acceptodds
Withdrawn from ICLR 2027

RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

31 views

Abstract

Wireless propagation modeling is foundational to the evaluation and design of modern communication systems. To achieve accurate wireless propagation prediction, recent learning-based approaches leverage scene geometry and materials, demonstrating promising generalization across transceiver locations and unseen scenes. However, these approaches remain largely confined to a fixed configuration of array size, receiver orientation, and transmit beam. In this work, we introduce RFWeave, a learned wireless propagation model designed to generalize across scenes, transceiver locations, and transceiver configurations. To accommodate varying array structures, RFWeave incorporates variable-length transceiver tokens that represent antenna elements individually. Furthermore, to capture the geometry between transmit and receive antenna elements, RFWeave aggregates their pairwise spatial relations into propagation tokens. A configuration-conditioned decoder predicts the spatial spectrum and received power. RFWeave is trained on a large-scale simulated dataset spanning 10,000 indoor scenes and more than 5 million channel records under diverse transceiver configurations. Extensive evaluations show that RFWeave achieves state-of-the-art performance across all evaluated generalization settings, improving spatial spectrum structural similarity (SSIM) by up to 15.7% and reducing received power mean absolute error (MAE) by up to 35.4% over the strongest baselines on unseen scenes. Additional experiments on measured DICHASUS data and in beam management further demonstrate RFWeave's broader applicability to real-world channels and downstream wireless tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
wireless propagation, 6G networks, transformer, spatial spectrum

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vx4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RFWeave is a learned wireless propagation model. Given a material-annotated scene, TX/RX locations, and a configuration (array size, RX orientation, TX beam), it predicts the RX-side spatial covariance, and hence the AoA spatial spectrum, together with the received power. Pairwise TX–RX element geometry is aggregated into element-indexed propagation tokens. A scene-conditioned transformer backbone refines these tokens, and RX orientation and TX beam enter only at the prediction stage, which outputs a low-rank covariance factor. The training data (10K indoor scenes, more than 5M records) are derived from the path records of two existing ray-traced datasets (WiPTR and Wi3R, both at 3.5 GHz). Far-field synthesis produces targets for 6 array sizes, 8 RX orientations, and 16 TX beams. The paper reports up to +15.7% spatial-spectrum SSIM and −35.4% received-power MAE over the strongest baselines on unseen scenes.

Critical strengths
  1. A practically relevant and clearly posed problem. The paper targets joint generalization across scenes, transceiver locations, and transceiver configurations, and the variable-length element tokenization is a natural way to handle different array structures.

  2. A broad experimental study. It covers a large dataset, single- and joint-factor configuration shifts, data and model scaling, ablations, a real-data adaptation study, and a downstream task. The paper is well written, and code is released.

Critical weaknesses

W1. The paper does not show that a learned model is needed, and in particular that the configuration mapping needs to be learned.

  • The targets are produced by deterministic far-field synthesis from ray-traced path records, so configuration changes are generated "without rerunning the ray tracer". Given a configuration-agnostic angular representation (per-path angles and gains, or a double-directional angular power spectrum), any array size, orientation, or beam follows in closed form. Learning this mapping end to end is unnecessary, and it is exactly where RFWeave is weakest: SSIM 0.566 under array-size extrapolation (Table 3), 6.65 dB power MAE on unseen beams (Table 2), and a sharp rise in power error beyond the training array sizes (Fig. 6). Known scaling, such as array gain, is not captured. The paper does not compare against a hybrid that learns only the scene/location-to-angular part and applies analytic synthesis.

  • The simulated ground truth is itself produced by ray tracing, so a learned model has to be justified by speed or by sim-to-real accuracy. Yet ray-tracing cost appears only as motivation; there is no runtime or accuracy comparison against ray tracing plus far-field synthesis. The real-data evidence is limited.

  • For received power, recent learned radio-map generators are natural baselines. The paper cites one (RadioDiff) as related work but does not compare against it.

W2. The headline gains come from seen configurations/environments. On unseen configurations the gains are small and absolute accuracy is low.

  • The abstract's "+15.7% SSIM / −35.4% power MAE" corresponds to the seen-configuration block of Table 3 (0.8060 vs. 0.6968; 2.28 vs. 3.53).
  • Under unseen beam directions (angular interpolation), the gains shrink to +7.5% SSIM and −5.2% power MAE.
  • In the single-factor study (Table 2), power MAE on unseen array sizes is 5.40 vs. 5.53, and on unseen beams 6.65 vs. 6.88.
  • Absolute errors remain large. On unseen beams, power MAE is 6.65–7.93 dB across all methods, and cross-scene power MAE on Wi3R is 8.30–11.46 dB without target-scene fine-tuning.
  • The above results give me a concern on whether this model works well on the unseen environments (e.g., different room w/ various other objects/materials).

W3. The training and test environments lack realism and diversity, and complex environments are not evaluated.

  • All scenes come from two existing ray-traced indoor datasets at a single frequency (3.5 GHz). Judging from Figs. 1–2 and the five material classes shown (drywall, wood, glass, concrete, floor), the scenes appear to be empty floor plans with walls, doors, and windows; the paper never mentions furniture or clutter. Please clarify the scene content.
  • Furnished indoor 3D datasets are widely available (e.g., 3D-FRONT, Structured3D, Matterport3D, HM3D, ScanNet, Infinigen Indoors). Furniture strongly affects multipath richness and therefore the AoA spatial spectrum that this paper targets.
  • Performance saturates beyond about 2,000 training rooms (Fig. 5), which suggests that diversity rather than quantity is the bottleneck.
  • The only structurally more complex environment, the simulated factory (with metal, chipboard, and marble), is used solely for downstream beam selection, after fine-tuning with an unspecified number of samples. No propagation accuracy (SSIM or power MAE) is reported for it.
  • No results are broken down by LoS/NLoS or by multipath richness, and all qualitative examples show floor-plan scenes. The paper therefore does not show how well RFWeave predicts propagation in cluttered or furnished environments.

W4. Minor. Statistical reporting and scope. No seeds or variance are reported, although many differences are small (e.g., 0.6357 vs. 0.6459 SSIM between the two single-factor variants). The dataset is a post-processing of existing public ray-tracing records.

Minor.

  • The "w/o configuration conditioning" ablation removes a necessary input, so its degradation is expected.A comparison with alternative conditioning schemes (e.g., injecting the configuration at the input) would be more informative.
  • The pairwise-geometry (−3.8% SSIM) and material (−1.6%) contributions are modest.
Questions for authors
  1. How does RFWeave compare with a hybrid that learns a configuration-agnostic angular representation and applies analytic far-field synthesis, especially for unseen array sizes and beams? What are the runtime and accuracy relative to ray tracing plus synthesis?

  2. For received power, how does RFWeave compare with recent learned radio-map generators (e.g., RadioDiff)?

  3. What do the WiPTR and Wi3R scenes contain (e.g., furniture)? Can you evaluate on furnished or real-scanned environments, report propagation accuracy in the factory scene, and break results down by LoS/NLoS?

  4. Exactly how do the baselines receive element positions, orientations, and beam weights?

  5. Please provide the full DICHASUS setup (frequency, array, scene model, covariance computation, split) and zero-shot results.

  6. For beam management, how does RFWeave compare with standard baselines? How many adaptation samples were used, what transmit power was assumed, and how many test positions were evaluated?

Reviewer_3datOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RFWeave for wireless propagation prediction across different scenes, transceiver locations, and antenna configurations. It encodes scene geometry and materials, represents individual antenna elements as tokens, and summarizes pairwise transmitter-receiver geometry. A configuration-conditioned decoder predicts received power and spatial covariance, which is used to compute the AoA spatial spectrum. Experiments on simulated channels, measured-data adaptation, and beam management report improvements over existing methods.

Critical strengths
  1. Existing cross-scene propagation models usually assume fixed transceiver configurations. RFWeave supports different array sizes, receiver orientations, and transmit beams within a single model.

  2. The model represents antenna elements individually and predicts spatial covariance through a low-rank factor, allowing the output to adapt to different receiver array sizes. RFWeave outperforms several baselines in simulation, and the ablations show that pairwise geometry and configuration conditioning contribute to its performance.

Critical weaknesses
  1. The joint generalization evaluation is incomplete. In the angular-interpolation setting, transmit beams are unseen, but array sizes and receiver orientations are seen. In the array-size extrapolation setting, only array sizes are outside the training range. The paper does not test all three factors being unseen together, so it is unclear how well the model handles simultaneous extrapolation. The measured-channel experiment also uses fine-tuning in a single hallway, leaving real-world zero-shot generalization untested.

  2. The baseline comparison does not clearly establish the benefit of the proposed antenna-token modeling. The training targets are synthesized from multipath records using known array-response equations. An alternative would be to learn configuration-independent propagation information and then apply these equations to different arrays and beams. Without this comparison, it is difficult to determine whether pairwise antenna tokens are necessary for the reported gains.

Questions for authors
  1. Have you tested unseen array sizes, receiver orientations, and transmit beams together on held-out scenes? If not, how should the joint generalization claim be interpreted?

  2. Is there a comparison with a model that learns configuration-independent propagation information and applies standard array-response equations at inference? What evidence isolates the benefit of pairwise antenna tokens?

  3. How does RFWeave perform on the measured DICHASUS data before fine-tuning?

Reviewer_Tb2EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RFWeave for wireless propagation prediction across scenes, transceiver locations, and configurations. The model represents individual antenna elements as variable-length tokens and aggregates pairwise TX–RX geometry into propagation tokens. A configuration-conditioned decoder predicts received power and a low-rank spatial covariance, from which the spatial spectrum is computed. Experiments on WiPTR and Wi3R evaluate scene and configuration generalization. The paper also reports adaptation to DICHASUS measurements and a downstream beam management experiment.

Critical strengths
  1. The paper studies a useful problem. Supporting different antenna sizes, receiver orientations, and transmit beams with one model can reduce the need to train separate propagation models. The element-level representation is a reasonable design for this purpose.

  2. The improvements on unseen scenes with seen configurations are substantial. In Table 3, RFWeave-B improves SSIM from 0.6968 to 0.8060 and reduces power MAE from 3.53 to 2.28 dB compared with the best respective baselines. The architectural ablations also support the usefulness of pairwise geometry and configuration conditioning.

Critical weaknesses
  1. The evidence for joint configuration generalization is incomplete. In Table 9, angular interpolation only uses unseen transmit beams, while array extrapolation only uses unseen array sizes. Neither setting tests unseen sizes, orientations, and beams together. Moreover, under array extrapolation, the power MAE improvement over Wi-GATr is only 4.93 to 4.79 dB, without uncertainty estimates. These results leave the robustness of the main generalization contribution unclear.

  2. The comparisons do not sufficiently establish the advantage over closely related approaches. Point2Radio is cited but not evaluated, even in the fixed-configuration cross-scene setting. The beam management experiment only compares against location-based steering and an oracle, without applying the same evaluation to other learned predictors. Therefore, the reported improvements do not fully establish the benefit of RFWeave over relevant alternatives. These evaluation gaps lead me to lean toward weak rejection.

Questions for authors
  1. Does any reported experiment simultaneously use unseen array sizes, receiver orientations, and transmit beams? Table 9 appears to hold out only one configuration factor in each OOD regime. Please clarify the scope of the joint generalization claim.

  2. Why is Point2Radio omitted from the fixed-configuration comparison? Are results available for this baseline, or for other learned predictors under the same beam management protocol?

  3. For DICHASUS, how are training and test positions separated, and how are covariance and spectrum targets constructed from measured channels? These details are needed to assess the reported adaptation with only 100 samples.

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

Related papers

Discussion (0)

Sign in to comment.