acceptodds
Under review as a conference paper at ICLR 2027

Learning Frequency-Transferable Radio Digital Twins with Gaussian Ray Tracing

Abstract

Emerging multi-band systems (e.g. WiFi 7) raise the importance of the multi-band radio-frequency digital twin (RFDT), a virtual replica of how radio signals propagate within a physical site, which serves numerous applications. However, existing approaches only focus on a single frequency band, and directly predicting the signal of unseen bands with these DTs does not work because RF propagation varies with frequency. Building an RFDT with multi-band data, on the other hand, incurs huge measurement overhead as the number of bands increases. We therefore propose FresGRT, the first one-to-many-band RFDT built on single-band RF measurements with 3D Gaussian ray tracing. FresGRT first learns the frequency-invariant representation based on the Fresnel equations to achieve cross-band prediction with only single-band data. FresGRT then learns from the frequency-selective fading to decouple the physically coupled representation parameters, and optimizes them separately with a staged training pipeline. We extensively evaluate FresGRT on real testbed and ray-tracing datasets. Results on the testbed show that FresGRT achieves unseen band prediction with a mean error of 4.4 dB, a 45% improvement over the best baseline, while reducing the survey time by 55%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
radio digital twin, Gaussian ray tracing, differentiable rendering, wireless channel prediction, cross-frequency generalization, Fresnel equations, material inversion equations/

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_pURZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FresGRT, a radio digital twin that predicts signals at unseen frequency bands using measurements from a single band. It represents scene geometry with 3D Gaussians and learns material parameters used in a differentiable Fresnel-based ray tracer. A staged training procedure fits line-of-sight attenuation, adjusts geometry using frequency-selective fading, and then fine-tunes the materials. Experiments on three real sites and 24 ray-tracing datasets report lower cross-band received-power errors than existing RF rendering methods.

Critical strengths
  1. The method learns material properties rather than a signal representation tied to one frequency. Combining these parameters with Fresnel equations and Gaussian ray tracing allows the model to predict received power at unseen bands without collecting new RF measurements.

  2. The evaluation includes both real measurements and simulated environments. FresGRT reduces cross-band received-power error compared with existing RF rendering methods, and the ablation results show improvements from the geometry and material training stages.

Critical weaknesses
  1. The baseline comparison does not fully isolate the benefit of the proposed learning method. Most compared RF representations are tied to their training frequency, so poor performance at unseen bands is expected. A conventional ray tracer with material parameters calibrated from the same single-band measurements and initial geometry is an important missing reference. Without it, the contribution of the Gaussian representation and staged training is difficult to separate from that of the frequency-dependent propagation equations.

  2. The evaluation focuses on band-mean received-power error, although the renderer models frequency-dependent channel responses and the training method uses frequency-selective fading. The accuracy of per-subcarrier power and complex CSI at unseen bands is not evaluated. The reported results support received-power prediction, but do not establish whether the model is accurate enough for applications requiring detailed channel information.

Questions for authors
  1. Is a comparison available with a conventional ray tracer whose material parameters are calibrated using the same single-band measurements and initial geometry? If not, how do the current comparisons separate the contribution of the proposed renderer and training procedure from the underlying propagation equations?

  2. Are unseen-band results available for per-subcarrier power or complex CSI, rather than only band-mean received power? Which intended applications can be supported by the reported evaluation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pURZ

(Q1, W1) Conventional ray tracer. We thank the reviewer for pointing out this missing reference. We now add Sionna RT, a conventional ray tracer whose ITU materials are calibrated with the same single-band links and the same initial geometry, to Tables 1, 2, 10, and 11 of the revision. Averaged over all cross-band testing, Sionna RT reaches 11.3 dB on the real sites and 11.4 dB on the ray-tracing datasets, while FresGRT reaches 4.4 dB and 4.3 dB. Since Sionna RT shares the frequency-dependent propagation equations with FresGRT, this gap comes from the Gaussian representation and the staged training, which correct the initial geometry with the RF measurements.

Table 1: MAE of the band-mean power (dB) averaged on the real sites (revised). *In-band.

Train → test (GHz)2.4 → 2.4*2.4 → 52.4 → 5.85 → 5*5 → 2.45 → 5.8
Sionna RT7.6713.9311.736.729.6610.01
FresGRT3.604.314.963.763.924.32

(Q2, W2) Per-subcarrier power and complex CSI. We thank the reviewer for this suggestion. We now evaluate the per-subcarrier power of FresGRT, i.e. the error of the power of each subcarrier averaged over the subcarriers, in Table 13 of the revision. On the real sites, the per-subcarrier error is 5.2 dB in-band and 5.5 dB cross-band, only 1.3 dB and 1.2 dB above the band-mean error. On the ray-tracing datasets, the per-subcarrier error is 4.5 dB in-band and 5.1 dB cross-band, only 0.5 dB and 0.5 dB above the band-mean error. FresGRT therefore also predicts the power of each subcarrier at unseen bands.

Table 13: MAE (dB) of the band-mean power and of the per-subcarrier power of FresGRT on the real sites (revised). *In-band.

Train → test (GHz)2.4 → 2.4*2.4 → 52.4 → 5.85 → 5*5 → 2.45 → 5.85.8 → 5.8*5.8 → 2.45.8 → 5
Band-mean power3.604.314.963.763.924.324.214.533.99
Per-subcarrier power5.305.506.064.915.295.455.385.625.01

We use the per-subcarrier amplitude instead of the complex CSI as the target, since the measured phase of commodity Wi-Fi hardware is corrupted by carrier frequency offset (CFO), sampling frequency offset (SFO), and other phase offsets. The target of FresGRT is the signal strength prediction at any bands. In practical deployments, most network management decisions, such as access point placement, coverage planning, band steering, and roaming, rely on the band-mean power, i.e. the received signal strength indicator (RSSI), rather than on the power of individual subcarriers. The well-trained cross-band digital twin simulates in software the signal strength between any transmitter and receiver configurations (any positions, any bands). Network planners can therefore place the devices, e.g., multi-band access points, with the DT before deployment, without collecting measurements for each transmitter and receiver configuration, which can take several hours of measurements.

I maintain my original score. The calibrated baseline and additional evaluations substantially address my concerns and support the reported cross-band gains. However, the Gaussian representation’s contribution remains difficult to isolate, and the shared simulation assumptions and limited real-world validation leave broader robustness uncertain.

AuthorsOpenReview ↗

Follow-up to Reviewer pURZ

We thank the reviewer for reading our response and for the positive assessment. We briefly clarify the remaining points.

Contribution of the Gaussian representation. With the initial geometry, as the reviewer asked, Sionna RT reaches a cross-band error of 11.3 dB on the real sites and 11.4 dB on the ray-tracing datasets, as reported in our first response. In the latest revision, we further give Sionna RT the faces of FresGRT after the LoS stage, i.e. with the spurious faces already pruned, with the same ITU material model and the same training links. Its cross-band error then becomes 7.8 dB and 7.9 dB, while FresGRT reaches 4.4 dB and 4.3 dB. The remaining difference is that Sionna RT traces fixed rectangular meshes, whereas FresGRT continues to adjust the position and the thickness of each face in its geometry stage.

Table 1: MAE of the band-mean power (dB) averaged on the real sites (revised). The initial-geometry row is from our first response. *In-band.

Train → test (GHz)2.4 → 2.4*2.4 → 52.4 → 5.85 → 5*5 → 2.45 → 5.8
Sionna RT (initial geometry)7.6713.9311.736.729.6610.01
Sionna RT (pruned faces)5.947.348.095.127.088.50
FresGRT3.604.314.963.763.924.32

Shared simulation assumptions. Although Sionna RT is the ray tracer that generated the datasets and shares every physical model with them, its cross-band error on these datasets is 7.9 dB, while FresGRT reaches 4.3 dB. We think that sharing the physics does not explain the better performance of FresGRT. More importantly, FresGRT outperforms the best baseline by 30% on the three real sites, which share no simulation assumption with FresGRT.

Reviewer_c8oNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces FresGRT for predicting RF propagation across frequency bands. It initializes a scene representation from visual/depth measurements, then fits geometry and material parameters using RF measurements from one band. Fresnel-based propagation equations allow the fitted representation to predict received power at other bands. Experiments on three real indoor sites and simulated environments show improved cross-band prediction over the evaluated baselines. The paper also reports reduced survey time by collecting RF training measurements at only one band.

Critical strengths
  1. The representation provides a physical basis for frequency transfer.

  2. The evaluation demonstrates real-world cross-band measurements and performance improvements.

Critical weaknesses
  1. Access to visual information is not clearly matched across methods. FresGRT uses visual/depth information to initialize scene geometry. The paper does not clearly explain whether the baselines receive equivalent information or how they are initialized. This makes it difficult to separate the benefit of the method from the benefit of additional geometric input.

  2. Robustness and maintenance under environmental changes are untested. The fitted model describes a static surveyed environment. People moving, doors opening, and furniture changes can alter propagation, but the paper does not evaluate the resulting errors or the cost of updating the model.

  3. The practical advantage over channel measurements is unclear. Given the long survey, additional modeling and compute, as well as the static environment assumption, it's not clear if the proposed system is any better than the straightforward on-demand channel measurement.

Questions for authors
  1. How much does accuracy degrade when people, doors, or furniture change? What additional measurements and computation are required to update the model?

  2. Do all baselines receive the same visual geometry and comparable initialization? If not, how much of the improvement comes from this additional information?

4 comments
AuthorsOpenReview ↗

Response to Reviewer c8oN

(Q1, W2) Environmental changes. We thank the reviewer for this practical question. We add an experiment on scene changes in the revision (Section 5.2, Table 5). The Gaussians of FresGRT are editable. When an object in the scene changes, e.g., a piece of furniture is moved, or a door is opened or closed, FresGRT detects the change with a camera and edits only the corresponding Gaussians, whose material parameters are already learned. This needs no new RF measurement and no training, and the editing takes at most 0.22 s on an Intel Core i9-10900X CPU.

We evaluate three changes on a ray-tracing scene (RT 7): closing a doorway, opening a new door, and moving a wall. Averaged over all training bands, the MAE over all links is 2.44 dB to 3.07 dB after the changes, compared with 2.60 dB on the original scene. On the links whose power changes by more than 3 dB, the MAE is 3.31 dB to 4.23 dB.

MAE (dB) of FresGRT after the scene changes, without new RF measurements or training.

SceneAll linksChanged links (>3 dB)
Original2.60–
(a) Close a doorway2.443.31
(b) Open a new door2.593.69
(c) Move a wall3.074.23

If a new object with an unknown material is added to the scene, only the material parameters of its Gaussians need to be fine-tuned, with new measurements at a single band on the links around the object. As for moving people, FresGRT is designed for applications in mostly static environments, such as access point placement. A moving person affects only the links that the person blocks, and only for a short time. We therefore do not model moving people in this paper, and we list it as future work in the revision.

(Q2, W1) Visual geometry of the baselines. We thank the reviewer for raising this point, and we apologize for not stating it clearly in the paper. We now detail the implementation of every baseline in the appendix of the revision. Except NeRF² and EchoNeRF, all baselines use the visual information as initialization. RFScape supervises its signed distance function with the vision point cloud, and GSpaRC and WRF-GS+ initialize their 3D Gaussians with the vision point cloud. The newly added COST-231 and Sionna RT use the faces of FresGRT as the geometry.

(W3) Advantage over on-demand measurement. We thank the reviewer for this comment. We agree that on-demand measurement is accurate at the measured positions and bands. However, it only covers the positions and bands that have been measured. In our testbed, measuring one band takes 30 s per position, including hardware reinitialization, reconnection between the transmitter and the receiver, and a period of measurement. FresGRT needs the measurements of only one band, which costs only O(1)O(1) of survey, while on-demand measurement requires O(N)O(N) of survey for NN bands. Once trained, FresGRT renders every link of the space in seconds, e.g., 27,378 links in 9.1 s in Table 17 of the revision, while on-demand measurement requires to take measurements at unseen positions.

Table 17: Rendering time of FresGRT on every scene (revised).

SceneGaussiansFacesSubcarriersOne positionAll pairs
RT 14,0922351100 ms3.46 s (12,223)
RT 22,635175179 ms490 ms (3,015)
RT 32,277215181 ms500 ms (2,192)
RT 42,1762251104 ms1.28 s (4,236)
RT 53,188155184 ms1.59 s (8,820)
RT 64,4942251106 ms9.10 s (27,378)
RT 71,637145166 ms100 ms (784)
RT 82,662185180 ms510 ms (4,081)
Site 14,2261511260 ms200 ms (441)
Site 23,346811258 ms110 ms (423)
Site 34,1961911287 ms340 ms (522)

This gives network planners a fast way to evaluate the coverage of every band before deployment. After deployment, the measurements from data transmissions could further fine-tune the twin for downstream services.

(Edited: table numbers updated to match the latest revision.)

Dear Reviewer c8oN,

We would like to check whether our response and the revision have addressed your concerns. In brief, the revision adds an experiment on scene changes (Section 5.2, Table 5), describes how every baseline uses the vision data (Appendix D), and reports the rendering time of every scene (Table 17).

We are happy to provide further clarification before the discussion period ends. Thank you for your time.

Thank you for the detailed response. It was not clear how the scene change detection will be automatically detected or edited and whether the reduced number of measurements can justify the cost/complexity here. I will maintain my oringal score.

AuthorsOpenReview ↗

Follow-up to Reviewer c8oN

We thank the reviewer for the follow-up. We apologize that our first response did not explain the scene-change procedure clearly, and we clarify the two points below.

Detection and editing of scene changes. We follow the approach of Argos [1]. The Gaussians are segmented into objects with SAGD [2], so that every Gaussian has an object ID. We use surveillance cameras of different views to track the objects in the scene. When the scene changes, SAM 2 [3] tracks the mask of each object in the camera streams. To obtain the motion of the object, ICP [4] aligns its 3D points before the change with those after the change, which gives the rotation and the translation of the object. FresGRT applies this motion to the Gaussians of the object, which keep their learned materials. A removed or added surface is handled in the same way by deleting or adding its Gaussians. Table 5 of the revision reports the error after the changes. We have added this procedure, from detecting the change to moving the Gaussians, to Section 5.2 and Appendix F of the revision.

We would like to emphasize again that the contribution of FresGRT is the one-to-many-band radio digital twin itself, rather than how to detect and track the changes of the scene. We therefore implement the object detection and tracking with existing methods. Once the new position of an object is known, FresGRT edits the Gaussians and predicts the channel by ray tracing.

Measurements and computation after a scene change. Since the moved Gaussians keep their learned materials, FresGRT needs no new RF measurement to fine-tune the model after a scene change. The only computation is editing the Gaussians, which takes at most 0.22 s on an Intel Core i9-10900X CPU. After the editing, FresGRT renders any position in at most 106 ms on one NVIDIA Quadro RTX 4000 (Table 18 of the revision). On-demand measurement, in contrast, has to measure each requested position, which takes 30 s per position and band in our testbed (i.e. 150 s per position for the five bands of the ray-tracing datasets). Depending on how many links the change affects, 48% to 70% of the positions have to be measured again in our experiment (Table 17 of the revision).

Table 17: Changed links and affected positions of the scene changes on RT 7 (784 links and 196 positions).

Scene changeChanged linksAffected positions
(a) Close a doorway176 (22.5%)137 (70%)
(b) Open a new door99 (12.6%)94 (48%)
(c) Move a wall148 (18.9%)112 (57%)

On-demand measurement in general. Beyond scene changes, on-demand measurement in general cannot serve real-time applications. In our testbed, measuring one band at one position takes 30 s (i.e. 90 s per position for the three bands), which includes the hardware reinitialization, the reconnection between the transmitter and the receiver, and the measurement. This cost is proportional to the number of requested positions. For example, access point placement compares many candidate locations of the access point. For each candidate, on-demand measurement has to measure all positions again, which takes 4.4 hours, e.g. on Site 3 (174 positions, 90 s each), while FresGRT renders each position in at most 106 ms. This speed comes at a one-time cost, i.e. the survey of one band and the training. The training takes 0.5 to 1.9 hours on the real sites on one NVIDIA Quadro RTX 4000 (Table 9 of the revision) and runs offline.

We hope that these clarifications and the additions to the revision address the remaining concerns, and we would be grateful if the reviewer could reconsider the score in light of them.

[1] A. Datta et al., "ARGOS: Leveraging Visual Priors for Scalable Wireless Navigation in Dynamic Environments," PerCom 2026.

[2] X. Hu et al., "SAGD: Boundary-Enhanced Segment Anything in 3D Gaussian via Gaussian Decomposition," arXiv:2401.17857, 2024.

[3] N. Ravi et al., "SAM 2: Segment Anything in Images and Videos," arXiv:2408.00714, 2024.

[4] P. J. Besl and N. D. McKay, "A Method for Registration of 3-D Shapes," IEEE TPAMI, 1992.

Reviewer_PmRZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FresGRT is a radio-frequency digital twin that predicts received power in frequency bands it was not trained on, using measurements from a single band. Gaussians are placed on planar faces extracted from an RGBD reconstruction, and each material group carries a relative permittivity and a conductivity that follow the ITU-R P.2040 frequency law. A differentiable Gaussian ray tracer renders the LoS path and specular reflections up to two bounces, with Fresnel/ITU slab transmission and reflection. Training proceeds in three stages: 1) LoS, 2) geometry, and 3) material fine-tuning. On three real setups and Sionna ray-tracing datasets, the paper reports cross-band MAE of around 4 dB, 45% and 52% below the best baseline, and a 55% reduction in survey time.

Critical strengths
  1. A principled design. Cross-band transfer comes from physics (Fresnel equations and the ITU material model) rather than from a band-specific radiance, so the route to unseen bands is clear.

  2. A real testbed. The authors built a robotic RGBD + Nexmon CSI platform, measured three sites in three bands, implemented the renderer in CUDA (about 100 ms per position), and released code and data. The paper is well written.

Critical weaknesses

The weaknesses that weigh most on my rating are W1, W3, and W4.

W1. All baselines are frequency-agnostic. The five baselines (RFScape, NeRF², EchoNeRF, GSpaRC, WRF-GS+) take no frequency input, so at an unseen band they can only reproduce their training-band output. Free-space path loss alone differs by 6.4 dB from 2.4 to 5 GHz and by 7.7 dB from 2.4 to 5.8 GHz, which accounts for much of their cross-band error (8–13 dB). The paper includes no frequency-aware baseline:

  • the learned baselines with a Friis correction, 20·log10(f′/f);
  • empirical indoor models with frequency and wall-loss terms (e.g., ITU-R P.1238, COST-231 multi-wall);
  • differentiable ray tracing with learnable ITU materials on the same reconstructed geometry. Because Gaussians are grouped into planar faces that share parameters, a mesh-based differentiable ray tracer is also the natural ablation of the Gaussian representation. Without these, the 45% gain cannot be separated from the generic advantage of any physics-based model.

W2. The claim to be "the first one-to-many-band RFDT" is a bit overstated. Predicting the channel in one band from measurements in another is well established in FDD systems (e.g., R2F2, FIRE, Diffusion^2). Recent learned radio-map work addresses cross-band signal prediction (e.g., CommUNext). Differentiable ray tracers with learnable ITU materials (e.g., Hoydis et al., "Learning Radio Environments by Differentiable Ray Tracing") are frequency-transferable by construction. None of these is cited or compared against. The contribution should be stated more precisely.

W3. The ray-tracing evaluation seems circular. The simulated data are generated by Sionna with ITU materials and slab-model transmission, and FresGRT uses the same ITU-R P.2040 frequency law and slab model. Good transfer on these datasets (52%) is therefore largely expected. The meaningful evidence is the real-site evaluation, which covers only three sites.

W4. The cross-band test includes training positions. The paper uses "80% of the positions as training data (rest for in-band testing, all positions for cross-band testing)." Most cross-band test links were therefore seen during training, in another band, and a model-free baseline (the measured training-band power at the same position plus a frequency correction) becomes possible. Cross-band errors on held-out positions are not reported, so joint spatial and frequency generalization is untested.

W5. The propagation model is incomplete, and this may bias transfer. The renderer models the LoS path and specular reflections up to two bounces, and drops two-bounce paths that leave the Tx/Rx room. It has no diffraction or diffuse scattering, and represents each scene with only 8–23 faces. The paper gives no reflection-order ablation, no LoS/NLoS breakdown, and no analysis of the mismatch with the ray-tracing data, which was generated with up to five reflections. Energy from unmodeled mechanisms is likely absorbed into the learned materials and the gain. Because diffraction and scattering depend on frequency differently from the slab model, this compensation can distort exactly the cross-band transfer the method relies on.

W6. Real-world calibration and the robustness mechanisms are not validated. It is unclear how K (transmit power, antenna, and radio-chain gain) is set for an unseen band; if it is fitted with target-band data, the prediction is not zero-shot. Also undescribed are AGC compensation for Nexmon CSI amplitudes and frequency-dependent antenna patterns. The coherence factors, the Fresnel-zone gate, and the learnable face offsets and occupancy are not ablated; only the training stages are.

W7. The scope and metrics are limited.

  • The real bands are 2.4, 5, and 5.8 GHz; 5 and 5.8 GHz are close, and the motivating 6 GHz band is absent.
  • The only metric is the MAE of band-mean power, although the method models frequency-selective fading.
  • FresGRT is not better in-band (e.g., 3.60 vs. 3.45 dB for NeRF² at 2.4 GHz).
  • Material recognition reaches 72–73%, and only 44–62% on exterior walls.
Questions for authors
  1. How does FresGRT compare with frequency-aware baselines: the learned baselines with a Friis correction, ITU-R P.1238 or multi-wall models, and differentiable ray tracing with ITU material calibration on the same geometry?
  2. What are the cross-band errors on the held-out 20% of positions only?
  3. How is K obtained for an unseen band, and how are AGC and antenna-pattern effects handled?
  4. How does performance change with the reflection order, and what are the cross-band errors for LoS and NLoS links separately?
  5. How accurate is the predicted per-subcarrier response (fading shape)? Can you add 6 GHz measurements?
7 comments
AuthorsOpenReview ↗

Response to Reviewer PmRZ (1/2)

(Q1, W1) Frequency-aware baselines. We thank the reviewer for the detailed suggestions on frequency-aware baselines. We add three frequency-aware baselines to Tables 1, 2, 10, and 11 of the revision, which are RFScape (the best learned baseline) with the Friis term 20log⁡10(f′/f)20\log_{10}(f'/f), the COST-231 multi-wall model, and Sionna RT with ITU materials calibrated on the same initial geometry and training links. Averaged over all cross-band testing on the real sites, they reach 6.3 dB, 7.2 dB, and 11.3 dB, while FresGRT reaches 4.4 dB. On the ray-tracing datasets, they reach 7.8 dB, 7.5 dB, and 11.4 dB, while FresGRT reaches 4.3 dB. COST-231 relies on empirical penetration losses, so it cannot learn the frequency-transferable properties of the objects or capture the multipath effects. Sionna RT traces the same faces as FresGRT, so it also serves as the mesh-based ablation of the Gaussian representation. RFScape takes the frequency as an input of its material network, and it does not learn a frequency trend from one band.

Table 1: MAE of the band-mean power (dB) averaged on the real sites (revised). *In-band.

Train → test (GHz)2.4 → 2.4*2.4 → 52.4 → 5.85 → 5*5 → 2.45 → 5.8
RFScape + Friis3.666.146.493.656.805.66
COST-2315.107.198.115.336.706.84
Sionna RT7.6713.9311.736.729.6610.01
FresGRT3.604.314.963.763.924.32

(Q2, W4) Held-out positions. We thank the reviewer for this question. We report the cross-band error on the held-out positions only, i.e. the 20% of the positions whose measurements are never used in training, in Table 12 in the revision. Averaged over all cross-band testing on the real sites, FresGRT reaches 4.4 dB on the held-out positions and 4.3 dB on all positions. On the ray-tracing datasets, FresGRT reaches 4.6 dB on the held-out positions and 4.6 dB on all positions. FresGRT therefore predicts the unseen bands at the unseen positions as accurately as at the training positions.

Table 12: MAE of the band-mean power (dB) of FresGRT in the cross-band scenario on the real sites. Simulation results are in the revision paper.

Train → test (GHz)2.4 → 52.4 → 5.85 → 2.45 → 5.85.8 → 2.45.8 → 5
All positions4.314.963.924.324.533.99
Held-out positions4.185.174.094.584.473.92

(Q3, W6) Gain, AGC, and antenna. These practical effects are indeed important for a real deployment, and we thank the reviewer bring up this. Since we use the same system and hardware to measure every band, we assume that the system gain is constant across bands and learn it as κ\kappa on the training band. Our metric is the band-mean received power, which corresponds to the commonly used RSSI. The receiver measures RSSI by subtracting the gain of its automatic gain control from the power of the amplified signal, so RSSI reflects the power received at the antenna regardless of the automatic gain control. As for the antenna directivity, our current design assumes antennas that radiate equally in all directions, so κ\kappa only captures the average antenna gain. Since FresGRT traces every path with its departure and arrival directions, a known antenna directivity at each frequency can be applied to each path before the paths are summed, which also extends FresGRT to multi-antenna systems. We leave this extension to future work.

(Q4, W5) Reflection order and LoS/NLoS. Table 14 of the revision reports the error of FresGRT with different reflection orders, averaged over all bands and all positions. The error decreases from 4.38 dB without reflections to 3.90 dB with the second-order reflections on the real sites, and from 4.72 dB to 4.52 dB on the ray-tracing datasets. The second-order reflections reduce the error by only 0.12 dB on the real sites, so the higher orders have a small effect. We list the higher-order reflections, the diffraction, and the diffuse scattering as future work in the revision.

Table 15 of the revision reports the LoS and NLoS links separately. The LoS links reach 3.78 dB on the real sites and 3.48 dB on the ray-tracing datasets, while the NLoS links, whose signals penetrate objects, reach 4.01 dB and 5.46 dB.

Tables 14 and 15: MAE (dB) of FresGRT with different reflection orders and on the LoS and NLoS links.

Refl. orderRealRT
04.384.72
14.024.59
2 (full)3.904.52
LinkRealRT
LoS3.783.48
NLoS4.015.46
AuthorsOpenReview ↗

Response to Reviewer PmRZ (2/2)

(Q5, W7) Per-subcarrier response and 6 GHz. Table 13 of the revision reports the per-subcarrier power error of FresGRT. On the real sites, this error is 5.2 dB in-band and 5.5 dB cross-band, 1.3 dB and 1.2 dB above the band-mean error. On the ray-tracing datasets, it is 4.5 dB in-band and 5.1 dB cross-band, 0.5 dB and 0.5 dB above the band-mean error. The additional error comes from the frequency-selective fading, which FresGRT approximates with reflections up to the second order.

Our IoT testbed is built on Nexmon CSI, which cannot capture the 6 GHz band. However, the 5.8 GHz band that we already evaluate is close to 6 GHz. In addition, a test band at 7.1 GHz, the upper end of the 6 GHz band of Wi-Fi 6E, is added to all 24 ray-tracing datasets (Tables 2 and 11 of the revision). Trained at 2.4 GHz and at 5 GHz, FresGRT reaches 5.35 dB and 5.58 dB at 7.1 GHz, while the best baseline reaches 9.09 dB and 7.53 dB.

(W2) The "first" claim. We thank the reviewer for pointing out these works. We cite all of them in the revision.

R2F2 and FIRE infer a link from the same link measured at a nearby band, so they cannot predict a position without measurements. CommUNext and Diffusion² are trained with the data of multiple bands, including the target band or the adjacent channels. Hoydis et al. calibrate the materials on a given fixed 3D model with the measurements of a single band. In contrast, FresGRT learns from the frequency-selective fading to adjust the 3D model together with the materials, which enables the one-to-many-band prediction. As a result, the Sionna RT baseline has a cross-band error of 11.3 dB on the real sites and 11.4 dB on the ray-tracing datasets, while FresGRT reaches 4.4 dB and 4.3 dB.

Given this, we still consider FresGRT the first one-to-many-band RFDT, but we are happy to tone down the claim.

(W3) Circular ray-tracing evaluation. We agree that the ray-tracing datasets and FresGRT share the ITU frequency law and the slab model. However, FresGRT does not receive the information that generated the data. It starts from the geometry reconstructed from the vision data instead of the ground-truth floor plan, learns the materials and the thicknesses from one band without knowing the material classes, and renders at most two reflections while the datasets contain up to five. To test whether sharing the physics alone explains the results, we add Sionna RT, the ray tracer that generated the datasets, with ITU materials calibrated on the same initial geometry and training links. Although Sionna RT shares every physical model with the data generator, its cross-band error is 11.4 dB on the ray-tracing datasets, while FresGRT reaches 4.3 dB. This shows that the advantage of FresGRT does not come from sharing the physics of the data generator.

Thank you for the thorough response and the substantial new experiments. The frequency-aware baselines and the held-out results address W1 and W4 well. However, I have three remaining points:

  1. Sionna RT baseline. I appreciate this addition, but the gap seems to come mainly from the calibration setup not from the Gaussian representation or the geometry correction. In App D, Sionna RT learns materials only for the reflecting faces, the remaining faces share one material, and thicknesses and spurious faces are not adjusted. Its in-band error (7.67 dB at 2.4 GHz) is already higher than that of COST-231 (5.10 dB), whereas FresGRT with the face positions locked (LoS stage only, Table 4) reaches 5.27 dB even cross-band. So, I think running Sionna RT with the same material groups and FresGRT's pruned faces would isolate the contribution of the representation.

  2. Reflection order. I am curious how much multipath contributes to the cross-band accuracy. Rendering the trained model without reflections raises the error only from 3.90 to 4.38 dB (Table 14), and this average mixes in-band and cross-band links. Cross-band numbers per reflection order, or a flat reference for the per-subcarrier error, would show the contribution of the frequency-selective fading.

  3. Circularity (W3). This is partly addressed. The Sionna RT result shows that the shared physics alone is not sufficient, but not how much it helps FresGRT. Since γ is fixed to 0.9 while the ray-tracing materials range from 0.78 to 1.34, the cross-band errors per interior material (like concrete, plasterboard, wood) from the existing datasets would show how sensitive the transfer is to a mismatch in the frequency law.

Reviewer_jHXhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FresGRT builds a frequency-transferable RF digital twin from measurements across subcarriers in one band and visually initialized geometry. Gaussians are grouped into planar faces and material groups, with Fresnel coefficients and an ITU slab model governing transmission and reflection. The model learns reference-frequency permittivity/conductivity and fixes the conductivity frequency-law exponent to 0.9. Training proceeds through LoS occupancy/material fitting, geometry fitting using frequency-selective fading, and material fine-tuning. On three real sites and 24 ray-tracing datasets, the main cross-band experiments report band-mean-power MAEs of approximately 4.4 and 4.0 dB, improving over the selected best baselines by 45% and 52%, with approximately 55% less survey time.

Critical strengths
  1. The representation fits cross-band prediction: material parameters with an explicit frequency law and Fresnel/ITU transmission/reflection models produce large cross-band gains over the selected neural-rendering/Gaussian-splatting baselines while maintaining comparable in-band accuracy. Three real sites and 24 simulated datasets provide valuable empirical support (Tables 1–2). A single source band includes multiple within-band subcarriers; this remains meaningfully different from cross-frequency prediction using multiple source frequencies or target-band anchors.
  2. Existing controls quantify training-stage contributions. Table 4 includes LoS, LoS plus geometry, LoS plus fine-tuning, and the full pipeline. The geometry-stage-omitted LoS-plus-fine-tuning variant is a useful internal control: adding the geometry stage reduces cross-band MAE from 4.81 to 4.64 dB on real sites and from 4.32 to 4.11 dB in simulation. Appendix C specifies stage iterations and learning rates, and Appendix D's 110-second per-position measurement breakdown supports the approximately 55% survey reduction.
Critical weaknesses
  1. The external comparisons do not isolate the gain over a physically calibrated ray-tracing alternative. Hoydis et al. (2023) already learn permittivity and conductivity from channel measurements. FresGRT's more specific contribution combines physical materials, Gaussian geometry, single-source-band transfer, and staged fitting. Tables 1–2 compare against the listed neural-rendering/Gaussian methods, while Table 4 usefully isolates stages inside FresGRT, including a fixed-geometry LoS-plus-fine-tuning control. Neither comparison quantifies the additional value over non-Gaussian physical calibration using comparable geometry information and source-band data. The reported 45%/52% gains should remain explicitly tied to the evaluated baseline set; a matched physical comparison would strengthen attribution.
  2. Equation (9), as printed, is not the stated Huber loss. For positive delta, x²/2 − (delta|x| − delta²/2) = (|x| − delta)²/2 ≥ 0, so min(x²/2, delta|x| − delta²/2) selects the linear branch for every x. This is a definite error in the written definition, but it does not establish an implementation error. Clarifying the actual loss is important because it is reused in the training pipeline (Equations 9–12). A smaller reporting issue is that Table 11 and Table 12 give different totals for corresponding scenes, and the strict 100 ms wording on page 9 does not hold for the 106–113 ms entries.
Questions for authors
  1. Are comparable-source-band results available for a physically calibrated non-Gaussian ray tracer using the same available geometry information? If not, please state explicitly that the quantitative advantage has been established against the selected neural/Gaussian baselines, and explain the remaining attribution limit.
  2. Is the min expression in Equation (9) a typesetting error? Please state the loss actually used in the experiments, preferably with the corresponding implementation reference, and indicate whether correcting the written expression changes any reported result.
  3. For the same scene, why do Tables 11 and 12 give different rendering totals—for example, 76 ms versus 60 ms for Site 1? Please identify any difference in reflection order, workload, batching, or measurement procedure, and qualify the strict 100 ms statement where necessary.
2 comments
AuthorsOpenReview ↗

Response to Reviewer jHXh

(Q1, W1) Physically calibrated ray tracer. We thank the reviewer for this suggestion, which helps attribute the gain of FresGRT. We include Sionna RT, a physically calibrated non-Gaussian ray tracer, to Tables 1, 2, 10, and 11 of the revision. Sionna RT uses the faces of FresGRT after the LoS stage as its geometry, and its ITU materials are calibrated on the same training links of the source band. Averaged over all cross-band testing, Sionna RT reaches 7.8 dB on the real sites and 7.9 dB on the ray-tracing datasets, while FresGRT reaches 4.4 dB and 4.3 dB. Sionna RT traces a fixed mesh and can only calibrate the materials. FresGRT also adjusts the geometry with the RF measurements, which a fixed mesh cannot do.

Table 1: MAE of the band-mean power (dB) averaged on the real sites (revised). *In-band.

Train → test (GHz)2.4 → 2.4*2.4 → 52.4 → 5.85 → 5*5 → 2.45 → 5.8
Sionna RT5.947.348.095.127.088.50
FresGRT3.604.314.963.763.924.32

(Q2, W2) Huber loss of the level loss, Eq. (9). We apologize for the typesetting error in the Huber loss of the level loss and thank the reviewer for catching this error. The correct equation is

huber(x)=12x2\mathrm{huber}(x)=\tfrac12 x^2 if ∣x∣≤δ|x|\le\delta, and huber(x)=δ∣x∣−12δ2\mathrm{huber}(x)=\delta|x|-\tfrac12\delta^2 if ∣x∣>δ|x|>\delta,

with δ=6\delta=6 dB, which is the loss used in our experiments and implemented in our released code. The same Huber loss is used in the depth and anchor losses of Eq. (12), so these are also correctly defined by the corrected equation. We have corrected this equation in the revision.

(Q3, W2) Rendering time. We appreciate the careful reading that caught this inconsistency, and we apologize for the confusion. Tables 11 and 12 came from two separate timing measurements of the same scenes, reflection order, and workload. Table 11 reported the median of 5 repetitions from an earlier run, while Table 12 reported the mean of 20 repetitions, which is why Site 1 shows 76 ms and 60 ms. In the revision, both tables report the same measurement, so Site 1 takes 60 ms in both tables. Tables 11 and 12 of the submission are Tables 18 and 19 of the revision, which are listed below.

Table 18: Rendering time of FresGRT on every scene (revised).

SceneGaussiansFacesSubcarriersOne positionAll pairs
RT 14,0922351100 ms3.46 s (12,223)
RT 22,635175179 ms490 ms (3,015)
RT 32,277215181 ms500 ms (2,192)
RT 42,1762251104 ms1.28 s (4,236)
RT 53,188155184 ms1.59 s (8,820)
RT 64,4942251106 ms9.10 s (27,378)
RT 71,637145166 ms100 ms (784)
RT 82,662185180 ms510 ms (4,081)
Site 14,2261511260 ms200 ms (441)
Site 23,346811258 ms110 ms (423)
Site 34,1961911287 ms340 ms (522)

Table 19: Rendering time breakdown of one position.

SceneDirect pathReflectionsOtherTotal
RT 127 ms (27%)64 ms (64%)9 ms (9%)100 ms
RT 223 ms (29%)49 ms (62%)7 ms (9%)79 ms
RT 322 ms (27%)51 ms (63%)8 ms (10%)81 ms
RT 427 ms (26%)70 ms (67%)7 ms (7%)104 ms
RT 520 ms (24%)53 ms (63%)11 ms (13%)84 ms
RT 626 ms (25%)68 ms (64%)12 ms (11%)106 ms
RT 719 ms (29%)41 ms (62%)6 ms (9%)66 ms
RT 821 ms (26%)52 ms (65%)7 ms (9%)80 ms
Site 113 ms (22%)46 ms (77%)1 ms (2%)60 ms
Site 219 ms (33%)38 ms (66%)1 ms (2%)58 ms
Site 324 ms (28%)62 ms (71%)1 ms (1%)87 ms

We also clarify that this time is for rendering a single transmitter-receiver link, which corresponds to real-time rendering at the position of a user. We have corrected the statement on the 100 ms requirement, since two ray-tracing scenes take 104 ms and 106 ms.

(Edited: the Sionna RT baseline was rerun on the faces of FresGRT after the LoS stage. Its numbers, its description, and the table numbers are updated to match the latest revision.)

Dear Reviewer jHXh,

We would like to check whether our response and the revision have addressed your concerns. In brief, the revision adds the physically calibrated ray tracer (Sionna RT) as a baseline in Tables 1, 2, 10, and 11, corrects the Huber loss in Eq. (9), and reports one consistent rendering-time measurement in Tables 18 and 19.

We are happy to provide further clarification before the discussion period ends. Thank you for your time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General response. We thank all reviewers for their valuable comments. Following the suggestions, the revision includes the following changes.

1. Frequency-aware baselines. Three baselines that account for the frequency are added to Tables 1, 2, 10, and 11: RFScape (the best learned baseline) with the Friis term 20log⁡10(f′/f)20\log_{10}(f'/f), the COST-231 multi-wall model, and Sionna RT with ITU materials calibrated on the faces of FresGRT after the LoS stage and the same training links.

2. A new test band. A test band at 7.1 GHz, in the 6 GHz band of Wi-Fi 6E, is added to all 24 ray-tracing datasets (Tables 2, 11, 12, and 13).

3. Baseline details. Appendix D describes the implementation of every baseline, including how each baseline uses the vision data.

4. Additional evaluations. The revision evaluates the reflection order (Table 14), the LoS and NLoS links (Table 15), the interior materials (Table 16), the held-out positions (Table 12), the per-subcarrier power (Table 13), and the scene changes (Tables 5 and 17).

5. Practical cost. The revision describes how FresGRT detects a scene change and edits the Gaussians (Appendix F), reports the training time of every scene (Table 9), and compares FresGRT with on-demand measurement (Appendix G).

With the new baselines, FresGRT still outperforms the best baseline by 30% on the real sites and 41% on the ray-tracing datasets in cross-band prediction.

(Edited: the Sionna RT baseline was rerun on the faces of FresGRT after the LoS stage. Its numbers and description are updated to match the latest revision. Item 5 is added.)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.