a height-diverse layout reduces the RMS height error by a factor of four.

I. INTRODUCTION

Physical 5G positioning campaigns bind measurements to one environment and antenna layout, while existing ray-traced OAI twins do not exercise positioning. This paper combines distributed-antenna ray tracing with the unmodified OAI UL-TDoA pipeline, then uses the twin to examine deployment choices unavailable to the physical testbed.

I. INTRODUCTION

  • Physical-testbed measurements bind results to one environment and antenna layout, making alternative deployment questions costly to answer through new field campaigns.Such campaigns require licensed spectrum, certified hardware, and a prepared site.Source: §I · p. 1
  • 0.71–1.20 m median horizontal error is achieved across the reported evaluation, while a height-diverse layout cuts RMS height error by a factor of four.The paper also reports that vertical accuracy is insensitive to a 2.5× bandwidth increase because it is limited by array geometry.Source: §I · p. 1
  • The twin extends ray-traced OAI emulation to eight spatially separated antennas, preserving each antenna’s absolute propagation delay for the unmodified UL-TDoA pipeline.It models each antenna as an independent radio device rather than as an element of a co-located array.Source: §I · p. 1

II. RELATED WORK

  • Prior ray-traced OAI twins target single-antenna links, which provide no arrival-time differences for a TDoA pipeline.The related work also describes hardware evaluations of OAI’s UL-TDoA pipeline, but not the combination of distributed-antenna ray tracing and positioning.Source: §II · pp. 1–2

III. DIGITAL TWIN ARCHITECTURE

The twin uses the GEO-5G deployment’s surveyed, distributed antenna geometry and two ray-tracing scenes. The campus scene is reconstructed and configured for propagation modelling, with a fixed solver setup chosen to accommodate multiply reflected paths.

A. Reference deployment

  • The eight GEO-5G elements form two orthogonal four-element sub-arrays, separated by 42.19 m between horizontal centroids and surrounding a courtyard.The elements are spaced roughly 60λ to 217λ apart.Source: A. Reference deployment · p. 2

B. Scene and propagation

  • The EURECOM scene is reconstructed from OpenStreetMap footprints, refined manually, assigned frequency-dependent radio materials, and exported for Sionna RT.The Munich scene uses the same array geometry under a distance-preserving transform.Source: B. Scene and propagation · p. 2, Figure 2 · p. 2
  • Solver settings are fixed, with depth 10 chosen for courtyard links that reach some antenna positions only through multiple reflections.No depth sensitivity study was run, so depth 10 is a conservative choice rather than a measured requirement; diffuse reflection is disabled and the SBR seed is fixed for bit-identical repeated taps.Source: B. Scene and propagation · p. 2

C. Distributed-antenna channel model

The channel model treats each spatially separated antenna as an independent radio device so the OAI positioning pipeline receives per-antenna absolute delays. The integration sends those channels through vrtsim while leaving the OAI stack unmodified.

C. Distributed-antenna channel model

  • Modelling each widely separated element as an independent radio device preserves its own visibility, reflectors, and absolute delay instead of collapsing the array to a common phase centre.The separations span 60–217λ at ranges of 24–57 m, where the compact-array plane-wave assumption is unsuitable.Source: C. Distributed-antenna channel model · p. 3
  • Eight independent 1×1 Sionna radio devices represent the gNB elements, with mirrored scenes used for uplink and downlink.Sionna’s array configuration is a scene property, motivating the mirrored scenes.Source: C. Distributed-antenna channel model · p. 3
  • Band-limited sampling converts paths to discrete taps over 71 taps per antenna pair, with absolute delays preserved on the same grid OAI uses.The farthest antenna can trail the nearest by up to 109 ns; the tap window is sized to accommodate absolute delay.Source: C. Distributed-antenna channel model · p. 3

D. Integration with OAI

  • Per-radio-device messages carry antenna-indexed complex taps through vrtsim, while the gNB and nrUE softmodems run unmodified above it.Scripted UE position updates trigger new solves and provide ground truth by construction.Source: D. Integration with OAI · p. 3, Figure 3 · p. 3

IV. UL-TDOA POSITIONING AND ITS CONDITIONING

UL-TDoA differences arrival times at known receivers to remove the unknown transmit instant, then estimates the UE position by least squares. The resulting position quality depends on antenna geometry, which the dilution figures summarize.

A. Measurement chain

  • The UE’s unknown transmit instant cancels when each arrival time is differenced against a reference TRP, yielding range differences.Each difference equals the difference between the UE-to-TRP ranges.Source: A. Measurement chain · p. 3
  • Eight TRPs provide seven range-difference constraints for a three-unknown least-squares position solve; the gNB reports SRS arrival times to the LMF over NRPPa.The LMF uses TRP coordinates and returns the position through the location-services interface.Source: A. Measurement chain · p. 3

B. Geometric conditioning

  • The position covariance is σ^2Q, with Q = (A^TA)^-1, under uncorrelated range-difference errors of common variance σ^2.A depends on the TRP directions relative to the UE, so the geometry determines the conditioning.Source: B. Geometric conditioning · p. 4
  • HDOP summarizes horizontal variances and VDOP the vertical variance from the same 3D solve; weak height geometry can also inflate HDOP.The dilution figures are dimensionless and compare conditioning, not exact error-propagation coefficients.Source: B. Geometric conditioning · p. 4

A. Two scenes

The study uses two propagation scenes while preserving the array geometry, then compares two complete bandwidth configurations. The position sweeps span differing geometric conditioning, and the configuration changes are not limited to occupied bandwidth.

B. Two bandwidths

  • Both configurations report correlation peaks on a 0.61 m bin spacing, but oversampling does not equalize their physical delay resolution, which follows occupied bandwidth.The tap and receiver delay-estimate quantizations are distinct; the receiver bin spacing is adjusted to 2.03 ns (0.61 m).Source: Table I · p. 4, B. Two bandwidths · pp. 4–5

A. Two scenes

  • The EURECOM and Munich sweeps use the same array geometry, but independently chosen test positions, so scene comparisons are aggregate rather than controlled substitutions.The array transform preserves element spacing, axis relations, and the 42.19 m centroid separation.Source: A. Two scenes · p. 4
  • Sixteen positions per scene span HDOP 2.24–7.79 at EURECOM and 2.06–7.36 in Munich, predicting about 3.5× greater error at each track’s far end than near end.The test tracks are at z = 1.7 m and run away from the array.Source: A. Two scenes · p. 4

B. Two bandwidths

  • Table I compares two complete operating points, not an isolated bandwidth change: PRB count also changes sample rate, tap spacing, and the physical duration of the 71-tap window.The absolute-delay window was not held constant, so its shorter duration in configuration B may contribute residually.Source: B. Two bandwidths · p. 4, Table I · p. 4
  • Changing the uplink power-control target by 6 dB reproduced all sixteen configuration-A fixes exactly when rerun at configuration B’s value.Four other configuration-file differences govern random access only.Source: B. Two bandwidths · p. 5

C. Determinism and what the statistics mean

Across the two scenes and bandwidths, most horizontal fixes are within the target range, but a poorly conditioned position produces a large error. The emulator is deterministic, so the reported track quantiles describe variation with geometry rather than repeated-measurement variability.

C. Determinism and what the statistics mean

  • Repeated emulator runs return identical fixes, so the track quantiles measure how deterministic error varies across positions, not variability across repeated trials.At a given position, averaging cannot reduce the same returned error.Source: C. Determinism and what the statistics… · p. 5

A. Horizontal accuracy

  • The 19.07 m Munich error occurs at the track’s worst-conditioned position, where a likely neighboring-tap assignment is amplified by HDOP 6.96.The implied range-difference error is 2.74 m, roughly half the 4.88 m tap spacing.Source: A. Horizontal accuracy · p. 5
  • 1.20 to 0.95 m at EURECOM and 0.98 to 0.71 m in Munich are the median horizontal errors as bandwidth widens; eleven to thirteen of sixteen fixes meet the 1.5 m target.The Munich outlier is 19.07 m at 40 MHz and 2.61 m at 100 MHz; excluding it, no fix exceeds 7 m.Source: A. Horizontal accuracy · p. 5, Figure 4 · p. 5

B. The error follows the geometry

The measured range-difference error is close to the UL-RTOA reporting granularity, while geometry explains substantial variation at poorly conditioned positions. Wider bandwidth mainly truncates large horizontal errors; it has little, inconsistent effect on height error.

B. The error follows the geometry

  • 0.31 m, or 1.02 ns, is the common estimated range-difference error in three of four sweeps, within 1% and about one UL-RTOA reporting step.The reporting granularity is Tc/2 = 1.017 ns, placing the pipeline close to that format’s floor.Source: B. The error follows the geometry · p. 5
  • Munich’s e2D–HDOP correlation is 0.71 and 0.66 across the two bandwidths, and its worse-conditioned half has 5.8× the mean error at 40 MHz.EURECOM’s weaker correlations occur where errors span only 0.24–3.81 m and positions are mostly near the measurement floor.Source: B. The error follows the geometry · p. 5

C. Effect of channel bandwidth

  • The Munich mean error in the worse-conditioned half falls from 4.47 to 1.44 m, while the better-conditioned half changes from 0.77 to 0.63 m as bandwidth widens.Five of eight positions improve in the worse-conditioned half, compared with two of eight in the better-conditioned half.Source: C. Effect of channel bandwidth · p. 6
  • The Munich 90th-percentile horizontal error falls from 5.39 to 2.32 m, while the median paired change is zero in both scenes.The largest error falls from 3.81 to 2.73 m at EURECOM and 19.07 to 3.70 m in Munich; bandwidth has no useful effect on vertical error, whose RMS changes by under 6% in opposite directions across scenes.Source: C. Effect of channel bandwidth · p. 6

D. Why the array cannot see height

The two-height layout gives the solver little independent height information, leaving vertical accuracy geometry-limited. Alternating antenna heights improves the predicted geometry and measured height accuracy, while horizontal gains are offset by increased range-difference error.

D. Why the array cannot see height

  • In the exact LMF system, three equations carry no height information and four repeat one coefficient, leaving height confounded with r0.The four elevated antennas average down the variance of one measurement combination but do not provide a second direction for resolving height.Source: D. Why the array cannot see height · p. 6
  • Median HDOP is 3.53 at EURECOM and 3.28 in Munich; making height observable halves HDOP even though horizontal coordinates are unchanged.The paper attributes that horizontal penalty to information being spent on the poorly observed height coordinate.Source: D. Why the array cannot see height · p. 6, E. Answering a question the testbed… · p. 6
  • At sixteen EURECOM positions and 40 MHz, normalized height residuals have mean −0.07 and standard deviation 1.02.The model predicts zero mean and unit standard deviation if height error is the same measurement error amplified by vertical geometry.Source: D. Why the array cannot see height · p. 6

E. Answering a question the testbed cannot

  • Alternating antenna heights creates six non-zero height-column coefficients with three distinct values, instead of one value repeated four times.The proposed layout inclines every baseline by more than 50°; spreading heights gradually within each sub-array would leave baselines nearly horizontal.Source: E. Answering a question the testbed… · p. 6
  • VDOP falls by a factor of five and HDOP by a factor of two after the height change, with horizontal antenna coordinates unchanged.Table III reports median dilution over sixteen test positions; the predicted height error is ˆσ VDOP at ˆσ = 0.31 m.Source: E. Answering a question the testbed… · p. 6, Table III · p. 7
  • At 40 MHz, RMS height error falls from 3.50 to 0.85 m at EURECOM and 2.31 to 0.95 m in Munich; worst error falls from 7.44 to 2.05 m.Fig. 5 shows the two-height array's height-error distribution spreading to 7 m, while the height-diverse array stays inside 2 m.Source: Table IV · p. 7, E. Answering a question the testbed… · p. 7, Figure 5 · p. 7
  • At EURECOM 100 MHz, the 90th-percentile horizontal error rises from 1.80 to 4.47 m despite improved median error at all four operating points.The predicted twofold horizontal geometry gain was not realised as implied range-difference error rose from 0.307 to 0.575 m at EURECOM and 0.308 to 0.535 m in Munich; the proposed rooftop-path explanation is a hypothesis.Source: E. Answering a question the testbed… · p. 7

F. Comparison with the physical testbed

The twin's closest comparison with the physical testbed is the EURECOM 100 MHz 90th-percentile result; other settings and methodological differences limit how broadly the match can be read.

F. Comparison with the physical testbed

  • At the matched EURECOM site and 100 MHz, the twin's 90th percentile is 1.80 m, within the physical testbed's reported 1–2 m range in 90% of cases [2].The Munich 90th percentile is 2.32 m and the 40 MHz sweeps reach 5.39 m, so the agreement is specific to the matched case.Source: F. Comparison with the physical testbed · p. 7
  • The twin is perfectly synchronized and has a sparser impulse response, while the physical comparison uses dedicated ToA/TDoA filtering across repeated measurements.Each twin result is a single fix, and its emulated tap grid places a floor under the arrival estimate.Source: F. Comparison with the physical testbed · p. 7
  • The comparison is indicative rather than like-for-like; within that limit, the twin reproduces the modeled system's accuracy regime.Source: F. Comparison with the physical testbed · p. 7

VII. LIMITATIONS

The twin reports metre-scale horizontal error and a vertical limit tied to array geometry, while the evaluation’s statistics, bandwidth interpretation, and model scope remain constrained. The conclusion identifies reproducible results and further validation needs.

VII. LIMITATIONS

  • With zero noise and 16 non-random positions on two tracks, the estimated 90th percentile represents effectively two points, not genuine error statistics.The authors say genuine statistics require noise realizations or UE-position perturbation sweeps, neither of which was performed.Source: §VII · p. 7
  • The two configurations change occupied bandwidth, sample rate, tap spacing, and the physical 71-tap window together, so the improvement cannot be attributed to bandwidth alone.That attribution would require a constant absolute-delay window and per-TRP delay profiles, which were not recorded.Source: §VII · pp. 7–8
  • The model omits diffuse scattering, coupling, RF impairments, antenna patterns, and absolute path loss; static snapshots also leave mobility, Doppler, and tracking untested.The distributed-geometry contribution was assumed rather than measured, and the height-diverse layout was tested at only one assignment.Source: §VII · p. 8

VIII. CONCLUSION

  • Median horizontal error was 0.71–1.20 m; vertical accuracy was bandwidth-insensitive with the two-height array, while a height-diverse layout cut RMS height error fourfold.The vertical limit is predictable from antenna coordinates alone; the layout change required one configuration file.Source: §VIII · p. 8
  • Per-position sweep outputs and their gNB configuration are available at, allowing every reported figure and table to be recomputed from raw fixes.Source: §VIII · p. 8
  • Future work includes noise and perturbation sweeps, per-TRP arrival-time recording, mapping height versus link quality across layouts, and comparison with physical-site measurements.Source: §VIII · p. 8
Loading 2610.13138v1…