{"id":"8c749a56-8af1-4fb5-9c3c-7444d7edc298","arxiv_id":"2507.08716","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Great-X reproduces Sionna's ray-tracing channel model inside Unreal Engine and provides Great-MSD, a multimodal UAV dataset for CSI localization.","lead":"A new simulation platform called Great-X generates synchronized wireless, camera, radar, and LiDAR data inside Unreal Engine for 6G research. It also releases a UAV dataset and shows a CSI-based localizer works, though cross-simulator transfer is limited.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-platform validation is circular: Great-X is compared only against Sionna, the engine it reimplements, and the NLOS fidelity claim is asserted without any independent reference.","rationale":"The reader's weakest_assumption correctly identifies that the validation is circular: Great-X is compared against Sionna, the very implementation it claims to reconstruct, with no real-world anchor. My stress-test agrees and sharpens the concern: the paper interprets NLOS disagreement as evidence of Great-X's superiority, but this is an interpretive leap unsupported by any external reference. The transfer numbers in Table II (8.107 m vs 0.647 m native, and 11.082 m in the reverse direction) are the only quantitative evidence, and they are presented without error bars, significance tests, or a clear statement of what error level would count as 'generalizable.' Since the paper's central contribution is a simulator and dataset, the lack of code/data availability further prevents independent verification. These issues do not invalidate the engineering effort, but they do mean the strongest claims—'reconstructs Sionna' and 'strong cross-platform generalization'—are not yet substantiated. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only with the release of artifacts and validation against an independent reference. My analysis does not move the verdict; it reinforces the conditions already stated. I agree with the reader's weakest_assumption and do not see a more load-bearing concern elsewhere, such as in the ray-tracing equations themselves, which are standard SionnaRT formulas and are not the source of risk. The risk is in the validation methodology, not the math.","tokens_in":6484,"tokens_out":1793,"duration_ms":25562,"concrete_test":"Release the Great-X and SionnaRT generation scripts and, for the same Lujiazui scene and identical antenna/geometry parameters, compute quantitative channel metrics (e.g., power delay profile, angular-delay profile, and path-loss) at many receiver positions. Then compare both simulators against either (a) a real-world measured channel in a similar low-altitude UAV scenario or (b) an independent, well-validated ray tracer such as Wireless InSite. If Great-X matches the independent reference better than Sionna in NLOS, the fidelity claim lands. Additionally, rerun the Table II localization transfer with at least 10 random seeds and report mean ± std to determine whether 8.107 m is significantly better than chance or than the reverse transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Great-X reconstructs Sionna's ray tracing and that its NLOS differences reflect higher physical fidelity—rests entirely on a comparison between Great-X and Sionna itself. Section IV's ADP comparison shows LOS agreement and NLOS divergence, but the paper attributes that divergence to 'higher geometric precision and more detailed texture information' in Unreal's renderer without any ground truth or independent simulator to support the direction of the interpretation. The discrepancy could equally arise from implementation errors, different scattering models, or unmodeled physics in Great-X. The only quantitative support for 'generalizability across different CSI simulation engines' is Table II: Great-MSD-trained model achieves 8.107 m on Sionna test data, versus 0.647 m on its native test set, while the reverse transfer is 11.082 m. This asymmetry and the absence of error bars or statistical tests make it unclear whether 8.107 m represents meaningful cross-platform generalization or simply reflects that both simulators share enough LOS-dominated features for a coarse localization task. Because no real-world channel measurements or third-party ray tracer anchors either simulator, the claim that Great-X is more accurate in NLOS is unfalsifiable as presented. This is the load-bearing weak point: the paper's headline evidence for fidelity and generalization is a self-comparison with an unverified reference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6718,"tokens_out":2652,"duration_ms":34655,"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":[{"comment":"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":"Section IV, Table II"},{"comment":"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":"Section II and Section IV"},{"comment":"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.","section":"Section IV, Fig. 7"}],"minor_comments":[{"comment":"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.","section":"Abstract and GitHub link"},{"comment":"The table headers contain typos: 'Sionna RT-trian' and 'Great-MSD-trian' should be 'train' or 'training'; also consider spelling out 'SionnaRT' consistently.","section":"Table II"},{"comment":"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":"Fig. 5 caption"},{"comment":"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":"Section IV"},{"comment":"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":"Section III"},{"comment":"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.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The central idea is timely, and an open-source single-engine multimodal ISAC simulator would be a useful community resource. However, the manuscript's key evidence for physical fidelity and cross-engine generalization is a self-comparison between the reimplementation and the reference implementation, with no external ground truth. I would encourage the authors to add an independent validation step—measured channels, a third-party ray tracer, or at least a carefully controlled sensitivity analysis—and to substantially strengthen the statistical treatment of the localization results. The dataset and code availability statement should also be verified, as the URL currently points to 'Great-MCD' while the paper describes 'Great-MSD'."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat'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.\n\nThe 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.\n\nCredit 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.\n\nWho 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.\n\nI'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.","headline":"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.","tokens_in":7265,"tokens_out":2766,"would_cite":false,"duration_ms":29867,"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":"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…","keywords":["multimodal ISAC simulation","ray tracing channel modeling","Unreal Engine","Sionna","CSI","UAV localization","integrated sensing and communication","digital twin"],"falsifier":"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.","tokens_in":6248,"feed_emoji":"📡","tokens_out":7584,"duration_ms":80584,"temperature":0.7,"pith_summary":"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.","feed_headline":"One engine now generates CSI, RGB, radar, and LiDAR data","feed_subtitle":"Cross-simulator tests show CSI trained on it localizes UAVs with 8.1 m average error.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the SionnaRT ray-tracing equations for polarization, reflection, refraction, and diffuse scattering that Great-X reimplements inside Unreal Engine.","marker":"[13]"},{"why":"Identifies Unreal Engine as the single engine in which the platform is built and through which high-fidelity geometry and rendering are provided.","marker":"[15]"},{"why":"Provides DeepMIMO as the established single-modality CSI dataset that the authors contrast with their multimodal, user-configurable platform.","marker":"[6]"},{"why":"Represents the closed-source, dual-engine multimodal ISAC dataset generation framework that Great-X aims to replace with an open single-engine architecture.","marker":"[8]"},{"why":"Represents the multi-engine heterogeneous simulation paradigm, coupling commercial ray tracing with AirSim, which the authors identify as limiting openness and extensibility.","marker":"[10]"},{"why":"Supplies the ResNet-50 backbone used in the CSI-based 3D localization baseline that demonstrates cross-platform generalization.","marker":"[16]"},{"why":"Provides the ITU-R material parameter definitions adopted in the channel model for the reflection and scattering calculations.","marker":"[14]"}],"fun_headline_variants":["One engine to simulate them all: CSI, RGB, radar, LiDAR","Unreal Engine replaces Sionna for multimodal ISAC data","Single-engine ISAC simulation yields cross-engine localization","Great-X: one engine for CSI, RGB, radar, and LiDAR","From Sionna to Unreal: multimodal ISAC data with one engine"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One engine to simulate them all: CSI, RGB, radar, LiDAR","Unreal Engine replaces Sionna for multimodal ISAC data","Single-engine ISAC simulation yields cross-engine localization","Great-X: one engine for CSI, RGB, radar, and LiDAR","From Sionna to Unreal: multimodal ISAC data with one engine"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000838,"raw_usage":{"total_tokens":3616,"prompt_tokens":871,"completion_tokens":2745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":2653}},"tokens_in":487,"tokens_out":2745,"duration_ms":20413,"temperature":1.0,"reasoning_tokens":2653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:10:31.480400+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"[Online]","cited_arxiv_id":null,"evidence_quote":"Supplies the SionnaRT ray-tracing equations for polarization, reflection, refraction, and diffuse scattering that Great-X reimplements inside Unreal Engine."},{"cited_title":"Unreal engine official website","cited_arxiv_id":null,"evidence_quote":"Identifies Unreal Engine as the single engine in which the platform is built and through which high-fidelity geometry and rendering are provided."},{"cited_title":"DeepVerse 6G: A dataset generation framework for multi-modal sensing and communica- tion digital twins,","cited_arxiv_id":null,"evidence_quote":"Represents the closed-source, dual-engine multimodal ISAC dataset generation framework that Great-X aims to replace with an open single-engine architecture."},{"cited_title":"Synthsom: A synthetic intelligent multi-modal sensing-communication dataset for synesthesia of machines (som),","cited_arxiv_id":null,"evidence_quote":"Represents the multi-engine heterogeneous simulation paradigm, coupling commercial ray tracing with AirSim, which the authors identify as limiting openness and extensibility."},{"cited_title":"Recommendation itu-r p.2040-3: Effects of building materials and structures on radiowave propagation above about 100 mhz,","cited_arxiv_id":null,"evidence_quote":"Provides the ITU-R material parameter definitions adopted in the channel model for the reflection and scattering calculations."}],"review_version":1}