Unreal is all you need: Multimodal ISAC Data Simulation with Only One Engine

Jun Jiang; Kongwu Huang; Shiyi Mu; Shugong Xu; Yuan Gao

REVIEW 3 major objections 6 minor 1 cited by

Unreal is all you need: Multimodal ISAC Data Simulation with Only One Engine

T0 review · 3 major / 6 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 The paper claims that a single Unreal Engine-based platform, Great-X, can reimplement Sionna's ray tracing and generate synchronized CSI, RGB, radar, and LiDAR data whose channel outputs transfer across simulators for CSI-based UAV…

desk verdict Great-X is a genuinely useful single-engine simulator with a new dataset, but the fidelity and generalization claims rest on a circular comparison against the very engine it reimplements. read the letter →

arxiv 2507.08716 v3 pith:KQ5NJCGU submitted 2025-07-11 cs.CV

Kongwu Huang , Shiyi Mu , Jun Jiang , Yuan Gao , Shugong Xu This is my paper

classification cs.CV

keywords multimodalISACsimulationraytracingchannelmodelingUnrealEngineSionnaCSIUAVlocalizationintegratedsensingandcommunicationdigitaltwin

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

The paper aims to establish that a single open-source simulation engine, Unreal Engine, can act as a complete multimodal data twin for integrated sensing and communication (ISAC), replacing the multi-engine and commercial pipelines currently used to generate CSI, RGB, radar, and LiDAR data. It does so by reimplementing the ray-tracing channel computations of Sionna inside Unreal Engine, then building the Great-MSD dataset of 100 low-altitude UAV trajectories with synchronized radio, vision, depth, and position outputs. To probe whether this reconstructed channel model is physically faithful, the authors train a CSI-based 3D localization network on their own generated data and test it on data from the original Sionna ray tracer. The reported transfer result, 8.107 m average error against 11.082 m in the reverse direction, supports the idea that the two simulators agree on line-of-sight propagation and that single-engine generated data is usable for cross-simulator learning.

What carries the argument

The load-bearing object is Great-X, an Unreal Engine-based reimplementation of Sionna's electromagnetic ray-tracing model. The mechanism that carries the argument is the set of propagation equations from SionnaRT, namely polarization basis decomposition, Fresnel reflection and transmission coefficients, and the diffuse scattering model with a scattering coefficient S controlling the split of reflected and scattered energy, rewritten in C++ inside Unreal Engine. Because the same engine renders optical images and traces radio paths over identical scene geometry, the platform can emit synchronized RGB, depth, CSI, and position data from one run; the authors argue the renderer's high-fidelity geometry and textures also improve non-line-of-sight multipath modeling compared with SionnaRT's Mitsuba-based renderer.

What would settle it

A concrete falsifier would be to generate CSI in a simple scene with a known analytic solution, such as a single dielectric slab, and compare Great-X's output against both SionnaRT and the closed-form Fresnel result across polarizations and incidence angles; if Great-X does not converge to the analytic solution as the mesh is refined, or if the two simulators diverge substantially on line-of-sight paths, the claim that the single-engine reconstruction is physically faithful would be contradicted.

Watch

Extended reading notes

Core claim

The central claim of the paper is that reconstructing Sionna's ray-tracing channel model inside Unreal Engine yields physically plausible, synchronized multimodal ISAC data, and that this data transfers across simulation engines. The authors implement the Sionna formulas for TE/TM polarization, Fresnel reflection and refraction, and diffuse scattering in Unreal Engine, and use the resulting platform, Great-X, to generate the Great-MSD dataset. A ResNet-50 localization model trained on Great-MSD reaches 0.647 m average error on native test data and 8.107 m on unseen SionnaRT test trajectories, while a SionnaRT-trained model reaches 0.309 m natively but 11.082 m on Great-MSD test data. Angle-delay profile comparisons show strong line-of-sight agreement between simulators, with divergences in non-line-of-sight paths that the authors attribute to Unreal Engine's higher geometric and texture fidelity.

Load-bearing premise

The load-bearing premise is that the Unreal Engine implementation faithfully reproduces Sionna's ray-tracing physics, so the observed line-of-sight agreement between the two simulators reflects genuine fidelity rather than a shared implementation error, and no real-world data is used to anchor either simulator.

Editorial extensions

If this is right

If the transfer result holds, generating large-scale multimodal ISAC datasets no longer requires coupling a wireless ray tracer to a separate driving or vision simulator; one engine produces synchronized radio and optical data from the same scene.
A model trained on Great-X data transfers to SionnaRT channels with 8.107 m average error, suggesting Great-X data can serve as a pretraining or augmentation source for CSI-based tasks even when deployment uses a different simulator.
The strong line-of-sight agreement between the two engines indicates the core wave-propagation physics is reproduced, so differences between simulators can be analyzed as renderer-fidelity effects rather than fundamental model mismatches.
Because Great-X and Great-MSD are open-source, they give the community a reproducible single-engine alternative to closed or multi-engine frameworks, supporting standardized evaluation of ISAC algorithms.

Reading between the lines

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

An implied extension of the paper's protocol is to use bi-directional cross-simulator evaluation as a general benchmark for any new ISAC simulator: the smaller the asymmetric transfer error, the more engine-independent the generated channel data.
Because no real-world measurement anchors either simulator, the 8.107 m transfer result establishes inter-simulator consistency, not physical accuracy; a field-measurement comparison would be needed to determine whether Unreal Engine's claimed NLOS fidelity is real.
If renderer fidelity is what drives NLOS differences, then the same single-engine platform could plausibly improve other ray-traced modalities such as mmWave radar and LiDAR, pointing toward multimodal consistency as a testable benefit rather than just a convenience.
A practical test suggested by the paper's logic: swapping the scene's material textures while keeping geometry fixed should change NLOS CSI in ways consistent with the scattering model; if it does not, the claimed fidelity mechanism would need revision.

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

Here's my take on the Great-X paper. The engineering contribution is real: a single Unreal Engine pipeline that outputs synchronized CSI, RGB, depth, and radar/LiDAR data is genuinely useful for ISAC research, and the Great-MSD dataset (100 UAV trajectories, 1000 frames each) is a concrete resource. That's the part worth publishing.

What's new is the integration, not the physics. Section II explicitly copies the reflection and scattering equations from SionnaRT, which is fine—but it means the later cross-platform comparison is a self-check, not an independent validation. The paper shows LOS agreement and NLOS divergence between Great-X and Sionna, then attributes the NLOS difference to Unreal's "higher geometric precision" with no external evidence. Implementation error is an equally plausible explanation. The load-bearing claim—that Great-X is more physically accurate—is unfalsifiable as presented.

The soft spots, in order of severity. First, no error bars or statistical tests on any of the transfer numbers, and 10 test trajectories is thin. Second, the generalization claim is overstated: 8.107 m on Sionna data versus 0.647 m on the native test set is a 12.5× gap, and the reverse direction is 11 m. That is a large domain shift, not "strong cross-platform generalization." Third, no real-world channel measurements or third-party simulator anchors either side, so the NLOS fidelity claim has no ground truth. Fourth, the code and dataset are promised but not yet public, so I can't verify the implementation or reproduce the numbers.

Credit where it's due: the single-engine design removes synchronization overhead, the authors are transparent that the equations come from SionnaRT, and the convergence curve shows the dataset supports learning. The paper is a building block, not a breakthrough.

Who is this for? People who need quick, synchronized synthetic ISAC data for baseline deep-learning experiments, especially low-altitude UAV scenarios. They should treat Great-X as a useful data generator, not as a validated physics model.

I'd accept this for peer review with expectation of major revision. The platform deserves publication, but the validation needs an external anchor—measured data or a third independent simulator—plus error bars and a tempered claim. If the authors release code and data as promised, the field will benefit.

Referee Report

3 major / 6 minor

Summary. The paper introduces Great-X, an Unreal Engine-based simulation platform that reimplements Sionna's ray-tracing channel equations in C++ and synchronizes them with RGB, depth, radar, and LiDAR simulation in a single engine. Using Great-X, the authors construct a large-scale multimodal dataset, Great-MSD, consisting of 100 low-altitude UAV trajectories in a Lujiazui urban scene, each with CSI, camera images, depth maps, and positional metadata. They then train a ResNet-50-based CSI regressor and evaluate bidirectional transfer between Great-X and SionnaRT, reporting native errors of 0.309 m and 0.647 m and cross-platform errors of 8.107 m and 11.082 m. The paper claims strong cross-platform generalization and attributes observed LOS/NLOS differences between the simulators to higher geometric and textural fidelity in Unreal's renderer.

Significance. If the central claims are substantiated, this would be a valuable open-source contribution: a single-engine, reproducible multimodal ISAC simulation pipeline with a public dataset, a baseline localization method, and the practical benefit of avoiding multi-engine co-simulation complexity. The authors are explicit about the source of the electromagnetic equations and provide a concrete dataset and GitHub link, which are strengths for reproducibility. However, the evidence for physical fidelity and cross-platform generalization is currently self-referential: Great-X is validated primarily by comparison against Sionna, which is the very system it reimplements. No real-world channel measurements or independent ray tracer are used as ground truth, and the only quantitative generalization evidence is a cross-simulator error that is more than an order of magnitude larger than the native error, computed on only 10 test trajectories. The significance of the platform and dataset is therefore real but conditional on stronger validation.

major comments (3)

[Section IV, Table II] The cross-platform generalization claim is not supported by the reported statistics. The Great-MSD-trained model's cross-platform error of 8.107 m is 12.5 times its native error of 0.647 m, and the reverse transfer error is 11.082 m versus 0.309 m natively. With only 10 test trajectories, no error bars, no per-trajectory distribution, and no comparison against trivial baselines (e.g., predicting the mean position), these numbers are equally consistent with both simulators sharing coarse LOS-dominated features that suffice for a rough localization, rather than with the simulators sharing high-fidelity channel physics. I recommend reporting confidence intervals, per-trajectory errors, paired statistical tests, and the performance of a non-informative baseline, as well as testing with more trajectories.
[Section II and Section IV] The validation is circular in its current form. The paper states in Section II that all propagation formulas are 'referenced from SionnaRT,' and Section IV's only quantitative fidelity evidence is agreement between Great-X and SionnaRT. If the implementation is correct, agreement with Sionna is partly built in by construction, so it verifies implementation consistency, not physical fidelity. The subsequent claim that Great-X's NLOS discrepancies reflect 'higher geometric precision and more detailed texture information' is an interpretation without an independent reference; the discrepancies could equally come from implementation errors, different scattering models, or unmodeled physics. To make the fidelity claim falsifiable, the authors should compare both simulators against measured channel data in a comparable scene or against an independent ray tracer such as Wireless InSite, or at minimum run controlled sensitivity studies varying material parameters and scene geometry.
[Section IV, Fig. 7] The ADP comparison is anecdotal and not quantified. Figure 7 shows one visual example of an angle-delay profile, with LOS agreement and NLOS divergence asserted qualitatively. No metric is defined for LOS/NLOS agreement, no aggregation across trajectories or time frames is provided, and the method for computing the ADP from CSI is not described. Without a quantitative evaluation, the visual divergence cannot support the conclusion that Great-X is more accurate in NLOS rather than simply different.

minor comments (6)

[Abstract and GitHub link] The abstract and full text give the repository as https://github.com/hkw-xg/Great-MCD, but the dataset is named Great-MSD throughout; please confirm the correct URL and dataset name.
[Table II] The table headers contain typos: 'Sionna RT-trian' and 'Great-MSD-trian' should be 'train' or 'training'; also consider spelling out 'SionnaRT' consistently.
[Fig. 5 caption] The caption lists '(a) RGB Image, (b) Depth Image' but the body text says the panels respectively show the depth map and RGB image; the order is inconsistent.
[Section IV] The ResNet-50 input representation of CSI is not described: how is the complex CSI tensor shaped, normalized, and fed into the network? This information is necessary for reproducibility.
[Section III] The dataset description should state whether the 100 trajectories are all in the same urban scene and whether the scene geometry, including vehicle traffic and dynamic objects, is static or time-varying; this affects the interpretation of 'scale' and generalization.
[Section IV] The claim that the SionnaRT-trained model 'reflects its overfitting' because it achieves 0.309 m natively but 11.082 m on Great-MSD data is plausible but not established without checking a validation set and measuring generalization within Sionna; consider reporting model capacity or regularization analysis.

Circularity Check

1 steps flagged · score 6.0 of 10

Cross-platform validation is self-referential: Great-X is defined as a reimplementation of SionnaRT, and its headline evidence of 'generalization' is agreement with SionnaRT itself.

self definitional [Section II ('The following formulas are referenced from SionnaRT') and Section IV (Table II and Fig. 7 ADP comparison).]
"The following formulas are referenced from SionnaRT [13]. ... As shown in Table II, the model trained on the Great-MSD exhibits strong cross-platform generalization capabilities. When evaluated on 10 unseen test trajectories generated by the SionnaRT, it achieves an average positioning error of 8.107 meters."

Great-X is constructed to reproduce SionnaRT's ray-tracing computation: the paper states it 'reconstructs the ray-tracing computation of Sionna' and that its EM formulas are 'referenced from SionnaRT'. The main quantitative validation is then a comparison with SionnaRT: a Great-MSD-trained CSI localizer is tested on SionnaRT-generated trajectories, and the resulting 8.107 m error is presented as 'strong cross-platform generalization'. Because SionnaRT supplies the defining equations of Great-X, agreement between the two is built in to the extent the reimplementation is faithful; the experiment tests internal consistency between two implementations of the same model, not generalization to an independent physical reference.

full rationale

The paper's central derivation is not a formal theorem chain; it is an engineering claim that Great-X faithfully reproduces SionnaRT's channel simulation inside Unreal Engine. That claim is supported by quoting SionnaRT's formulas and by showing LOS-agreeing ADPs and a localization transfer error of 8.107 m between Great-MSD and SionnaRT. This is self-referential validation: the reference simulator is the same source from which the equations were taken, so the comparison is an implementation-fidelity check rather than an independent confirmation of physical fidelity or cross-engine generality. The NLOS claim that Unreal's renderer gives 'higher geometric precision and more detailed texture information' and therefore 'improved accuracy' is asserted without any ground-truth channel measurements or third-party ray tracer; that is an unsupported correctness claim, though not itself a circular reduction. No load-bearing self-citation by the authors is involved, and the paper is honest about releasing code and data, but the advertised 'generalizability across different CSI simulation engines' rests on comparing a reimplementation to its own template.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the correctness and faithful implementation of SionnaRT's equations, on the representativeness of the simulated urban scene, and on the unquantified assumption that renderer fidelity explains cross-simulator differences. No free parameters are fitted in the paper; the trained localizer's weights are not treated as model parameters here.

assumptions (3)

domain assumption The SionnaRT ray-tracing equations for reflection, refraction, and diffuse scattering are physically correct and applicable to the simulated scenarios.
Section II states the formulas are referenced from SionnaRT [13] and uses them without independent derivation or experimental validation.
ad hoc to paper The Unreal Engine's geometric and texture detail yields higher-fidelity NLOS multipath modeling than Sionna's Mitsuba renderer.
Section IV uses this to explain ADP differences between the simulators, but no quantitative evidence is given that the differences come from rendering fidelity rather than implementation differences.
domain assumption The simulated Lujiazui scene and flight trajectories are representative of real low-altitude UAV communication environments.
Section III and Table I define the scene and flight parameters; no comparison to real urban measurements is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unreal is all you need: Multimodal ISAC Data Simulation with Only One Engine." pith.science (2026). https://pith.science/paper/KQ5NJCGU

@misc{pith2026250708716,
  author       = {Pith},
  title        = {Pith review of: Unreal is all you need: Multimodal ISAC Data Simulation with Only One Engine},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KQ5NJCGU}},
  note         = {Machine review of arXiv:2507.08716}
}

read the original abstract

Scaling laws have achieved success in LLM and foundation models. To explore their potential in ISAC research, we propose Great-X. This single-engine multimodal data twin platform reconstructs the ray-tracing computation of Sionna within Unreal Engine and is deeply integrated with autonomous driving tools. This enables efficient and synchronized simulation of multimodal data, including CSI, RGB, Radar, and LiDAR. Based on this platform, we construct an open-source, large-scale, low-altitude UAV multimodal synaesthesia dataset named Great-MSD, and propose a baseline CSI-based UAV 3D localization algorithm, demonstrating its feasibility and generalizability across different CSI simulation engines. The related code and dataset will be made available at: https://github.com/hkw-xg/Great-MCD.

Figures

Figures reproduced from arXiv: 2507.08716 by the authors.

**Figure 2.** Comparing Great-X and DeepVerse 6G, the data generation of SynthSoM. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Reflection and refraction of a plane wave at a plane [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Diffuse and specular reflection of an incoming wave [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: (a)–(d) respectively illustrate the depth map, RGB image, CSI beam visualization, and the corresponding 3D spatial [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Relationship between the number of training trajecto [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Comparison of ADPs from different simulators: (a) for SionnaRT, and (b) for Great-X. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Towards channel foundation models (CFMs): Motivations, methodologies and opportunities
eess.SP 2025-07 conditional novelty 4.0 of 10

A survey and position paper proposing channel foundation models, with experiments on two pretrained CSI models showing gains over a vanilla ViT baseline.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications,

L. Alzubaidi, J. Bai, A. Al-Sabaawi, J. Santamar ´ıa, A. S. Albahri, B. S. N. Al-Dabbagh, M. A. Fadhel, M. Manoufali, J. Zhang, A. H. Al-Timemy et al. , “A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications,” Journal of Big Data , vol. 10, no. 1, p. 46, 2023

work page 2023
[2]

On the performance of an integrated communication and localization system: an analytical framework,

Y . Gao, H. Hu, J. Zhang, Y . Jin, S. Xu, and X. Chu, “On the performance of an integrated communication and localization system: an analytical framework,” IEEE Transactions on Vehicular Technology, vol. 73, no. 7, pp. 10 845–10 849, 2024

work page 2024
[3]

A stochastic geometry-based analytical framework for integrated localization and communication systems,

Y . Gao, H. Du, Z. Jiang, H. Hu, J. Zhang, S. Zhang, J. Du, F. R. Yu, and S. Xu, “A stochastic geometry-based analytical framework for integrated localization and communication systems,” IEEE Internet of Things Journal, 2025

work page 2025
[4]

Joint channel estimation and data detection for otfs systems: A lightweight deep learning framework with a novel data augmentation method,

Y . Gao, X. Xu, Y . Jin, W. Yuan, J. Zhang, and S. Xu, “Joint channel estimation and data detection for otfs systems: A lightweight deep learning framework with a novel data augmentation method,” IEEE Internet of Things Journal , 2025

work page 2025
[5]

Multimodal transformers for wireless communications: A case study in beam pre- diction,

Y . Tian, Q. Zhao, F. Boukhalfa, K. Wu, F. Bader et al. , “Multimodal transformers for wireless communications: A case study in beam pre- diction,” arXiv preprint arXiv:2309.11811 , 2023

arXiv 2023
[6]

Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,

A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,” arXiv preprint arXiv:1902.06435, 2019

arXiv 1902
[7]

[Online]

Wireless insite propagation modeling software. [Online]. Available: https://www.remcom.com/wireless-insite-propagation-software

work page
[8]

DeepVerse 6G: A dataset generation framework for multi-modal sensing and communica- tion digital twins,

U. Demirhan, A. Taha, S. Jiang, and A. Alkhateeb, “DeepVerse 6G: A dataset generation framework for multi-modal sensing and communica- tion digital twins,” preprint, Feb 2025

work page 2025

Show all 16 references

[9]

Carla: An open urban driving simulator,

A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “Carla: An open urban driving simulator,” in Conference on robot learning . PMLR, 2017, pp. 1–16

2017
[10]

Synthsom: A synthetic intelligent multi-modal sensing-communication dataset for synesthesia of machines (som),

X. Cheng, Z. Huang, Y . Yu, L. Bai, M. Sun, Z. Han, R. Zhang, and S. Li, “Synthsom: A synthetic intelligent multi-modal sensing-communication dataset for synesthesia of machines (som),” Scientific Data , vol. 12, 2025

2025
[11]

Wavefarer automotive radar simula- tion software

Remcom Inc. Wavefarer automotive radar simula- tion software. [Online]. Available: https://www.remcom.com/ wavefarer-automotive-radar-software
[12]

Airsim: High-fidelity visual and physical simulation for autonomous vehicles,

S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for autonomous vehicles,” in Field and Service Robotics: Results of the 11th International Conference . Springer, 2018, pp. 621–635

2018
[13]

[Online]

Sionna. [Online]. Available: https://nvlabs.github.io/sionna/rt
[14]

Recommendation itu-r p.2040-3: Effects of building materials and structures on radiowave propagation above about 100 mhz,

ITU, “Recommendation itu-r p.2040-3: Effects of building materials and structures on radiowave propagation above about 100 mhz,” International Telecommunication Union, Tech. Rep., Sep. 2023, series P

2023
[15]

Unreal engine official website

Epic Games. Unreal engine official website. [Online]. Available: https://www.unrealengine.com
[16]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

2016

Pith tools

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

[1] [1]

A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications,

L. Alzubaidi, J. Bai, A. Al-Sabaawi, J. Santamar ´ıa, A. S. Albahri, B. S. N. Al-Dabbagh, M. A. Fadhel, M. Manoufali, J. Zhang, A. H. Al-Timemy et al. , “A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications,” Journal of Big Data , vol. 10, no. 1, p. 46, 2023

work page 2023

[2] [2]

On the performance of an integrated communication and localization system: an analytical framework,

Y . Gao, H. Hu, J. Zhang, Y . Jin, S. Xu, and X. Chu, “On the performance of an integrated communication and localization system: an analytical framework,” IEEE Transactions on Vehicular Technology, vol. 73, no. 7, pp. 10 845–10 849, 2024

work page 2024

[3] [3]

A stochastic geometry-based analytical framework for integrated localization and communication systems,

Y . Gao, H. Du, Z. Jiang, H. Hu, J. Zhang, S. Zhang, J. Du, F. R. Yu, and S. Xu, “A stochastic geometry-based analytical framework for integrated localization and communication systems,” IEEE Internet of Things Journal, 2025

work page 2025

[4] [4]

Joint channel estimation and data detection for otfs systems: A lightweight deep learning framework with a novel data augmentation method,

Y . Gao, X. Xu, Y . Jin, W. Yuan, J. Zhang, and S. Xu, “Joint channel estimation and data detection for otfs systems: A lightweight deep learning framework with a novel data augmentation method,” IEEE Internet of Things Journal , 2025

work page 2025

[5] [5]

Multimodal transformers for wireless communications: A case study in beam pre- diction,

Y . Tian, Q. Zhao, F. Boukhalfa, K. Wu, F. Bader et al. , “Multimodal transformers for wireless communications: A case study in beam pre- diction,” arXiv preprint arXiv:2309.11811 , 2023

arXiv 2023

[6] [6]

Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,

A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,” arXiv preprint arXiv:1902.06435, 2019

arXiv 1902

[7] [7]

[Online]

Wireless insite propagation modeling software. [Online]. Available: https://www.remcom.com/wireless-insite-propagation-software

work page

[8] [8]

DeepVerse 6G: A dataset generation framework for multi-modal sensing and communica- tion digital twins,

U. Demirhan, A. Taha, S. Jiang, and A. Alkhateeb, “DeepVerse 6G: A dataset generation framework for multi-modal sensing and communica- tion digital twins,” preprint, Feb 2025

work page 2025