TRACE: Learning to Self-Calibrate Wireless Digital Twins from ISAC Measurements

Saad Masrur; Saeed R. Khosravirad; Ismail Guvenc

cs.AISep 26, 2026

TRACE: Learning to Self-Calibrate Wireless Digital Twins from ISAC Measurements

Authors: Saad Masrur, Saeed R. Khosravirad, Ismail Guvenc

Organizations: Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC · Radio Systems Research, Nokia Bell Labs, Murray Hill, NJ, USA

Abstract

Wireless digital twins (DTs) rely on 3D environment models to predict radio propagation and support wireless-network decisions, yet these models are often initialized from imperfect 3D maps. Errors in building position, height, footprint, and orientation can therefore cause a high-fidelity propagation engine to simulate the wrong physical environment. In this paper, we study how a deployed wireless network can repair an existing DT using its own radio frequency (RF) measurements. In particular, we introduce Twin Residual Alignment and Calibration Engine (TRACE), a physics-grounded learning-based self-calibration framework that treats twin maintenance as residual alignment between the physical world and the current DT. Using the same sensing configuration as the physical measurements, TRACE ray-traces the current DT, coherently backprojects the measured and simulated RF onto a common world grid, and extracts the same local region around each building's current DT position. A multi-view corrector then fuses evidence across sensing nodes and neighboring buildings to predict a gated six-parameter correction per building, without relying on absolute layout or sensor ordering, and supports iterative correction through re-rendering. On 5,400 held-out samples from unseen simulated scenes at 28 GHz, TRACE reduces 3D position RMSE from 2.202 m to 0.302 m and yaw RMSE from 4.978° to 0.894°, outperforming ViT and U-Net baselines under changes in layout, building count, sensing-node count, and SNR. On measured 28 GHz RF data from the NIST outdoor courtyard, a model trained only on synthetic RF reduces mean planar wall-position error from 1.00 m to 7.8 cm, without measured-data fine-tuning or geometric labels. These results show that the discrepancy between measured and twin-rendered RF can serve as a learning signal for repairing a wireless DT.

Figures & tables

Figure 1: TRACE at a glance. (a) A twin with wrong building geometry. (b) Network and twin observe the same acquisition; both are backprojected into one world frame, and GeoCrop makes each building a local alignment problem; the corrector emits a gate and 6D residual, written back and re-rendered. (c) 86.3% of simulated 3D position error removed; $7.8\text{\,}\mathrm{c}\mathrm{m}$ on measured NIST RF.

Figure 2: TRACE architecture: from RF alignment to twin correction. The front end aligns measured and twin-rendered RF in a common world frame and extracts paired GeoCrops. The corrector compares the two RF branches, fuses sensing-node and neighboring-building context, and predicts a gated six-parameter geometric correction.

Table 3

		Input blinding			Scene shuffling
	TRACE $(\boldsymbol{I},\boldsymbol{I}^{\mathrm{twin}})$	$\boldsymbol{I}$ only	$\boldsymbol{I}^{\mathrm{twin}}$ only	Both blinded	Shuffled $\boldsymbol{I}^{\mathrm{twin}}$	Shuffled $\boldsymbol{I}$
3D RMSE [m]	0.302	0.835	2.630 †	2.460 †	0.889	3.723 †

Table 3: Falsifying the mechanism on the frozen model. 3D position RMSE when RF branches are blinded (set to zero before GeoCrop) or replaced by those of a different scene, with the twin prior

\boldsymbol{P}^{(t)}

unchanged. The uncorrected twin is

2.202\text{\,}\mathrm{m}

; † worse than no correction.

Figure 3: Three shifts, one trained model, no retraining. 3D RMSE improvement (% of initial error removed) as (a) building count, (b) SNR, and (c) active sensing nodes move away from training.

Figure 4: Measured-RF validation and downstream impact. (a) Wall-position error using measured $28\text{\,}\mathrm{G}\mathrm{H}\mathrm{z}$ NIST RF. (b) P10 spectral-efficiency retention of beam selection using corrupted and TRACE -corrected twins; the dashed line marks the $4.13\text{\,}\mathrm{m}$ mean OSM footprint offset of Fan et al. (2014) .

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\boldsymbol{p}_{i}^{(t)}=[x_{i},y_{i},h_{i},\ell_{i},w_{i},\psi_{i}]$	state of building $i$ in the current twin
$\boldsymbol{P}^{(t)}=\{\boldsymbol{p}_{i}^{(t)}\}_{i=1}^{N}$	current twin state, $N$ buildings
$\boldsymbol{P}^{\star}$	physical scene, expressed in the same parameterization
$\Delta\boldsymbol{p}_{i}^{(t)}=\boldsymbol{p}_{i}^{\star}-\boldsymbol{p}_{i}^{(t)}$	target residual for building $i$
$\boldsymbol{\xi}_{s}$	known acquisition configuration of sensing node $s$
$\boldsymbol{H}_{s},\ \boldsymbol{H}^{\mathrm{twin}}_{s}$	measured and twin-rendered CSI at node $s$

Appendix

Table 4: Symbols used in the main text and appendix.

Panel of Figure 5 — Figure 5: Representative scene and complementary world-frame backprojections at $z=9m$ . (a) A six-building synthetic scene with three separated rooftop sensing nodes. (b)–(d) Corrupted-twin backprojections from Sensing node 1–3 on the same metric world grid. Dashed boxes indicate the current-twin footprint priors and solid cyan boxes indicate the fixed GeoCrop windows, which are applied identically to the corresponding measured backprojections. The RF response of a building depends strongly on sensing viewpoint: structures that are prominent from one node can be weak or poorly resolved from another, motivating per-node processing followed by cross-sensor fusion.

Input	3D RMSE [m] $\downarrow$
Amplitude only ( TRACE )	0.350
Complex ( TRACE )	0.302

Appendix

Table 5: Complex versus amplitude-only backprojection. 3D position RMSE on the scene-disjoint test split with everything else in the pipeline held fixed; lower is better. The uncorrected twin is

2.202\text{\,}\mathrm{m}

Figure 6: Correction accuracy improves as more backprojection focus heights are available. Mean 3D position RMSE as the number of available focus heights increases. Each point averages over all subsets of the corresponding size. All models are trained with six heights and evaluated with frozen weights.

Method	2D [m] $\downarrow$	Height [m] $\downarrow$	3D [m] $\downarrow$
Uncorrected twin	1.878	1.151	2.202
TRACE , $K=1$	0.720	1.268	1.489
TRACE , $K=2$	0.430	1.157	1.250
TRACE , $K=3$	0.312	0.942	0.999
TRACE , $K=4$	0.255	0.675	0.728
TRACE , $K=5$	0.224	0.455	0.512

Appendix

Table 6: Effect of the number of backprojection focus heights

K

on TRACE . The model is trained with all six focus heights; at test time, only

K

heights are available and the rest are set to zero. Results are averaged over all subsets of each size. Lower is better.

Method	Type	2D RMSE [m] $\downarrow$	Error removed [%] $\uparrow$
Corrupted twin	no correction	1.878	–
Phase correlation	training-free	1.925	$-2.5$
NCC	training-free	1.802	$4.0$
ViT	learned	1.483	$21.0$
U-Net	learned	0.553	$70.6$
TRACE	learned	0.206	$\mathbf{89.0}$

Appendix

Table 7: Classical registration baselines for planar correction. 2D position RMSE on the 5,400-sample scene-disjoint test split at

15\text{\,}\mathrm{d}\mathrm{B}

. Phase correlation and NCC are training-free and operate on the same building-centered spatial regions used by TRACE . Error removed is relative to the uncorrected twin.

Component removed	Replaced by	Par. [M]	RMSE	Cost	Infl. [%]	Removed [%]
uncorrected twin	no correction	–	2.202	–	–	0.0
none	full TRACE	3.98	0.302	–	–	86.2
Cross-sensor attention	mean pooling only	2.93	0.375	$+0.072$	$+23.8$	83.0
Movement gate	raw residual applied	3.98	0.332	$+0.029$	$+9.6$	84.9
Sensor-pose FiLM	unconditioned encoder	3.85	0.317	$+0.014$	$+4.6$	85.6
Coherent residual stream	four-cue CueFusion	3.91	0.315	$+0.012$	$+4.1$	85.7

Appendix

Table 8: Effect of removing individual learned-corrector components. Each row removes one component of the learned corrector and retrains the resulting model from scratch. Par. [M] is the number of trainable parameters, in millions. RMSE is the resulting 3D position error in meters. Cost is the absolute increase in RMSE relative to full TRACE , while Infl. [%] expresses the same increase as a percentage of the full- TRACE RMSE. Removed [%] is the fraction of the initial 3D position error of the corrupted twin that is eliminated after correction.

Corrector	$-5\text{\,}\mathrm{d}\mathrm{B}$	$0\text{\,}\mathrm{d}\mathrm{B}$	$5\text{\,}\mathrm{d}\mathrm{B}$	$10\text{\,}\mathrm{d}\mathrm{B}$	$15\text{\,}\mathrm{d}\mathrm{B}$	$20\text{\,}\mathrm{d}\mathrm{B}$
U-Net on GeoCrops	25.6	55.4	72.6	80.4	83.1	83.7
TRACE	52.8	72.6	81.3	84.9	86.2	86.6

Appendix

Table 9: Same GeoCrop front end, different corrector. Fraction of initial 3D position error removed across evaluation SNR; both models are trained at

15\text{\,}\mathrm{d}\mathrm{B}

. Higher is better.

Model	Worst-case error removed [%] $\uparrow$
TRACE	79.6
U-Net	54.8
ViT	18.1

Appendix

Table 10: Generalization to unseen sensing poses. Minimum fraction of initial 3D position error removed across all tested pose conditions using frozen

15\text{\,}\mathrm{d}\mathrm{B}

checkpoints; higher is better.

Figure 7: Input blinding across geometric parameters. The frozen model is evaluated with the full measured–rendered pair, measured RF only, twin-rendered RF only, and both RF branches blinded. Dashed lines show the corresponding error of the corrupted twin before correction. The 3D position result is reported in Sec. 6.2 ; the remaining panels show footprint length, footprint width, and yaw.

Method	OOD-Layout	OOD-Error
	Removed [%]	Removed [%]
ViT	23.59	10.60
U-Net	56.33	54.59
TRACE , step 1	88.70	79.16
TRACE , step 2	–	93.08

Appendix

Table 11: Performance under out-of-distribution layout and error shifts. Fraction of initial 3D position RMSE removed when either the scene layout or the initial twin error lies outside the training support; higher is better. TRACE step 2 applies the same frozen corrector after updating and re-rendering the twin from step 1.

Panel of Figure 8 — Figure 8: Measured-data benchmark, NIST outdoor courtyard National Institute of Standards and Technology () . (a) Measurement setup with fixed TX and a mobile RX region. (b) Mobile RX platform carrying RF, LiDAR, and camera sensors. (c) The planar recovery task: $M=8$ RX poses sampled in Area 1 with displaced twin walls. LiDAR provides the reference wall geometry; TRACE is trained only on synthetic twin corruptions and evaluated on measured $28\text{\,}\mathrm{G}\mathrm{H}\mathrm{z}$ RF.

Figure 9: Robustness to the number of measured receiver positions. Fraction of the initial wall-position error removed as the number of receiver positions $M$ fused into each estimate varies; the mean planar error is given in parentheses. The same frozen checkpoint, trained with $M=8$ , is evaluated without retraining. The uncorrected twin has $1.00\text{\,}\mathrm{m}$ mean error.

Quantity	Value	Quantity	Value
Carrier frequency	$28\text{\,}\mathrm{G}\mathrm{H}\mathrm{z}$	Sensing nodes $S$	3, monostatic
Bandwidth	$122.88\text{\,}\mathrm{M}\mathrm{H}\mathrm{z}$	Node height	$15.5\text{\,}\mathrm{m}$ (rooftop)
TX array	$8\times 8$	Node downtilt	$12\text{\,}\mathrm{\SIUnitSymbolDegree}$
RX array	$8\times 8$	Node aim point	$(30,0,5)$ m
Ray tracer	Sionna RT	Sensing-node building	$8\times 8\times$ 15\text{,}\mathrm{m}$$
Stage 1: true-scene building parameters

Appendix

Table 12: Sensing configuration and synthetic scene distribution.

Quantity	Value	Quantity	Value
Grid $N_{x}\times N_{y}$	$260\times 198$ px	Grid resolution	$0.4066\text{\,}\mathrm{m}$ / px
Extent in $x$	$[-25.0,80.3]$ m	Extent in $y$	$[-40.0,40.1]$ m
Focus heights $K$	6	Focus height values	$z=0,3,6,9,12,15$ m
GeoCrop size	$56\times 56$ px	GeoCrop extent †	$22.8\text{\,}\mathrm{m}$ square
Image scaling	per scene, shared	Scaling region	town crop, all heights

Appendix

Table 13: World-frame backprojection grid and GeoCrop.

Quantity	Value	Quantity	Value
Token dimension $D_{\mathrm{emb}}$	256	Tokens per crop $T^{\dagger}$	$7\times 7=49$
Complex CNN stages	4	CNN channels (complex)	$6/16/32/64/128$
Cross-attention	1 layer, 4 heads	CueFusion projection	$5D_{\mathrm{emb}}\to D_{\mathrm{emb}}$
Sensor fusion	2 layers, 4 heads	Building attention	3 layers, 4 heads
FiLM initialization	zero (identity at start)	Dropout	0.2
Parameters (full TRACE )	3,979,911 (3.98 M)	Absolute $(x,y)$ in graph	excluded

Appendix

Table 14: Architecture of the learned TRACE corrector

f_{\theta}

Quantity	Value	Quantity	Value
Optimization
Optimizer	AdamW	Learning rate	$1\times 10^{-4}$
Weight decay	$1\times 10^{-3}$	Batch size	128
Schedule	ReduceLROnPlateau	Warmup	8 epochs
Max epochs	200	Early-stop patience	20 epochs
Gradient clipping	1.0 (max norm)	Mixed precision	enabled

Appendix

Table 15: Training and evaluation settings.

Shared
Quantity	Value	Quantity	Value
Input channels †	$S\times 3\times 2\times K=108$	Input size	$198\times 260$
Head trunk	$256+6\to 256$	Heads	$g_{i}$ , $\boldsymbol{r}_{i}$ (6)
Missing sensors	zero-filled channels	Sub-pixel correction	not applicable
U-Net
Base channels	44	Depth	3 enc / 3 dec, bottleneck

Appendix

Table 16: Configurations of the ViT and U-Net baselines.

Explore similar work

KODAMA: Multimodal Digital Twin Reconstruction for Urban RF Propagation Modelling
Sep 7, 2026Maximiliano Wardle, A. Ryo KoblitzRadio Map Estimation Cognitive Digital Twins
Nokia Bell Labs · University of York
Tri-Band Channel Measurement-Enabled Multi-Layer Digital Twin for Terahertz Wireless Data Centers
Sep 1, 2026Mingjie Zhu, Ziming Yu, Guangjian Wang +1Digital Twins Wireless Communications
Terahertz Wireless Communications (TWC) Interdisciplinary Research Center, Shanghai Jiao Tong University, Shanghai 200240, China · Huawei Technologies Company Ltd., Chengdu 611731, China · Terahertz Wireless Communications (TWC) Interdisciplinary Research Center, and Cooperative Medianet Innovation Center (CMIC), School of Information Science and Electronic Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
TWIST: Closed-Loop token Synchronization for Application-Aware Wireless Digital Twins
May 26, 2026Sige Liu, Kezhi WangDigital Twins Wireless Communications
Department of Computer Science, Brunel University London, Uxbridge UB8 3PH, U.K.

Sep 7, 2026, cs.CVJ/K move · Enter open · S save

KODAMA: Multimodal Digital Twin Reconstruction for Urban RF Propagation Modelling

Maximiliano Wardle, A. Ryo Koblitz

Nokia Bell Labs · University of York

Radio Map Estimation Cognitive Digital Twins Digital Twins Digital Elevation Models Random Forest

Open paper AskPDF arXiv

3D reconstruction typically strives for geometric fidelity or visual plausibility. Radio frequency digital twins (RFDT) are instead judged by whether communication channels behave in them as they do in the real world. RFDTs promise site-specific channel prediction but current practice forces a choice between coarse automated scenes and hand-built, measurement-calibrated models that take weeks to construct per-site. We present KODAMA, an automated pipeline that reconstructs ray tracing-ready RFDTs at city scale from off-the-shelf geospatial data alone: aerial imagery, LiDAR, and photogrammetry yield terrain and watertight building meshes, while exposure-weighted multi-view fusion of street-level imagery recovers façade relief, electromagnetic materials, and clutter---all without site visits or calibration. Across three sites spanning 3.6 to 28 GHz, KODAMA's uncalibrated predictions achieve single-digit RMSE, reducing point-to-point error by up to 5.35 dB over automated baselines and coming within 0.22 dB of a measurement-calibrated, hand-built RFDT.

Abstract

Explore similar work

KODAMA: Multimodal Digital Twin Reconstruction for Urban RF Propagation Modelling

Tri-Band Channel Measurement-Enabled Multi-Layer Digital Twin for Terahertz Wireless Data Centers

TWIST: Closed-Loop token Synchronization for Application-Aware Wireless Digital Twins