PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Wireless Digital Twin Construction and Calibration Using Real-World Measurements

Fan-Hao Lin, Kun-Li Shih, Chu-Hsiang Huang, Hui Chen, Chao-Kai Wen, Henk Wymeersch

arXiv:2610.01437v1cs.ITmath.IT

TL;DR

High-fidelity wireless digital twins require accurate propagation modeling, while collecting sufficient real-world channel data is costly and difficult to scale. AWDT coordinates agents to construct and calibrate a WDT using public environmental information, commercial smartphone measurements, and ray-tracing feedback. It reduces RSRP prediction MAE from 11.46 dB to 5.25 dB and demonstrates cross-device transfer with lightweight device-specific bias adaptation.

  • Problem

    High-fidelity WDT construction requires accurate site-specific propagation modeling, while collecting enough real-world data is costly, time-consuming, and difficult to scale.

    Source: §I · p. 1

  • Method

    AWDT coordinates specialized agents to build the environment, infer operator configurations, and calibrate radio materials using commercial smartphone measurements and ray-tracing feedback.

    Source: §I · p. 1

  • Results

    RSRP prediction MAE falls from 11.46 dB to 5.25 dB, and the calibrated WDT transfers across measurement devices with lightweight device-specific bias adaptation.

    Source: §I · p. 1, §IV · p. 6

  • Takeaways & Limitations

    The results highlight the potential of agentic AI for scalable and practical WDT construction and calibration with limited prior information.

    Source: §IV · p. 6

Abstract

from arXiv · show

Wireless digital twins (WDTs) are promising enablers for developing and evaluating AI-native radio access networks, yet constructing a high-fidelity WDT typically requires substantial manual effort to integrate heterogeneous information and infer unknown propagation-related parameters. This paper proposes Agentic WDT (AWDT), an end-to-end agentic framework for autonomous WDT construction and calibration using readily available environmental information and measurements readily obtainable from commercial smartphones. AWDT comprises three agents: EnvAgent constructs the propagation environment, OpAgent infers BS and sector configurations, and MatAgent calibrates radio material properties. The agents iteratively refine the WDT using discrepancies between LTE/NR reference signal received power (RSRP) measurements and ray-tracing predictions. Real-world experiments show that AWDT reduces the RSRP prediction MAE from 11.46 to 5.25 dB, demonstrating a substantial improvement in ray-tracing fidelity. Evaluation with an independent measurement system further demonstrates cross-device transferability with lightweight device-specific bias adaptation, highlighting the potential of agentic AI for scalable WDT construction and calibration with limited prior knowledge.

I. INTRODUCTION

High-fidelity wireless digital twins can support AI-RAN development, but constructing them requires accurate site-specific propagation modeling amid costly real-world data collection and unknown propagation parameters. AWDT uses specialized agents and ray-tracing feedback to construct and calibrate a WDT from commercial smartphone measurements.

I. INTRODUCTION

  • Accurate WDTs can support scalable channel-data generation for AI-RAN, while collecting sufficient real-world data is costly, time-consuming, and difficult to scale.A WDT must accurately reproduce the physical environment’s propagation characteristics.Source: §I · p. 1
  • Ray-tracing fidelity depends on 3D geometry, operator configurations, antenna patterns, and radio materials, while prior agentic simulation frameworks lacked validation against real-world propagation measurements.Earlier measurement-driven calibration relied on dedicated CSI measurements and a preconstructed environment rather than end-to-end construction from commercial smartphone measurements.Source: §I · p. 1

II. MULTI-AGENT FRAMEWORK FOR AWDT

  • AWDT coordinates EnvAgent, OpAgent, and MatAgent to iteratively construct and calibrate a shared WDT using ray-tracing feedback.The agents address environment, operator-side configurations, and radio-material properties, respectively.Source: §I · p. 1, §II · p. 1

I. INTRODUCTION

  • AWDT uses publicly available environmental data and commercial smartphone measurements without requiring detailed BS configurations or dedicated channel-measurement equipment.Its agents operate on physically interpretable WDT components and collaboratively refine propagation-relevant parameters through ray-tracing feedback.Source: §I · p. 1
  • AWDT evaluates cross-device transfer using independently collected measurements and lightweight device-specific bias adaptation.The calibrated WDT is transferred across measurement devices with this adaptation.Source: §I · p. 1

A. EnvAgent: Propagation Environment Construction

EnvAgent constructs a propagation-ready scene from public environmental data and visual observations through an adaptive, goal-driven workflow. OpAgent infers base-station and sector configurations from cell identities, RSRP measurements, and ray-tracing feedback.

A. EnvAgent: Propagation Environment Construction

  • EnvAgent integrates public building footprints, heights, terrain elevation, and other data whose resolution, coverage, representation, and reliability may differ.Streetview and ground-level photographs add local structural details that may be missing or inaccurate in public datasets.Source: A. EnvAgent: Propagation Environment… · p. 2
  • EnvAgent selects scene-specific information sources and tools, refining the scene as needed instead of following a fixed processing pipeline.Its output is a propagation-ready 3D scene with reconstructed terrain and building geometry for ray tracing and AWDT calibration.Source: A. EnvAgent: Propagation Environment… · p. 2

B. OpAgent: Operator-side Configuration Inference

  • OpAgent infers BS locations, antenna heights, and sector parameters, then refines them using cell identities, RSRP distributions, and ray-tracing feedback.The sector parameters include yaw, horizontal and vertical half-power beamwidths, downtilt, and sector-specific power bias.Source: B. OpAgent: Operator-side… · p. 2
  • OpAgent groups cells by physical cell identity, estimates sector centroids weighted by RSRP, and averages co-located sector centroids into a coarse BS anchor.Nearby rooftops around the anchor become candidate BS sites.Source: B. OpAgent: Operator-side… · p. 2
  • OpAgent uses spatial RSRP residual patterns to choose which sector parameter to adjust and in which direction, retaining updates only when agreement improves.Yaw mismatch shifts the predicted high-gain region; beamwidth and downtilt affect its coverage and residuals across UE distances and elevation angles.Source: B. OpAgent: Operator-side… · p. 2

C. MatAgent: Radio Material Property Calibration

MatAgent tackles the underconstrained problem of calibrating radio materials from commercial RSRP measurements by prioritizing buildings with local mismatch and retaining only simulation-tested updates.

C. MatAgent: Radio Material Property Calibration

  • MatAgent addresses the underconstrained material-calibration problem through spatially prioritized, simulation-tested updates.Different material settings can produce similar received powers because each measurement aggregates multiple propagation contributions.Source: C. MatAgent: Radio Material Property… · p. 2
  • Each building is a calibration unit with relative permittivity ϵr, conductivity σ, and scattering coefficient S.Source: C. MatAgent: Radio Material Property… · p. 2
  • Because valid NR gNB identifiers and NCI are unavailable, target NR PCIs are assumed to be co-located.Source: C. MatAgent: Radio Material Property… · p. 2
  • The distance-weighted score q_b ranks buildings by local RSRP mismatch, with larger values indicating higher calibration priority.The weighting uses the training set, residuals, distance to the building footprint, and a spatial scale ℓ=90.Source: C. MatAgent: Radio Material Property… · p. 3
  • Buildings are processed in descending q_b order, and each material update is kept only if ray tracing improves the validation objective.Scores and residuals are recomputed after an accepted update; calibration stops at the validation-based stopping criterion.Source: C. MatAgent: Radio Material Property… · p. 3

A. Experiment Setup

The experiments use the National Sun Yat-sen University campus in Taiwan, a varied propagation setting with flat and mountainous terrain.

A. Experiment Setup

2) Commercial RSRP measurement:

Commercial smartphone measurements provide LTE and NR RSRP with GPS locations, while an SDR-based system supports cross-device evaluation. The data are spatially split for calibration and evaluation, and all WDT predictions use a fixed ray-tracing configuration.

2) Commercial RSRP measurement:

  • A commercial smartphone running Cellular-Pro collected LTE B3 RSRP and NR n78 SS-RSRP with GPS locations for AWDT construction and calibration.The measured signals are LTE B3 at 1.86 GHz and NR n78 at 3.46 GHz.Source: 2) Commercial RSRP measurement · p. 3
  • An SDR-based sniffer with fixed, explicitly configured receiver gain collected NR SS-RSRP to test transfer of propagation characteristics across devices.The motivation is that smartphone-reported RSRP can vary with RF front-ends and measurement implementations.Source: 2) Commercial RSRP measurement · p. 3

3) RSRP Data Preprocessing:

  • Across 34 runs, measurements were aggregated into 4,228 occupied 5 m×5 m cells and split by road segment into training, validation, and test sets.The split contains 3,382 (80%), 423 (10%), and 423 (10%) cells; cell medians reduce smartphone-orientation and short-term variation.Source: 3) RSRP Data Preprocessing · p. 3

4) Agent Platform:

  • EnvAgent, OpAgent, and MatAgent use OpenAI Codex as the agent environment, with role-specific instructions, tools, and workflows.The environment supports context-aware reasoning, code execution, and external tool interaction.Source: 4) Agent Platform · p. 3
  • All WDT predictions use NVIDIA Sionna RT with reflection, diffraction, scattering, and refraction enabled and maximum interaction depth 10.The same configuration is used throughout construction and calibration.Source: 4) Agent Platform · p. 3

B. Propagation Environment Construction

EnvAgent combines public geospatial data and reference photographs to refine a propagation-ready scene. In the reported comparison, that refinement improves agreement between ray-tracing predictions and measured RSRP.

B. Propagation Environment Construction

  • EnvAgent supplements OSM building footprints with building heights from the Taiwan 3D Map Service and terrain elevation from NASA SRTM.The OSM initialization has incomplete height information and insufficient terrain representation, creating geometric mismatches.Source: B. Propagation Environment Construction · p. 3

5) Ray-Tracing Configuration:

  • Reference photographs let EnvAgent identify and correct missing or inaccurate structures, including corridors and building-specific geometries.The refinement is intended to better reproduce the actual propagation environment.Source: 5) Ray-Tracing Configuration · p. 3
  • After EnvAgent refinement, LTE and NR predictions become more consistent with measurements than the OSM-based scene, whose errors mainly reflect underestimated building heights and missing terrain blockage.Source: 5) Ray-Tracing Configuration · p. 3

C. Commercial RSRP Calibration

Commercial RSRP residuals guide calibration of radio materials and operator-side settings after environment construction, improving agreement between ray-tracing predictions and measurements.

C. Commercial RSRP Calibration

  • MatAgent refines building-wise permittivity, conductivity, and scattering, with calibrated buildings showing distinct dielectric-property groups but relatively narrow scattering variation.The text attributes the limited scattering variation to diffuse scattering being less distinguishable from aggregate received-power measurements.Source: C. Commercial RSRP Calibration · pp. 4–5
  • Calibration retains validation-set updates only when L = EP68 + 0.2EP95 decreases, stopping after five consecutive unsuccessful candidate updates.The weighted objective prioritizes typical error while retaining sensitivity to large tail errors.Source: C. Commercial RSRP Calibration · p. 4
  • OpAgent infers BS locations and sector configurations from measurements; for PCI 276, a greater antenna height restores the observed LoS geometry and reduces RSRP underestimation.Without operator-side ground truth, the inferred height is a physically consistent configuration, not an exact reconstruction.Source: C. Commercial RSRP Calibration · p. 4
  • After environment refinement, residual discrepancies remain from unknown operator configurations and material properties; OpAgent and MatAgent improve agreement by calibrating both.The section evaluates inferred configurations through resulting RSRP predictions because exact sector parameters cannot be directly verified.Source: C. Commercial RSRP Calibration · pp. 4–5
  • MAE falls from 11.46 dB for the OSM-based scene to 5.25 dB after calibration, outperforming the U-Net and linear-interpolation baselines.The comparison covers LTE B3 and NR n78 measurements across construction and calibration stages.Source: C. Commercial RSRP Calibration · p. 5, Figure 4 · p. 5

D. Cross-Measurement-Device Evaluation

Because smartphone receivers can have device-dependent RSRP variations, the study evaluates the calibrated AWDT using independent SDR measurements after adapting only a device-specific bias.

D. Cross-Measurement-Device Evaluation

  • Device-dependent RSRP variation can arise from differences in RF front-ends and receiver implementations, while internal gain states and compensation procedures are generally inaccessible.This motivates a separate measurement system for the transfer evaluation.Source: D. Cross-Measurement-Device Evaluation · pp. 5–6
  • An SDR-based sniffer with fixed, explicitly configured receiver gain supplies an independent NR SS-RSRP reference for testing whether smartphone-calibration gains transfer across devices.Its SS-RSRP estimate uses resource elements carrying the secondary synchronization signal.Source: D. Cross-Measurement-Device Evaluation · p. 6
  • With only a global device-specific bias correction, calibrated AWDT achieves consistently lower prediction errors on the remaining SDR measurements.The propagation-related parameters are kept fixed; 20% of SDR measurements estimate the bias, and the remainder evaluate performance.Source: D. Cross-Measurement-Device Evaluation · p. 6

IV. CONCLUSION

AWDT combines three agents to construct and calibrate a WDT from commercial measurements with limited prior information, reducing prediction error and showing cross-device transferability.

IV. CONCLUSION

  • AWDT jointly uses EnvAgent, OpAgent, and MatAgent to construct the environment, infer operator-side configurations, and calibrate radio materials using commercial RSRP measurements.The framework operates with limited prior information.Source: §IV · p. 6
  • MAE decreases from 11.46 dB to 5.25 dB after AWDT calibration.Source: §IV · p. 6
  • AWDT demonstrates cross-device transferability with lightweight device-specific bias adaptation.The conclusion describes this as evidence of potential for scalable and practical WDT construction and calibration.Source: §IV · p. 6, Table I · p. 6
Loading 2610.01437v1…