{"id":"6951be3e-309f-4210-9286-b76f25b9a776","arxiv_id":"2508.06054","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MM-LSCM fuses RSRP and LiDAR point clouds in a NeRF-style volume renderer to predict angular power spectra in unexplored areas, outperforming WNOMP and a LiDAR-free ablation on a lab dataset.","lead":"This paper builds a neural network that combines radio signal strength readings with LiDAR scans to predict signal conditions across a whole coverage area. It reports large gains over prior channel-modeling baselines on a lab dataset, though the radio measurements used for evaluation are synthesized rather than collected.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No derivation connects volume-rendered integral (Eq. 12) to discrete APS (Eq. 1); Eq. (5) also misdefines APS as a complex expectation, so the forward model is not established.","rationale":"The reader's weakest assumption matches the main concern I find: the volume-rendered integral in Eq. (12) is the only link between the neural network's free outputs and the physical APS of Eq. (1), and the paper provides no derivation for this link. If that mapping is unfaithful, the network cannot recover the true APS even with unlimited data, and the empirical results in Table I become uninterpretable. The Eq. (5) error—defining APS as an expected complex gain rather than an expected power—strengthens the concern, because it makes the formal model internally inconsistent. I do not think this forces rejection: the experimental design is a reasonable proof-of-concept (real LiDAR and real measured spatial spectrum, with synthetic RSRP), and a corrected derivation or a synthetic ground-truth validation could settle the issue. Therefore the reader's CONDITIONAL verdict is appropriate; my analysis does not move it. I would ask the authors to correct Eq. (5) to E|α|^2 and to supply the missing derivation or the proposed three-path experiment.","tokens_in":9135,"tokens_out":10609,"duration_ms":135506,"concrete_test":"Re-derive Eq. (12) from Eq. (1) for a minimal three-scatterer channel: assign each path's α_{n,l,t} to a voxel along ray n, and check whether any choice of S_{l,n}(t) and σ_n(t) makes ∫νS dt equal α_{n,l,t} for arbitrary complex gains and delays. Then train MM-LSCM on noiseless y=Φx generated from this known sparse x for a synthetic scene with three paths; if the inferred \\hat x_l by Eq. (13) does not converge to the true sparse peaks when the RSRP residual is near zero, the volume-rendered forward model is unfaithful and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Sections III-C/D: APS estimates \\hat x_{l,n}=|∫_0^{t_max} ν_n(t)S_{l,n}(t)dt|^2 (Eqs. 12–13) are asserted to represent the sparse angle-only APS of Eq. (1). The complex gain α_{n,l,t} in Eq. (1) is a single scalar per discrete angle; the volume rendering instead coherently integrates over range t with stopping density ν_n(t)=T_n(0,t)σ_n(t). No derivation shows that this integral can equal any finite sum of path gains or their expected squared modulus. Since S_{l,n}(t) and σ_n(t) are free network outputs, minimizing the self-supervised RSRP loss (17) can drive \\hat x_l to any vector satisfying Φ\\hat x_l≈y_l; the reported MSE advantage over WNOMP may then reflect curve-fitting of the synthetic RSRP rather than physical APS reconstruction. The concern is compounded by Eq. (5), which defines the APS as x_l=E_T[α_{l,t}] (a complex vector) even though Eqs. (4) and (6) require x_l to be the real nonnegative power spectrum E_T|α_{n,l,t}|^2; the absolute square is missing from the formal model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MM-LSCM, a dual-branch neural network that combines beam-wise RSRP and LiDAR point clouds to predict angular power spectra (APS) over a coverage area. A NeRF-style volume rendering representation is used: each voxel is assigned a hitting probability and a direction-dependent complex radiance, and integration along AoD rays yields an APS estimate. Training is self-supervised through an RSRP reconstruction loss, with an auxiliary depth loss supervised by LiDAR occupancy. Experiments on the NeRF2 indoor dataset, using synthetic RSRP generated from measured spatial spectra, report lower MAE/MSE than WNOMP and a single-modality ablation.","tokens_in":9539,"tokens_out":8316,"duration_ms":98083,"significance":"The self-supervised formulation and the use of LiDAR as a geometric prior are genuinely attractive: they avoid labeled APS data and are aimed at predicting unexplored regions, which per-grid compressed-sensing baselines cannot do. The depth-supervision idea, tying the radio stopping probability to LiDAR occupancy, is a novel and reasonable soft-alignment mechanism. If the formal model and evaluation are made sound, the framework would be a useful contribution to data-driven localized channel modeling. However, the current manuscript does not yet establish the claimed reconstruction accuracy because several load-bearing definitions and mappings are unproven or inconsistent.","major_comments":[{"comment":"Equation (5) defines the APS as x_l = E_T[alpha_{l,t}] with alpha_{l,t} the complex path-gain vector. Under the paper's own uniform-phase assumption, E_T[alpha_{l,t}] = 0, and the quantity is complex, not a real nonnegative APS. Equations (4) and (6), and the loss (17), require x_l = E_T[|alpha_{n,l,t}|^2]. This is not a notational slip: as written, the formal model predicts zero RSRP. Please correct the definition and give the short derivation from Eq. (3) rather than citing [6] alone.","section":"Sec. II, Eq. (5)"},{"comment":"The mapping from the volume-rendered integral hat_r_{l,n} = integral nu_n(t) S_{l,n}(t) dt to the discrete sparse APS in Eq. (1) is asserted, not derived. Since sigma_n(t) and S_{l,n}(t) are unconstrained network outputs, minimizing (17) can fit Phi hat_x_l to y_l without recovering the true APS. Please provide a representational argument, or clearly state a modeling assumption under which the integral equals the desired path sum, or add a synthetic experiment with known sparse APS to demonstrate that hat_x_l is the underlying APS rather than a curve fit.","section":"Sec. III-D, Eqs. (12)-(13)"},{"comment":"The RSRP inputs are not measured. They are generated as y_l = Phi x_l and y_l^r = Phi_r x_l from the downsampled ground-truth spatial spectrum. Thus the abstract's 'real-world dataset' claim applies only to LiDAR and spatial spectrum; all radio training targets are synthetic and noiseless. Moreover, MSE in Eq. (23) is evaluated against exactly these x_l, so the comparison rewards models that invert the same generative model used to create the data. Please present real RSRP measurements, or explicitly reframe the study as a controlled synthetic benchmark and discuss the conditions under which the conclusions transfer.","section":"Sec. IV-A, Eqs. (21)-(22)"},{"comment":"Using the first nonzero-density voxel along each ray as the depth supervision target for the radio stopping probability is a strong physical assumption. In a multipath environment, the dominant contribution to a given AoD can arrive from a reflection beyond the first occupied voxel; forcing nu_n(t) to concentrate at t_{d*} will suppress legitimate later paths and bias hat_x_l. The paper provides no evidence that first-hit depth matches the radio-path depth. Please justify this assumption or relax it, for example by supervising only LoS rays.","section":"Sec. III-E, Eqs. (18)-(19)"},{"comment":"The robustness experiment is not fully specified. 'Introduce 3 dB noise to both the measurement matrix and the RSRP data' could mean noise is added to the Phi used in training, to the Phi used at test time, or to the RSRP only. If the model is trained with clean Phi but evaluated with noisy Phi, the degradation may be an artifact of train/test mismatch; if noisy Phi is used in the loss (17), the forward model itself changes. State precisely how noise is added, at which stage, and how the WNOMP baseline is treated under the same perturbation.","section":"Sec. IV-C"}],"minor_comments":[{"comment":"The APS dimension is inconsistent: 'N=16201' and 'N=1620' appear in the same paragraph, while the footnote states 90 x 18 = 1620.","section":"Sec. IV-A"},{"comment":"There is an unmatched parenthesis: '|W^H S|^2)' should presumably be '(|W^H S|^2)'.","section":"Sec. II, Eq. (4)"},{"comment":"The spherical-harmonics notation sh(tau, mu) is not defined. Please specify how tau encodes the SH coefficients and how mu is evaluated with the basis.","section":"Sec. III-C, Eq. (8)"},{"comment":"The MAE formula appears to be a sum over M beams, not a mean. Clarify whether the reported numbers are averaged over the M beams.","section":"Sec. IV-B, Eq. (23)"},{"comment":"MSE values differ by orders of magnitude across methods (e.g., 0.0685 vs 153.93). State the normalization/scaling of x_l so the reader can interpret the comparison.","section":"Tables I and II"},{"comment":"Clarify that the receiver in the NeRF2 dataset is treated as the BS while the original setup uses a fixed receiver and a moving tag; the coordinate transformations and codebook orientation used in the experiments should be specified.","section":"Sec. IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising direction and the authors are clearly aware of the relevant literature. I chose major_revision rather than reject because the central issues, while load-bearing, are in principle addressable: correcting Eq. (5), adding a representational or empirical justification for the volume-rendering APS mapping, and reworking the evaluation so that the RSRP data are not synthesized from the same x_l used for MSE evaluation. If the authors cannot provide real RSRP or a convincing controlled benchmark, the empirical claims should be substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Haibo, this one is worth a look, with caveats. The core idea — using NeRF-style volume rendering to turn sparse RSRP plus LiDAR into a full angular power spectrum, and to extrapolate to unexplored regions — is a real and potentially useful step for radio map estimation. The architecture is a sensible adaptation of NeRF2: a dual-branch network where LiDAR density conditions the attenuation branch, and a spherical-harmonics radiance branch produces complex signals. The stopping-probability coherence weight that links radio rendering and depth prediction is a nice touch, and the self-supervised loss over RSRP is a natural way to avoid labeled APS.\n\nWhat the paper actually does well is give a concrete trainable model and show, on the NeRF2 dataset, that it beats the two implemented comparisons (WNOMP and a LiDAR-free ablation) on both explored and unexplored grids. The gap on unexplored regions is the most interesting result.\n\nNow the soft spots. They are not cosmetic. Eq. (5) defines the APS as a complex expectation of the channel gains, which is inconsistent with Eq. (4) and (6) where it must be the power spectrum E|α|^2. That looks like a dropped absolute square, but it is in the formal model and needs fixing. More importantly, the volume-rendered integral in Eqs. (12)–(13) is asserted as the APS without any derivation connecting it to the discrete path model in Eq. (1). The claim that it follows from Huygens-Fresnel is not a derivation, and it matters because the entire predictive power of the model rests on this representation being faithful. The experiments are also synthetic: RSRP is generated as y=Φx from the ground-truth APS, not measured. That is acceptable for a proof of concept, but the robustness test then adds noise to this synthetic data and the already-known matrix, which is weaker evidence than real-world noisy RSRP. Finally, the paper gives no code, no hyperparameters, no error bars, and no comparison to simpler interpolation or Gaussian-process baselines. The WNOMP comparison is only fair on Sub-Task 1 because WNOMP is per-grid, but a stronger extrapolation baseline would make the claim more convincing.\n\nAll that said, the central hypothesis — that geometry from LiDAR improves channel extrapolation — is plausible, and the paper is honest about its limitations. It is not a finished result, but it deserves a serious referee. An editor should send it out with the expectation of major revision: clean up Eq. (5), add a real derivation or clear argument for the volume-to-APS mapping, release code and settings, and add error bars. If the forward model can be justified, this would be a solid journal paper. For now, read it to track the ideas, but don't cite it yet; wait for the revision.","headline":"A promising NeRF-based channel modeling idea whose forward model is not derived and whose experiments are synthetic; deserves a serious referee but needs major revision.","tokens_in":9960,"tokens_out":4148,"would_cite":false,"duration_ms":51284,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A self-supervised volume-rendering network fuses LiDAR and RSRP to reconstruct channel angular power spectra across a coverage area, including unmeasured locations, and outperforms RSRP-only compressed sensing.","keywords":["localized statistical channel modeling","angular power spectrum","neural radiance field","LiDAR point cloud","RSRP","self-supervised learning","volume rendering","spherical harmonics"],"falsifier":"Take a synthetic scene whose true APS is a few known point sources generated from Eq. (1), train MM-LSCM on RSRP produced by Eq. (21), and test whether the recovered APS localizes those source angles and gains with small error. If the model matches the RSRP but assigns energy to wrong angles, or if its APS estimate shifts when voxel resolution or ray discretization changes, the volume-rendering representation is not faithful to the channel model.","tokens_in":9062,"feed_emoji":"📡","tokens_out":10065,"duration_ms":100736,"temperature":0.7,"pith_summary":"MM-LSCM claims that fusing LiDAR point clouds with beam-wise RSRP measurements in a self-supervised neural radiance field reconstructs the channel angular power spectrum (APS) across an entire coverage area, including locations never measured. Existing LSCM treats each grid as a standalone compressed-sensing problem, so unvisited grids cannot be predicted and environmental structure is ignored. The paper's network renders the APS by volume integration: each voxel emits a learned direction-dependent complex signal, and an environment-informed stopping probability weights which voxels actually contribute. Compared with the RSRP-only baseline, the method cuts MAE from 1.85 dB to 0.87 dB on measured grids and reaches 0.56 to 0.90 dB on unmeasured grids, and it tolerates 3 dB noise on both measurement matrix and RSRP. If this holds, operators could replace expensive drive tests with sparse RSRP plus LiDAR sweeps.","feed_headline":"LiDAR + RSRP predicts radio channels in unvisited areas","feed_subtitle":"A self-supervised neural radiance field beats RSRP-only solvers and holds up under 3 dB noise.","key_machinery":"The load-bearing object is the stopping probability $\\nu_n(t) = T_n(0,t)\\,\\sigma_n(t)$, where $\\sigma_n(t)$ is the network-predicted hit probability at a voxel and $T_n(0,t)=\\exp(-\\int_0^t \\sigma_n(t')\\,dt')$ is the pass probability. This one weight aligns the radio and environment modalities: it controls how much each voxel's radiated signal contributes to the synthesized APS, and it also renders the obstacle depth used as a geometric self-supervision signal. The APS element for direction $n$ is formed as $|\\int_0^{t_{\\max}}\\nu_n(t)\\,S_{l,n}(t)\\,dt|^2$, with $S_{l,n}$ the learned spherical-harmonics radiance from the voxel toward grid $l$; the whole pipeline is a ray-marching approximation","core_discovery":"The central claim is that environmental geometry, injected as per-voxel density from a LiDAR point cloud, supplies the missing spatial prior for localized statistical channel modeling. The proposed network learns two coupled fields along rays from the base station: a hitting probability (volume density) that encodes where the signal is blocked or scattered, and a direction-dependent complex radiance represented by spherical harmonics. Volume rendering combines these into a complex gain for each discrete angle, whose squared magnitude is the APS; the same rendering, weighted by a stopping probability, also produces an expected obstacle depth. Training is self-supervised: on measured grids the","pith_inferences":["Since the performance gain comes from injecting geometry as a prior, the MAE gap over RSRP-only solvers should widen in cluttered environments and nearly vanish in open areas; this is testable by re-running the same pipeline on synthetic scenes with controlled scatterer density.","The stopping probability is a single shared weight for radio contribution and physical depth; this suggests using predicted depth in unexplored regions as an audit signal to check whether the model's radio predictions are grounded in real geometry rather than fitted to the training RSRP.","The architecture only needs a per-voxel density prior, so the same self-supervised scheme could ingest other sensors, such as cameras, coarse ray-traced maps, or building footprints, without changing the rendering equations.","Because the learned field is continuous, one could simulate antenna rotations or relocations by re-evaluating the measurement matrix against the same radiance field, enabling rapid what-if network planning without new measurements."],"forward_implications":["Operators can estimate full angular power spectra across a coverage area from sparse drive-test RSRP plus LiDAR, without labeled channel data.","Unvisited grids become predictable, so network optimization can simulate changes in antenna tilt and azimuth in areas never measured.","The method keeps its accuracy when both the codebook response and the RSRP measurements are noisy, which matches real-world conditions.","Removing the environment branch, as in the single-modality ablation, raises MAE and increases MSE by orders of magnitude on the measured-grid task, so the LiDAR prior is doing real work.","The depth-rendering side output means the model produces a coarse environmental map as a byproduct, potentially reusable for site planning."],"supporting_citations":[{"why":"Supplies the LSCM measurement model $y_l=\\Phi x_l$ and the WNOMP baseline that MM-LSCM compares against.","marker":"[6]"},{"why":"Supplies the real-world laboratory dataset (LiDAR point clouds and angular spectra) and the single-modality neural baseline adapted as SM-LSCM.","marker":"[10]"},{"why":"Supplies the volume-rendering pass-probability and stopping-probability equations used in synthesis.","marker":"[11]"},{"why":"Supports the claim that point-cloud geometry captures propagation semantics useful for prediction.","marker":"[12]"},{"why":"Supplies the discretized multipath channel model and steering-vector form behind Eq. (1).","marker":"[14]"},{"why":"Supplies the Fourier position encoding $\\gamma(\\cdot)$ used to improve spatial resolution.","marker":"[15]"},{"why":"Supplies the spherical-harmonics representation used to model direction-dependent radiance.","marker":"[16]"},{"why":"Supplies the discrete volume-rendering weights used to approximate the continuous integrals in Eqs. (12) and (14).","marker":"[17]"}],"fun_headline_variants":["Self-supervised neural field fuses LiDAR and RSRP for channel stats","LiDAR-neural radiance field predicts radio channels","Multi-modal neural field for channel modeling without labels","Self-supervised LiDAR+RSRP predicts 3D radio coverage","Neural radiance field for wireless channel maps from LiDAR+RSRP"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes the volume-rendered integral of per-voxel radiance times stopping probability faithfully represents the discrete, sparse multipath angular power spectrum of Eq. (1); no derivation links the two, so if the mapping is unfaithful even unlimited training data cannot recover the true APS.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised neural field fuses LiDAR and RSRP for channel stats","LiDAR-neural radiance field predicts radio channels","Multi-modal neural field for channel modeling without labels","Self-supervised LiDAR+RSRP predicts 3D radio coverage","Neural radiance field for wireless channel maps from LiDAR+RSRP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000424,"raw_usage":{"total_tokens":1965,"prompt_tokens":649,"completion_tokens":1316,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":393,"completion_tokens_details":{"reasoning_tokens":1226}},"tokens_in":393,"tokens_out":1316,"duration_ms":11458,"temperature":1.0,"reasoning_tokens":1226,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:58:37.547777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic scene whose true APS is a few known point sources generated from Eq. (1), train MM-LSCM on RSRP produced by Eq. (21), and test whether the recovered APS localizes those source angles and gains with small error. If the model matches the RSRP but assigns energy to wrong angles, or if its APS estimate shifts when voxel resolution or ray discretization changes, the volume-rendering representation is not faithful to the channel model.","supporting_citations":[{"cited_title":"A physics-based and data-driven approach for localized statistical channel modeling,","cited_arxiv_id":null,"evidence_quote":"Supplies the LSCM measurement model $y_l=\\Phi x_l$ and the WNOMP baseline that MM-LSCM compares against."},{"cited_title":"Nerf2: Neural radio-frequency radiance fields,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world laboratory dataset (LiDAR point clouds and angular spectra) and the single-modality neural baseline adapted as SM-LSCM."},{"cited_title":"Nerf: Representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the volume-rendering pass-probability and stopping-probability equations used in synthesis."},{"cited_title":"How to define the propagation environment semantics and its application in scatterer-based beam prediction,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that point-cloud geometry captures propagation semantics useful for prediction."},{"cited_title":"Fourier features let networks learn high frequency functions in low dimensional domains,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier position encoding $\\gamma(\\cdot)$ used to improve spatial resolution."},{"cited_title":"Plenoxels: Radiance fields without neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the spherical-harmonics representation used to model direction-dependent radiance."}],"review_version":1}