Differentiable High-Performance Ray Tracing-Based Simulation of Radio Propagation with Point Clouds

Janne Heikkil\"a; Miguel Bordallo L\'opez; Niklas Vaara; Pekka Sangi

REVIEW 4 major objections 5 minor 26 references

Differentiable High-Performance Ray Tracing-Based Simulation of Radio Propagation with Point Clouds

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A differentiable ray tracer simulates multi-bounce radio paths directly on point clouds in under 90 milliseconds and learns material parameters by gradient descent.

desk verdict A fast, differentiable point-cloud ray tracer with real performance gains, but the path-completeness claim is unquantified and the material learning is only self-referential. read the letter →

arxiv 2507.04021 v1 pith:4KG2K25S submitted 2025-07-05 eess.SP cs.CV

Niklas Vaara , Pekka Sangi , Miguel Bordallo López , Janne Heikkilä This is my paper · ORCID

classification eess.SPcs.CV

keywords raytracingradiopropagationsimulationpointcloudsdifferentiablemulti-bouncepathsspecularreflectiondiffusescatteringmaterialparameterlearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a ray tracing simulator for radio propagation that works directly on 3D point clouds, skipping the usual conversion to triangle meshes. The authors claim it is the first point-cloud ray tracing method that computes multi-bounce paths—specular reflections, diffuse scattering, and single-bounce diffraction—fast enough for real-time use, with two indoor scenarios completing in under 90 ms. The simulator is differentiable, meaning the electromagnetic parameters assigned to each material label can be adjusted by gradient descent to match measured channel impulse responses. If true, this would let an automated pipeline reconstruct an environment from sensor data, label it semantically, and calibrate its radio properties from measurements without manual tuning.

What carries the argument

The central machinery is the visibility matrix-aided path computation: a matrix of visibilities from each receiver to each voxel's ray reception point is built in milliseconds, and after a constant number of specular ray casts from the transmitter, the matrix is queried to determine which interaction points can see the receiver, yielding multi-bounce path candidates without increasing ray count. The second piece is the intersection model: each point in a voxel is represented as a planar disk of radius $r_p$ aligned with its surface normal, and the ray-surface intersection is the weighted average of all hit disks, with weights $\omega_i = \exp(-\|q_i - p_i\|^2 / 2 r_p^2) \exp(-\lambda_d \|q_i - q_{\min}\|)$ favoring hits near the disk center and close to the ray origin. This replaces iterative signed-distance-function ray marching with explicit, differentiable ray-disk intersections.

What would settle it

Simulate a scene with a known strong specular reflector that is not aligned with any voxel reception point, then compare the simulator's path list against an exhaustive image-method ray tracer or against measured channel impulse responses: if the reflector's path is missing while geometry and materials are known, the constant-count casting plus receiver-only visibility matrix has missed a significant path. Conversely, if learned material parameters on a scene with known ground-truth materials fail to converge to those values while the path candidates are unchanged, the disk-weighted intersection model is biasing the electromagnetic coefficients.

Watch

Extended reading notes

Core claim

The paper's central claim is that propagation paths can be simulated directly on point clouds in real time by combining a voxel-based discretization with a visibility matrix computed only from each receiver to voxel ray reception points. Instead of letting the number of launched rays grow after each bounce, every interaction is treated as a specular reflection, and candidate paths are recovered by querying the visibility matrix; diffuse scattering is handled as a final interaction and validated by a single receiver ray. Intersections with the point cloud are computed by representing each point as a surface-aligned planar disk and taking a Gaussian-weighted average of ray-disk hits, replacing the slower signed-distance-function ray marching of the authors' earlier method. The resulting paths are passed to Sionna for differentiable electromagnetic computation, and the paper demonstrates that relative permittivity, conductivity, and scattering coefficients of material labels can be recovered from synthetic ground-truth channel impulse responses in about 5000 Adam iterations.

Load-bearing premise

The load-bearing premise is that a visibility matrix computed only from each receiver to voxel reception points, plus a constant number of specular ray casts from the transmitter, finds every propagation path that matters, and that the weighted disk average with hand-set radius and depth attenuation represents the true reflecting surface.

Editorial extensions

If this is right

Real-time applications such as robot odometry links can recompute propagation paths on every movement, as in the motivating example.
Environment model construction becomes fully automated: raw sensor point clouds plus semantic labels feed the simulator without manual mesh cleanup.
Material parameters can be learned from channel measurements per material label, so radio digital twins can be calibrated automatically rather than relying on sparse literature databases.
Because the electromagnetic computation is differentiable and path computation is hard-coded, the learning loop can optimize material properties without re-running the path finder every iteration.
The method also supports triangle meshes, so it can be used wherever current mesh-based ray tracers run, with point clouds as the preferred input.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the visibility matrix truly captures all significant paths with a constant ray count, the same approach might extend to dynamic scenes where the environment changes frame to frame, since only the matrix and intersection weights need recomputation.
The hand-set disk radius $r_p$ and depth attenuation $\lambda_d$ are likely scene-dependent; a testable extension would be to make them learnable parameters in the same gradient-descent loop, since the intersection is differentiable.
The 90 ms figure is for a single transmitter with up to six receivers in a room-scale scene; scaling to city-scale point clouds or many simultaneous links may require hierarchical voxel structures or multi-GPU partitioning.
Since the method differentiates through Sionna's electromagnetic model only, not through the path computation, future work could differentiate through the geometry itself to optimize sensor placement or scene reconstruction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Niklas, you should know two things about 2507.04021. First, it actually looks like a working, fast differentiable point-cloud ray tracer: the authors ship code, the 81–88 ms Replica numbers are plausible, and the RX-only visibility matrix plus constant-ray multi-bounce is a genuinely new combination. Second, the central claim of multi-bounce completeness is not supported by the experiments. The stress-test note is right: with one ray per DPS and no branching, the search space is linear in the number of DPSs and bounce depth, not combinatorial, so the method can silently miss significant paths. The paper never reports recall against an image-method or mesh-based reference, so we can't tell whether it finds the important paths or just some paths.

What the paper does well: the disk-based intersection (Eqs. 3–4) is a sensible improvement over the earlier SDF ray marching, and the two-stage refinement with convergence thresholds is a practical engineering choice. The Sionna integration is clean, and the whole pipeline is well-documented with parameters. The performance table shows clear gains over their prior work [7], and the claim of near-linear complexity with bounce depth is consistent with the algorithm described.

The soft spots are real but addressable. The baseline comparison is muddled: they cite [7]'s voxel size of 0.5 m as 'matching' their 0.0625 m, which it does not; that inflates the apparent advantage. The material-learning experiment is self-referential: ground-truth CIRs come from the same path engine, so it is an identifiability test, not a validation against reality. That is fine as a demo, but the abstract's phrasing suggests more. Also, there are no error bars on the timing numbers, and the path counts vary widely with max depth, which hints at sensitivity to the DPS grid.

The biggest unresolved issue is the one the stress-test highlights: no recall analysis. For an RT simulator, finding the right paths matters more than raw speed. I would like to see a comparison against a conventional mesh-based ray tracer (e.g., Sionna RT) on the same scenes, with a measure of how many significant paths (say, above a power threshold) are found. That would either confirm the method's usefulness or expose its blind spots.

My take: the paper is a serious engineering contribution with code and a plausible design. It deserves a proper peer review, but it needs at least one additional validation experiment—ideally a path-recall comparison—before it can support the 'physically accurate multipath' claim. I would not cite it for the material-learning result, but I might cite it as a differentiable point-cloud RT baseline. Bring it to the reading group if you want to see what a well-executed but under-validated systems paper looks like.

Referee Report

4 major / 5 minor

Summary. The paper proposes a differentiable ray-tracing-based radio propagation simulator that operates directly on point clouds. The method voxelizes the point cloud, builds a visibility matrix from each RX to the voxel ray reception points, casts a constant number of rays from each TX, and computes multi-bounce paths consisting of specular reflections, diffuse scattering, and single-bounce diffractions. The paths are passed to Sionna for differentiable electromagnetic computations. Experiments report sub-90 ms path computation for up to five interactions in two indoor scenes, and a demonstration of learning material parameters from synthetically generated channel impulse responses.

Significance. If the efficiency and completeness claims hold, the simulator would be a useful building block for real-time radio digital twins and differentiable material estimation. The paper has concrete strengths: the code is made available, the method integrates with Sionna, and it directly consumes point clouds without mesh conversion. The reported wall-clock times are encouraging. However, the experimental support is incomplete: the baseline comparison mixes voxel sizes, the multi-bounce path search is not evaluated for recall against an exhaustive reference, and the material-learning demonstration is a closed-loop synthetic identifiability test. These issues must be resolved before the central claims can be accepted.

major comments (4)

[III-A, Table II] The performance comparison against [7] is not apples-to-apples. The text states that the baseline uses a voxel size of 0.5 m 'since the smallest discretization size matches our voxel size,' but Table I sets our voxel size to rho = 0.0625 m; these values differ by a factor of 8 in linear dimension and about 512 in voxel volume. Trace time, refinement time, and the set of discovered paths all depend on voxel size, so the speed advantage shown in Table II may be partly an artifact of the coarser baseline discretization. Please rerun [7] at rho = 0.0625 m or otherwise provide a controlled comparison, and correct the contradictory sentence.
[II-B] The paper does not demonstrate that the constant-ray, no-branching path search finds all significant multi-bounce paths. Rays are launched from the TX to each DPS and then reflected; because the number of traced rays does not increase, a two-bounce path via surfaces A and B is found only if the ray from A happens to hit a B voxel that is RX-visible. The refinement in Eq. (1) optimizes points on already selected infinite planes and cannot recover a path whose intermediate reflection missed every DPS. Table II reports path counts and wall-clock times but never compares against an exhaustive image-method or mesh-based reference, so the reader cannot tell whether missing paths bias the 'physically accurate multipath components' claimed in the abstract. Please add a recall/completeness evaluation or explicitly document the coverage limitation.
[III-B] The material-learning experiment is self-referential. The ground-truth CIRs are generated by the same path finder and EM model whose parameters are then recovered by gradient descent; the experiment therefore tests identifiability within the simulator, not the ability to learn true environmental material properties. The sentence 'the parameters can be recovered accurately' should be restricted to this closed-loop setting, and the paper should either add an external validation (e.g., an independent ray tracer or measured channel data) or clearly frame the result as a differentiable-optimization demonstration.
[II-C, Eqs. (3)-(4)] The weighted-disk intersection model has two hand-set parameters, rp = 0.015 m and lambda_d = 100, with no sensitivity analysis or comparison to a mesh-based ground truth. Because the weighted average q directly determines path geometry and hence all downstream EM coefficients, a systematic bias in q would invalidate the simulated CIRs even if material parameters are perfectly fitted. Please quantify how the PDP or path geometry changes as rp and lambda_d vary over plausible ranges, and justify the chosen values from point-cloud density.

minor comments (5)

[Table I] The header 'V oxel Size' contains a stray space and should read 'Voxel Size'.
[II-A] The phrase 'An high level illustration' should be 'A high-level illustration'.
[Introduction] The novelty claim 'first ray tracing-based method for simulating propagation paths directly on point clouds that is capable of computing multi-bounce paths in time' should be supported by a direct comparison with prior point-cloud ray-launching methods such as [5] and [20], or softened to reflect the specific real-time setting.
[Table II] The number of specular-plus-scattered paths is not monotonic in depth (393, 681, 637, 403). Please explain whether this is due to duplicate removal, varying visibility, or stochastic effects.
[Fig. 3] The axes and legends in Fig. 3 are difficult to read in the provided rendering; please enlarge the font and include units for conductivity.

Circularity Check

1 steps flagged · score 4.0 of 10

Core ray-tracing derivation is self-contained; material-learning evidence is a self-referential identifiability test, not external validation.

fitted input called prediction [Section III-B (Learning Electromagnetic Properties from Material Labels), Eq. (5), Fig. 3]
"As the ground truth for training, channel impulse responses (CIRs) are synthetically generated by computing channel frequency responses (CFRs) using 129 equally spaced samples over a 1.5 GHz bandwidth centered at 8 GHz. ... The results are illustrated in Fig. 3, which showcases the parameters at each training iteration, and that the parameters can be recovered accurately."

The ground-truth CIRs are not measurements but outputs of the same differentiable simulator (same path computation plus Sionna EM) whose material parameters are then optimized to match them. Since the fitted forward model is identical to the generative model and the paths are fixed, the loss is zero at the true parameter values by construction. Recovering those values therefore demonstrates only that the optimizer can invert the simulator's own forward model, not that the simulator's paths or learned materials are physically accurate or that material properties can be learned from real channel data. The paper itself defers real-world measurements to future work, but the presented evidence for learning EM properties reduces to a self-consistency check.

full rationale

The central simulator contribution — efficient multi-bounce ray tracing on point clouds with a constant-ray-count visibility-matrix path search — is an algorithmic construction, not a fitted prediction. Its performance claims are supported by wall-clock times and path counts compared against the earlier baseline [7]; the absence of an exhaustive reference for path recall is an evaluation gap, not circularity. The disk-averaged intersection model (Eqs. 3–4) is a stated modeling choice with explicit parameters, not a hidden reuse of the target result. The cited prior work [7], [23] is used for standard path-minimization and hashing techniques, and the relevant equations are reproduced in the paper, so those self-citations are not load-bearing reductions. The only circular content is in the material-learning demonstration: synthetic ground truth is generated by the same differentiable pipeline that is then fitted, making the recovery a self-referential identifiability test rather than external validation. This does not undermine the ray-tracing derivation itself, but it does mean the learning claim's evidence is closed-loop. Overall score 4 reflects one partial circularity while the core method remains independently constructed.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The surface-aligned disks and visibility matrices are numerical constructs, not claimed physical objects. The central claim rests primarily on hand-set geometric parameters and on the completeness of the visibility-matrix path candidate search.

free parameters (5)

Voxel size rho = 0.0625 m
Discretization resolution for the point cloud; chosen by the authors and controls path candidate granularity and diffuse scattering area (Table I, Section II-A).
Point disk radius r_p = 0.015 m
Radius of surface-aligned disks used in intersection weighting (Eqs. 3-4, Table I); set by hand based on point cloud density.
Distance attenuation lambda_d = 100
Depth attenuation weight in Eq. 4; chosen by the authors to adapt to point cloud noise (Section II-C, Table I).
Refinement thresholds and iterations = xi=50, kappa=10, t_c=0.0001, t_a=1.0, t_d=0.01
Convergence and validity criteria for path refinement (Table I); hand-set algorithmic hyperparameters.
Material parameters for five labels = Recovered after 5000 iterations; values shown in Fig. 3
Relative permittivity, conductivity, and scattering coefficient for five material labels are optimized to match synthetic ground truth CIRs in Section III-B; these are fitted values, not independent predictions.

assumptions (6)

domain assumption Ray-optical approximation is valid at 8 GHz in indoor scenes.
The simulator models propagation as discrete rays; the paper does not validate this approximation against measurements or full-wave simulation.
domain assumption Point cloud normals and disk representation accurately approximate the true surface.
Eqs. 3-4 assume each point can be represented by a planar disk and that the weighted average of disk intersections is the true surface; accuracy depends on hand-set r_p and lambda_d.
domain assumption RX-to-DPS visibility matrix contains all interaction candidates needed for significant paths.
Section II-B derives path candidates by querying visibility from RXs to voxel reception points while casting constant-count rays from TX; completeness is assumed, not proven.
domain assumption Sionna's electromagnetic computations and gradients are correct.
The paper delegates all electromagnetic calculations and differentiability to the external Sionna library [21] without auditing its implementation.
ad hoc to paper Synthetic CIRs generated by the same simulator are a valid training target for the learning demonstration.
Section III-B generates ground truth CFRs with the same path engine and then recovers the parameters used to generate them; this is self-referential and not a test against measured channels.
domain assumption Diffuse scattering can be modeled as the final interaction only.
Section II-B1 states diffuse scattering is supported as the last interaction point; this limits generality and is a stated simplification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Differentiable High-Performance Ray Tracing-Based Simulation of Radio Propagation with Point Clouds." pith.science (2026). https://pith.science/paper/4KG2K25S

@misc{pith2026250704021,
  author       = {Pith},
  title        = {Pith review of: Differentiable High-Performance Ray Tracing-Based Simulation of Radio Propagation with Point Clouds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4KG2K25S}},
  note         = {Machine review of arXiv:2507.04021}
}

read the original abstract

Ray tracing is a widely used deterministic method for radio propagation simulations, capable of producing physically accurate multipath components. The accuracy depends on the quality of the environment model and its electromagnetic properties. Recent advances in computer vision and machine learning have made it possible to reconstruct detailed environment models augmented with semantic segmentation labels. In this letter, we propose a differentiable ray tracing-based radio propagation simulator that operates directly on point clouds. We showcase the efficiency of our method by simulating multi-bounce propagation paths with up to five interactions with specular reflections and diffuse scattering in two indoor scenarios, each completing in less than 90 ms. Lastly, we demonstrate how the differentiability of electromagnetic computations can be combined with segmentation labels to learn the electromagnetic properties of the environment.

Figures

Figures reproduced from arXiv: 2507.04021 by the authors.

**Figure 1.** Overview of how the simulator transforms the input point cloud into a ray traceable representation, followed by path [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Scenes and antenna positions used in the experiments. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Relative permittivity, conductivity, and scattering coefficient at each training iteration. The rightmost figure shows the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 21 canonical work pages

[7]

Ray launching- based computation of exact paths with noisy dense point clouds,

N. Vaara, P. Sangi, M. Bordallo L ´opez, and J. Heikkil¨a, “Ray launching- based computation of exact paths with noisy dense point clouds,” IEEE Transactions on Antennas and Propagation , 2025

work page 2025
[1]

Demonstration of a continuously updated, radio-compatible digital twin for robotic integrated sensing and communications,

V . C. Andrei, P. Susarla, A. Djuhera, N. Vaara, J. Mustaniemi, C. ´A. Casado, X. Li, U. J. M ¨onich, H. Boche, and M. Bordallo L ´opez, “Demonstration of a continuously updated, radio-compatible digital twin for robotic integrated sensing and communications,” in 2025 IEEE 5th International Symposium on Joint Communications & Sensing (JC&S) . IEEE, 2025, pp. 1–2

work page 2025
[2]

Real-time digital twins: Vision and research directions for 6g and beyond,

A. Alkhateeb, S. Jiang, and G. Charan, “Real-time digital twins: Vision and research directions for 6g and beyond,” IEEE Communications Magazine, 2023

work page 2023
[3]

Simulating specular reflections for point cloud geometrical database of the environment,

U. T. Virk, J.-F. Wagen, and K. Haneda, “Simulating specular reflections for point cloud geometrical database of the environment,” in 2015 Loughborough Antennas & Propagation Conference (LAPC) . IEEE, 2015, pp. 1–5

work page 2015
[4]

Indoor propagation chan- nel simulations at 60 GHz using point cloud data,

J. J ¨arvel¨ainen, K. Haneda, and A. Karttunen, “Indoor propagation chan- nel simulations at 60 GHz using point cloud data,” IEEE Transactions on Antennas and Propagation , vol. 64, no. 10, pp. 4457–4467, 2016

work page 2016
[5]

Point cloud ray-launching simulations of indoor multipath channels at 60 GHz,

P. Koivum ¨aki and K. Haneda, “Point cloud ray-launching simulations of indoor multipath channels at 60 GHz,” in 2022 IEEE 33rd Annual International Symposium on Personal, Indoor and Mobile Radio Com- munications (PIMRC). IEEE, 2022, pp. 01–07

work page 2022
[6]

Ray-optics simulations of outdoor-to-indoor multipath channels at 4 and 14 ghz,

P. Koivum ¨aki, A. Karttunen, and K. Haneda, “Ray-optics simulations of outdoor-to-indoor multipath channels at 4 and 14 ghz,” IEEE Transac- tions on Antennas and Propagation , 2023

work page 2023
[8]

A gpu-based radio wave propagation prediction with progressive processing on point cloud,

M. Pang, H. Wang, K. Lin, and H. Lin, “A gpu-based radio wave propagation prediction with progressive processing on point cloud,” IEEE Antennas and Wireless Propagation Letters , vol. 20, no. 6, pp. 1078–1082, 2021

work page 2021

Show all 26 references

[9]

Wip: Real-world 3d models derived from mobile mapping for ray launching based propagation loss modeling,

T. Wahl, D. Borrmann, M. Bleier, A. N ¨uchter, T. Wiemann, T. H ¨anel, and N. Aschenbruck, “Wip: Real-world 3d models derived from mobile mapping for ray launching based propagation loss modeling,” in 2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Mu...

2022
[10]

Indoor radio map construction via ray tracing with RGB-D sensor-based 3D reconstruction: Concept and experiments in WLAN systems,

N. Suga, Y . Maeda, and K. Sato, “Indoor radio map construction via ray tracing with RGB-D sensor-based 3D reconstruction: Concept and experiments in WLAN systems,” IEEE Access , vol. 11, pp. 24 863– 24 874, 2023

2023
[11]

Environment- driven mmwave beamforming for multi-user immersive applications,

A. Kamari, N. Wu, H. K. Narsani, B. Han, and P. Pathak, “Environment- driven mmwave beamforming for multi-user immersive applications,” in Proceedings of the 1st ACM Workshop on Mobile Immersive Computing, Networking, and Systems , 2023, pp. 268–275

2023
[12]

Path loss prediction in urban environments with sionna- rt based on accurate propagation scene models at 2.8 ghz,

G. Xia, C. Zhou, F. Zhang, Z. Cui, C. Liu, H. Ji, X. Zhang, Z. Zhao, and Y . Xiao, “Path loss prediction in urban environments with sionna- rt based on accurate propagation scene models at 2.8 ghz,” IEEE Transactions on Antennas and Propagation , 2024

2024
[13]

Rgb-d sensor- aided radio map estimation using materials classification,

N. Suga, N. Yoshida, R. Gozono, Y . Maeda, and K. Sato, “Rgb-d sensor- aided radio map estimation using materials classification,”IEEE Wireless Communications Letters, 2025

2025
[14]

ITU-R P.2040,

“ITU-R P.2040,” https://www.itu.int/rec/R-REC-P.2040/en, accessed: 24.04.2025

2025
[15]

Learning radio environments by dif- ferentiable ray tracing,

J. Hoydis, F. A. Aoudia, S. Cammerer, F. Euchner, M. Nimier-David, S. Ten Brink, and A. Keller, “Learning radio environments by dif- ferentiable ray tracing,” IEEE Transactions on Machine Learning in Communications and Networking , 2024

2024
[16]

Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,

S. Jiang, Q. Qu, X. Pan, A. Agrawal, R. Newcombe, and A. Alkhateeb, “Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,” IEEE Open Journal of the Communications Society, 2025

2025
[17]

Mask3d: Mask transformer for 3d semantic instance segmentation,

J. Schult, F. Engelmann, A. Hermans, O. Litany, S. Tang, and B. Leibe, “Mask3d: Mask transformer for 3d semantic instance segmentation,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 8216–8223

2023
[18]

Oneformer3d: One transformer for unified point cloud segmentation,

M. Kolodiazhnyi, A. V orontsova, A. Konushin, and D. Rukhovich, “Oneformer3d: One transformer for unified point cloud segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 943–20 953

2024
[19]

Ray tracing for radio propagation modeling: Principles and applications,

Z. Yun and M. F. Iskander, “Ray tracing for radio propagation modeling: Principles and applications,” IEEE Access, vol. 3, pp. 1089–1100, 2015

2015
[20]

A discrete environment-driven GPU-based ray launching algorithm,

J. S. Lu, E. M. Vitucci, V . Degli-Esposti, F. Fuschini, M. Barbiroli, J. A. Blaha, and H. L. Bertoni, “A discrete environment-driven GPU-based ray launching algorithm,” IEEE Transactions on Antennas and Propagation, vol. 67, no. 2, pp. 1180–1192, 2018

2018
[21]

Sionna RT: Differentiable ray tracing for radio propagation modeling,

J. Hoydis, F. A. Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” arXiv preprint arXiv:2303.11103 , 2023

2023 arXiv
[22]

OptiX: A general purpose ray tracing engine,

S. G. Parker, J. Bigler, A. Dietrich, H. Friedrich, J. Hoberock, D. Luebke, D. McAllister, M. McGuire, K. Morley, A. Robison et al. , “OptiX: A general purpose ray tracing engine,” ACM Transactions on Graphics , vol. 29, no. 4, pp. 1–13, 2010

2010
[23]

A refined path generation pipeline for radio channel propagation modeling,

N. Vaara, P. Sangi, J. Pyhtil ¨a, M. Juntti, and J. Heikkil ¨a, “A refined path generation pipeline for radio channel propagation modeling,” in 2023 17th European Conference on Antennas and Propagation (EuCAP) . IEEE, 2023, pp. 1–5

2023
[24]

Geometrical theory of diffraction,

J. B. Keller, “Geometrical theory of diffraction,” Josa, vol. 52, no. 2, pp. 116–130, 1962

1962
[25]

The replica dataset: A digital replica of indoor spaces,

J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Verma et al. , “The replica dataset: A digital replica of indoor spaces,” arXiv preprint arXiv:1906.05797 , 2019

1906 arXiv
[26]

Mea- surement and modelling of scattering from buildings,

V . Degli-Esposti, F. Fuschini, E. M. Vitucci, and G. Falciasecca, “Mea- surement and modelling of scattering from buildings,”IEEE transactions on antennas and propagation , vol. 55, no. 1, pp. 143–153, 2007

2007

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [7]

Ray launching- based computation of exact paths with noisy dense point clouds,

N. Vaara, P. Sangi, M. Bordallo L ´opez, and J. Heikkil¨a, “Ray launching- based computation of exact paths with noisy dense point clouds,” IEEE Transactions on Antennas and Propagation , 2025

work page 2025

[2] [1]

Demonstration of a continuously updated, radio-compatible digital twin for robotic integrated sensing and communications,

V . C. Andrei, P. Susarla, A. Djuhera, N. Vaara, J. Mustaniemi, C. ´A. Casado, X. Li, U. J. M ¨onich, H. Boche, and M. Bordallo L ´opez, “Demonstration of a continuously updated, radio-compatible digital twin for robotic integrated sensing and communications,” in 2025 IEEE 5th International Symposium on Joint Communications & Sensing (JC&S) . IEEE, 2025, pp. 1–2

work page 2025

[3] [2]

Real-time digital twins: Vision and research directions for 6g and beyond,

A. Alkhateeb, S. Jiang, and G. Charan, “Real-time digital twins: Vision and research directions for 6g and beyond,” IEEE Communications Magazine, 2023

work page 2023

[4] [3]

Simulating specular reflections for point cloud geometrical database of the environment,

U. T. Virk, J.-F. Wagen, and K. Haneda, “Simulating specular reflections for point cloud geometrical database of the environment,” in 2015 Loughborough Antennas & Propagation Conference (LAPC) . IEEE, 2015, pp. 1–5

work page 2015

[5] [4]

Indoor propagation chan- nel simulations at 60 GHz using point cloud data,

J. J ¨arvel¨ainen, K. Haneda, and A. Karttunen, “Indoor propagation chan- nel simulations at 60 GHz using point cloud data,” IEEE Transactions on Antennas and Propagation , vol. 64, no. 10, pp. 4457–4467, 2016

work page 2016

[6] [5]

Point cloud ray-launching simulations of indoor multipath channels at 60 GHz,

P. Koivum ¨aki and K. Haneda, “Point cloud ray-launching simulations of indoor multipath channels at 60 GHz,” in 2022 IEEE 33rd Annual International Symposium on Personal, Indoor and Mobile Radio Com- munications (PIMRC). IEEE, 2022, pp. 01–07

work page 2022

[7] [6]

Ray-optics simulations of outdoor-to-indoor multipath channels at 4 and 14 ghz,

P. Koivum ¨aki, A. Karttunen, and K. Haneda, “Ray-optics simulations of outdoor-to-indoor multipath channels at 4 and 14 ghz,” IEEE Transac- tions on Antennas and Propagation , 2023

work page 2023

[8] [8]

A gpu-based radio wave propagation prediction with progressive processing on point cloud,

M. Pang, H. Wang, K. Lin, and H. Lin, “A gpu-based radio wave propagation prediction with progressive processing on point cloud,” IEEE Antennas and Wireless Propagation Letters , vol. 20, no. 6, pp. 1078–1082, 2021

work page 2021