Pith. sign in

REVIEW 4 major objections 5 minor 29 references

A signal-level indoor mmWave radar twin can recover most geometry-supported responses and turn the rest into diagnosable causes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 17:35 UTC pith:H2UPZKBK

load-bearing objection Useful systems paper: shared FMCW receive-channel twin plus honest residual ledger in one office—not a calibrated EM model, and the authors mostly say so. the 4 major comments →

arxiv 2607.28108 v1 pith:H2UPZKBK submitted 2026-07-30 cs.CV

mmRadarTwin: A Measurement-Calibrated Signal-Level Digital Twin Platform for Indoor mmWave Radar

Jianyi Zhou , Chenghao Zhang , Yanli Li , Dong Yuan This is my paper
classification cs.CV
keywords mmWave radar sensingdigital twinssignal-level simulationUnreal Enginepath attributionrange-angle processingindoor multipathFMCW radar
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Indoor millimeter-wave radar maps are hard to reproduce: the same room can look different from nearby poses because geometry, materials, multipath, hardware conventions, and processing all shape the measured range-angle field. Most digital-twin and ray-tracing tools stop at channels or paths, not at the complex receive signals a real FMCW radar processes. mmRadarTwin links a commodity radar measurement branch to an Unreal Engine scene branch through one shared receive-channel and range-angle pipeline, and keeps per-path records that name the actor, material, bounce event, and output bin behind each simulated return. In one office, across 154 poses at 22 locations, a physics-only path simulator recalled 70.8% of measurement-active geometry-supported response regions in the central field of view, while routing leftovers to missing path support, weak materials, unsupported peaks, geometric shifts, or missing physics. The point is not perfect radar-map cloning or cross-room generalization; it is a reusable workflow for building, comparing, and diagnosing indoor radar twins so calibration targets the real cause.

Core claim

mmRadarTwin shows that a measurement-calibrated, signal-level digital twin—shared complex receive grids, identical range-angle processing, and path-attributed contribution records—can recover a substantial share of geometry-supported indoor mmWave responses (70.8% region recall in the central FOV over 154 poses) and convert residuals into actionable discrepancy classes rather than opaque heatmap mismatch.

What carries the argument

Path-attributed residual diagnosis: each simulated return writes a complex multi-channel contribution plus metadata (actor, material tag, propagation event, range-angle bin), so residuals in the shared range-angle domain route to five mutually exclusive classes—missing path support, weak supported path, unsupported anchor, shifted response, or missing physical mechanism—each mapped to a specific calibration action or model-limit tag.

Load-bearing premise

The load-bearing premise is that a high-frequency bounce-ray model plus hand-set material tags on reconstructed room geometry is good enough to explain the main indoor radar map once both sides share the same processing chain.

What would settle it

Rebuild the same office twin with independent material measurements or full-wave checks, or move to a second room with the same workflow: if geometry-supported recall collapses and residual classes no longer map to repairable causes, the diagnostic contract fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Radar digital twins should expose complex receive-channel grids and path records, not only path-loss or CSI, if the goal is sensing comparison.
  • Material tuning is only valid when a supported path exists but is weak; other mismatches need geometry, pose, beamforming, or system-response fixes.
  • Scene-grounded region recall is a fairer success metric than global Top-K peaks when dominant returns flip with heading.
  • Unmatched structural returns should stay labeled as missing mechanisms (diffraction, diffuse scatter, floor-ceiling paths) rather than be hidden by free residual correction.
  • The same platform components can be reused for a new indoor scene by swapping actors, poses, and material tags under fixed hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Coupling this residual ledger to a downstream detector or tracker would test whether fixing C1–C4 errors improves real perception, not just map similarity.
  • The large C4 (shifted) and C5 (missing mechanism) shares suggest geometry registration and diffraction/diffuse models may unlock more recall than further RCS sweeps alone.
  • A learned refinement head on top of this physics twin could be constrained to residual classes the path basis cannot explain, keeping sim-to-real edits interpretable.
  • Multi-room rebuilds would reveal whether the five-class ledger is stable taxonomy or office-specific.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. mmRadarTwin is a systems platform that couples a commodity monostatic FMCW mmWave radar measurement branch to an Unreal Engine 5 scene-simulation branch through a shared multi-channel receive-grid and range-angle FFT interface. The simulator uses shooting-and-bouncing rays with per-actor material tags, writes complex receive-channel grids, and exports path-attributed contribution records (actor, material, bounce event, bin support). Residuals in the shared RA domain are routed into five mutually exclusive discrepancy classes (C1–C5) that map to scene repair, material/RCS update, system-response handling, pose/beamforming correction, or missing-mechanism tags. In one office deployment (TI AWR2243, 154 poses at 22 locations), the physics-only simulator recalls 70.8% of measurement-active geometry-supported response regions in the central FOV (near/mid/far 82.8/60.7/38.4%), with median matched-region amplitude error 1.01 dB, while reporting substantial Q1/Q2 mismatch mass and residual routing dominated by shifted responses, weak paths, and missing mechanisms. The paper explicitly does not claim full map reconstruction or cross-room generalization.

Significance. If the workflow holds, the paper fills a real gap between wireless ray-tracing digital twins (channel/path-loss outputs) and radar sensing, which needs complex receive products comparable to FMCW ADC chains. The shared-domain interface, complex-convention alignment, and path-contribution primitive are concrete systems contributions that make sim-to-measurement residuals diagnosable rather than purely visual. Careful non-claims, range-stratified recall, Q1/Q2 energy, and C1–C5 routing are strengths relative to heatmap-similarity papers. The work is primarily a platform and evaluation methodology contribution for indoor radar twins, not a new propagation theory result; its value depends on whether residual attribution is reproducible and action-guiding beyond this tuned office deployment.

major comments (4)
  1. [§5.2, Table 2; §4.2; Table 4] Table 2 and §5.2 state that BaseRCS/DiffBack/Mix/etc. are “deployment-fixed simulator parameters, not independently measured material constants,” but the manuscript never states whether these scalars were fixed a priori (e.g., from literature or a held-out calibration set) or adjusted while inspecting the same office RA maps used in Table 4. Because C2 is defined as the only admissible material/RCS-update class (§4.2) and residual routing is a central claim (R3/R5, §6.5), the diagnostic contract is load-bearing on this point. Please report the fitting protocol: freeze Table 2 before evaluation, or provide a held-out pose/location split and show that C2 vs C5 assignments are stable under that freeze.
  2. [§4.2; §6.1–6.4; Table 4] Matched-region recall (70.8%, Table 4) is defined on “geometry-supported expected response regions” where reconstructed geometry, pose, and expected object extent jointly support a strong structural response (§4.2, §6.1–6.2). No annotation protocol, inter-annotator check, automatic region proposal rule, or count of regions per pose/range band is given. Without that, the headline metric cannot be independently reproduced and may preferentially score large planar reflectors that SBR handles well. Specify how regions are proposed and accepted/rejected, how many regions enter the denominator by range band, and whether region labels were frozen before simulator tuning.
  3. [§4.4; §6.5; R5 in §2.3] §4.4 and R5 claim that C1–C5 route residuals to actionable calibration or limitation tags, but the evaluation never closes the loop: there is no before/after experiment showing that applying the designated action (scene repair, material update, pose correction, etc.) reduces the corresponding residual class on held-out poses. Without at least one controlled repair case per major class (especially C2 vs C4 vs C5), the taxonomy remains a descriptive ledger rather than a validated diagnostic workflow. A small closed-loop study would substantially strengthen the central systems claim.
  4. [§6.1; §4.2 Class 2; Table 4 Q3] §6.1 independently peak-normalizes each pose to 0 dB before comparison. Under that convention, region recall mainly tests geometric placement of surviving paths once absolute scale is free; Q3’s 1.01 dB median residual is conditional on both branches already placing strong energy in the same region. Absolute or single global-gain calibration is not reported, yet C2 (weak supported path) is an amplitude-class decision. Clarify how peak normalization interacts with C2 thresholding, and report at least one globally gain-aligned or unnormalized amplitude comparison so readers can separate placement success from radiometric fidelity.
minor comments (5)
  1. [Table 4; §6.1] Top-K peak agreement (~0.26) is appropriately demoted to an auxiliary metric, but Table 4 would be clearer if the denominator and matching rule for Top-K were stated in the table caption, not only in §6.1.
  2. [§3.4] Eq. (5) uses a specific conjugation convention; a one-sentence note on how convention mismatch was detected on real ADC data (phase-slope sign check) would help implementers.
  3. [§6.2–6.3] Figure 5–6 boxes are helpful but the main text should state whether white/magenta boxes are exhaustive for those poses or illustrative only.
  4. [§7] Related work cites Sionna RT and RF Genesis appropriately; a brief explicit contrast table (output domain: CIR vs RC grid vs spectrogram; attribution: yes/no) would sharpen positioning.
  5. [Abstract; §5.2] Typos/style: “mm-RadarTwin” hyphenation is inconsistent in the abstract vs title; “paper-facing material tags” (§5.2) is unclear wording.

Circularity Check

0 steps flagged

No circular derivation: 70.8% recall is an empirical sim-vs-measurement comparison, not a quantity forced by its own inputs.

full rationale

mmRadarTwin is a systems/platform paper, not a first-principles derivation. The load-bearing quantitative claim (70.8% matched-region recall over 154 poses; Table 4, §6.4) compares an independent real FMCW measurement branch to a forward SBR simulator after a shared FFT chain (§3.4, §5.4). That match rate is not defined in terms of itself, not algebraically forced by the receive-grid equations (Eqs. 1–8), and not justified by a self-citation uniqueness theorem. Material-tag scalars (Table 2) are explicitly labeled deployment-fixed simulator parameters rather than predicted EM constants, and the paper does not present those scalars—or the C1–C5 taxonomy—as predictions derived from data. Peak-normalization and author-specified geometry-supported regions (§4.2, §6.1) weaken what the metric can claim about absolute calibration and annotation independence; those are evaluation-validity limits, not circular reductions of a claimed derivation to its inputs. Residual routing is an operational diagnostic workflow, not a renamed known law. No step reduces Eq. X to Eq. Y by construction or fits a parameter and re-reports it as an independent prediction. Score 0 is the proportionate finding.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central systems claim rests on standard FMCW/virtual-array signal processing and high-frequency ray optics, plus deployment choices (hand-set material scalars, reconstructed meshes, match tolerances, residual taxonomy) that are not derived from first principles. No new physical particles or forces; invented structure is methodological (contribution records, C1–C5 classes, shared RC/RA contract). Free parameters materially affect amplitude and residual routing; geometry fidelity and path-basis completeness are load-bearing domain assumptions.

free parameters (5)
  • Material-tag scalars (BaseRCS, BackExp, Mix, DiffBack, NormExp) for Metal/Concrete/Wood/TableTop/Carpet/Glass/Plastic = e.g. Metal BaseRCS=-3.0 dB; Carpet=-25.0 dB; others per Table 2
    Table 2 values are deployment-fixed simulator parameters, explicitly not independently measured EM constants; they control specular/diffuse strength and thus Class-2 residuals and matched amplitudes.
  • Strong/medium/weak RA thresholds = -15 / -20 / -25 dB
    Evaluation cells classified at -15/-20/-25 dB after per-pose peak normalization; directly defines match/mismatch mass.
  • Region and peak match tolerances = ±0.20 m / ±5° (region); 0.35 m / 10° (peaks)
    Matched-region recall uses ±0.20 m or one range bin and ±5° or one angle bin; Top-K uses 0.35 m and 10°—author-chosen windows that gate the headline metric.
  • SBR tracing hyperparameters (bounce depth, path-power cutoff, ray budget) = fixed but unspecified numerically in text
    Fixed across poses but numeric values not fully reported; determine path support and Class-1 vs Class-5 outcomes.
  • Per-pose peak normalization and [-50,0] dB clip = peak→0 dB; clip [-50,0] dB
    Removes absolute power calibration; comparison is relative structure only, affecting all reported dB residuals.
axioms (6)
  • domain assumption High-frequency path-basis (SBR) coherent superposition adequately represents dominant indoor mmWave receive-channel contributions for RA comparison.
    §3.3 explicitly adopts path basis over full-wave solution; diffraction, dispersive materials, and some rough-surface/elevation mechanisms are deferred.
  • domain assumption Flat-smooth materials follow Fresnel-like reflection; surfaces above λ/8 RMS use a diffuse-backscatter threshold model.
    §3.3 and §5.2 cite Fresnel-dominant and Rayleigh-roughness regimes; library collapses real materials into seven tags.
  • domain assumption LiDAR SLAM + mobile photogrammetry meshes with collision proxies are accurate enough for path extraction and pose alignment to the radar frame.
    §5.2 scene reconstruction; centimeter-level Polycam claims are cited from prior work, not re-validated as radar-phase-accurate here.
  • domain assumption Identical fast-time and virtual-array FFTs plus one fixed complex baseband convention make measured vs simulated RA residuals attributable to scene/physics rather than processing asymmetry.
    R1 and §3.4; convention validation is magnitude-preserving conjugate check, not full hardware-chain identity.
  • standard math Standard FMCW beat-frequency and virtual-array phase model (Eqs. 1–5) with TDM-MIMO M=8.
    Classical radar signal model; deployment-specific chirp numbers in §5.1.
  • ad hoc to paper Five mutually exclusive residual classes with lowest-number priority suffice as calibration decision categories for this office deployment.
    §4.1–4.2; authors state the taxonomy is extensible and not exhaustive of all mmWave residuals.
invented entities (3)
  • mmRadarTwin shared receive-channel / range-angle digital-twin interface independent evidence
    purpose: Force measurement and UE simulation branches into one comparable complex signal product for indoor monostatic FMCW radar.
    Systems construct combining known RT and DSP pieces; evaluated empirically in one office.
  • Per-path complex contribution record primitive (actor, material, event, bin support) no independent evidence
    purpose: Attribute RA residuals to named scene entities and path subsets rather than aggregate heatmap error.
    Methodological primitive (§3.5); consistency invariant Eq. 8 checked at machine precision on validation poses only.
  • C1–C5 residual attribution taxonomy no independent evidence
    purpose: Map sim-measurement disagreements to calibration actions or mechanism-extension tags.
    Operational categories defined by the paper’s match-window cues; not an externally standardized fault taxonomy.

pith-pipeline@v1.2.0-daily-grok45 · 22270 in / 4517 out tokens · 99893 ms · 2026-07-31T17:35:56.457279+00:00 · methodology

0 comments
read the original abstract

Indoor mmWave radar perception is difficult to reproduce because measured range-angle responses depend on scene geometry, material response, multipath, hardware conventions, and signal processing. Existing ray-tracing and digital-twin tools often expose rendering, channel, or path-level quantities, while radar sensing requires complex signal products that can be processed and compared in the same domain as real FMCW measurements. We present mmRadarTwin, a signal-level and path-attributed digital-twin platform for indoor mmWave radar. mmRadarTwin links a real radar measurement branch with an Unreal Engine scene-simulation branch through a shared receive-channel and range-angle processing interface. The simulator writes complex multi-channel receive grids and exports per-path contribution records that identify the actor, material tag, propagation event, and output-bin support of each simulated return. We evaluate mmRadarTwin in an office deployment using a commodity monostatic mmWave radar and mobile scene-capture hardware. Across 154 measured poses spanning 22 radar locations, the current physics-only path-basis simulator recalls 70.8% of measurement-active geometry-supported response regions in the central usable field of view while exposing residuals caused by weak or missing path support, shifted responses, unsupported anchors, and missing physical mechanisms. Rather than claiming complete radar-map reconstruction or cross-room generalization, mmRadarTwin establishes a practical systems workflow for constructing, comparing, and diagnosing indoor radar digital twins.

Figures

Figures reproduced from arXiv: 2607.28108 by Chenghao Zhang, Dong Yuan, Jianyi Zhou, Yanli Li.

Figure 1
Figure 1. Figure 1: Motivation for mmRadarTwin. Standard game [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The system overview for mmRadarTwin contains real-world capture, signal-level radar simulation, shared comparison, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hardware rig and TDM-MIMO antenna layout. The [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Acquisition coverage in the reconstructed indoor [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Scene-grounded real-to-sim comparison for one representative pose. Column (a) shows the real scene and the measured [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative match and mismatch examples. Each row compares the measured radar response (left) with the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pose-dependent peak dominance in the measured [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 1 linked inside Pith

  1. [1]

    David Borts, Erich Liang, Tim Broedermann, Andrea Ramazzina, Stefanie Walz, Edoardo Palladin, Jipeng Sun, David Brueggemann, Christos Sakaridis, Luc Van Gool, Mario Bijelic, and Felix Heide. 2024. Radar fields: frequency-space neural scene representations for FMCW radar. InACM SIGGRAPH (SIGGRAPH ’24). ACM SIGGRAPH, New York, NY, USA, 1–10

    2024

  2. [2]

    Xingyu Chen, Zihao Feng, Ke Sun, Kun Qian, and Xinyu Zhang. 2024. RFCanvas: Modeling RF Channel by Fusing Visual Priors and Few-shot RF Measurements. In Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems (Hangzhou, China)(SenSys ’24). Association for Computing Machinery, New York, NY, USA, 464–477. doi:10.1145/3666025.3699351

    arXiv 2024

  3. [3]

    Xingyu Chen and Xinyu Zhang. 2024. RF Genesis: Zero-Shot Generalization of mmWave Sensing through Simulation-Based Data Synthesis and Generative Diffusion Models. InProceedings of the 21st ACM Conference on Embedded Net- worked Sensor Systems(Istanbul, Turkiye)(SenSys ’23). Association for Computing Machinery, New York, NY, USA, 28–42. doi:10.1145/3625687.3625798

    arXiv 2024

  4. [4]

    Epic Games. 2022. Epic Games releases Unreal Engine 5 for all cre- ators. https://www.epicgames.com/site/en-US/news/epic-games-releases- unreal-engine-5-for-all-creators. Accessed 2026-05-26

    2022

  5. [5]

    Xueqiang Han, Tianyue Zheng, and Jun Luo. 2026. Stereo-Fi: Free-Form 3D Reconstruction via Generatively Co-Trained Inverse RF Rendering. InProceedings of the 2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys ’26). Association for Computing Machinery, New York, NY, USA, 334–346. doi:10.1145/3774906.3802762

    arXiv 2026

  6. [6]

    Danping He, Ke Guan, Bo Ai, Zhangdui Zhong, Junhyeong Kim, Heesang Chung, and Andrej Hrovat. 2023. Channel Measurement and Ray-Tracing Simulation for 77 GHz Automotive Radar.IEEE Transactions on Intelligent Transportation Systems24, 7 (2023), 7746–7756. doi:10.1109/TITS.2022.3208008

    arXiv 2023

  7. [7]

    Jakob Hoydis, Fayçal Aït Aoudia, Sebastian Cammerer, Florian Euchner, Merlin Nimier-David, Stephan Ten Brink, and Alexander Keller. 2024. Learning radio en- vironments by differentiable ray tracing.IEEE Transactions on Machine Learning in Communications and Networking2 (2024), 1527–1539

    2024

  8. [8]

    Jakob Hoydis, Faycal Ait Aoudia, Sebastian Cammerer, Merlin Nimier-David, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. 2023. Sionna RT: Dif- ferentiable Ray Tracing for Radio Propagation Modeling. In2023 IEEE Globecom Workshops (GC Wkshps). 317–321. doi:10.1109/GCWkshps58843.2023.10465179

    arXiv 2023

  9. [9]

    Agrawal, Richard Newcombe, and Ahmed Alkhateeb

    Shuaifeng Jiang, Qi Qu, Xiaqing Pan, Abhishek K. Agrawal, Richard Newcombe, and Ahmed Alkhateeb. 2025. Learnable wireless digital twins: reconstructing electromagnetic field with neural representations.IEEE Open Journal of the Communications Society6 (2025), 1568–1590

    2025

  10. [10]

    Hao Kong, Cheng Huang, Jiadi Yu, and Xuemin Shen. 2025. A Survey of mmWave Radar-Based Sensing in Autonomous Vehicles, Smart Homes and Industry.IEEE Communications Surveys & Tutorials27, 1 (2025), 463–508

    2025

  11. [11]

    Zhongzheng Lai, Dong Yuan, Huaming Chen, Yu Zhang, and Wei Bao. 2023. WirelessDT: A Digital Twin Platform for Real- Time Evaluation of Wire- less Software Applications. In2023 IEEE/ACM 45th International Conference on Software Engineering: Companion Proceedings (ICSE-Companion). 146–150. doi:10.1109/ICSE-Companion58688.2023.00043

    arXiv 2023

  12. [12]

    Wenbin Li, Danping He, Ke Guan, Xiangyu Shi, and Zhangdui Zhong. 2022. Millimeter-wave radar measurement and ray-tracing simulation for urban street environment. In2022 3rd URSI Atlantic and Asia Pacific Radio Science Meeting (AT-AP-RASC). IEEE, Gran Canaria, Spain, 1–4

    2022

  13. [13]

    Yuan Liu, Moein Ahmadi, Johann Fuchs, Mohammad Alaee-Kerahroodi, and M. R. Bhavani Shankar. 2025. Dynamic Indoor mmWave MIMO Radar Simula- tion: An Image Rendering-Based Approach.IEEE Transactions on Antennas and Propagation73, 4 (2025), 1984–1999. doi:10.1109/TAP.2024.3412269

    arXiv 2025

  14. [15]

    Haofan Lu, Christopher Vattheuer, Baharan Mirzasoleiman, and Omid Abari

  15. [16]

    Tribhuvanesh Orekondy, Pratik Kumar, Shreya Kadambi, Hao Ye, Joseph Soriaga, and Arash Behboodi. 2023. WiNeRT: Towards Neural Ray Tracing for Wireless Channel Modelling and Differentiable Simulations. InThe Eleventh International Conference on Learning Representations. OpenReview.net, Kigali, Rwanda. https: //openreview.net/forum?id=tPKKXeW33YU

    2023

  16. [17]

    Sujeet Milind Patole, Murat Torlak, Dan Wang, and Murtaza Ali. 2017. Auto- motive radars: A review of signal processing techniques.IEEE Signal Processing Magazine34, 2 (2017), 22–35. doi:10.1109/MSP.2016.2628914

    arXiv 2017

  17. [18]

    Clement Ruah, Osvaldo Simeone, Jakob Hoydis, and Bashir Al-Hashimi. 2024. Calibrating Wireless Ray Tracing for Digital Twinning Using Local Phase Er- ror Estimates.IEEE Transactions on Machine Learning in Communications and Networking2 (2024), 1193–1215. doi:10.1109/TMLCN.2024.3448391

    arXiv 2024

  18. [19]

    Batool Salehi, Utku Demir, Debashri Roy, Suyash Pradhan, Jennifer Dy, Stratis Ioannidis, and Kaushik Chowdhury. 2024. Multiverse at the Edge: Interacting Real World and Digital Twins for Wireless Beamforming.IEEE/ACM Transactions on Networking32, 4 (2024), 3092–3110. doi:10.1109/TNET.2024.3377114

    arXiv 2024

  19. [20]

    Emerson Sie, Xinyu Wu, Heyu Guo, and Deepak Vasisht. 2024. Radarize: Enhanc- ing Radar SLAM with Generalizable Doppler-Based Odometry. InProceedings of the 22nd Annual International Conference on Mobile Systems, Applications and Services(Minato-ku, Tokyo, Japan)(MOBISYS ’24). Association for Computing Machinery, New York, NY, USA, 331–344. doi:10.1145/364...

    arXiv 2024

  20. [21]

    Xinghua Sun and Akshay Gadre. 2026. Towards Practical Bi-Static Polarimetric Imaging Using Commodity mmWave Radars for Material Sensing. InProceedings of the 2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys ’26). Association for Computing Machinery, New York, NY, USA, 172–185. doi:10.1145/3774906.3800470

    arXiv 2026

  21. [22]

    The track seen with Polycam

    Federica Vacatello and Alessandro Melega. 2026. “The track seen with Polycam”: a comparative study of SfM-based 3D documentation applied to architectural structures and decorative motifs at the Holy Sepulchre (Jerusalem).Digital Applications in Archaeology and Cultural Heritage40 (2026), e00480. doi:10.1016/ j.daach.2025.e00480

    2026

  22. [23]

    Xiyu Wang, Danping He, Ke Guan, Hongyu Duan, Jianwu Dou, Najah Abed AbuAli, and Zhangdui Zhong. 2024. Channel Measurement and Modeling for Millimeter-Wave Automotive Radar.IEEE Transactions on Intelligent Transporta- tion Systems25, 3 (2024), 2933–2943

    2024

  23. [24]

    Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. 2022. FAST-LIO2: Fast Direct LiDAR-Inertial Odometry.IEEE Transactions on Robotics38, 4 (2022), 2053–2073. doi:10.1109/TRO.2022.3141876

    arXiv 2022

  24. [25]

    Iskander

    Zhengqing Yun and Magdy F. Iskander. 2015. Ray Tracing for Radio Propagation Modeling: Principles and Applications.IEEE Access3 (2015), 1089–1100. doi:10. 1109/ACCESS.2015.2453991

    arXiv 2015

  25. [26]

    Lihao Zhang, Haijian Sun, Samuel Berweger, Camillo Gentile, and Rose Qingyang Hu. 2026. RF-3DGS: Wireless Channel Modeling With Radio Radiance Field and 3D Gaussian Splatting.IEEE Transactions on Wireless Communications 25 (2026), 10419–10433. doi:10.1109/TWC.2026.3652154

    arXiv 2026

  26. [27]

    2023.WiSegRT: Dataset for Site-Specific Indoor Radio Propagation Modeling with 3D Segmentation and Differentiable Ray-Tracing

    Lihao Zhang, Haijian Sun, Jin Sun, and Rose Qingyang Hu. 2023.WiSegRT: Dataset for Site-Specific Indoor Radio Propagation Modeling with 3D Segmentation and Differentiable Ray-Tracing. Technical Report arXiv:2312.11245. arXiv

    Pith/arXiv arXiv 2023

  27. [28]

    Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. 2023. NeRF2: Neural Radio-Frequency Radiance Fields. InProceedings of the 29th Annual International Conference on Mobile Computing and Networking. ACM MobiCom, Madrid, Spain, 1–15

    2023

  28. [29]

    Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. 2024. NeRF2.GetMobile: Mobile Computing and Communications28, 1 (2024), 29–34

    2024

  29. [2024]

    InProceedings of the 41st International Conference on Machine Learning (ICML’24)

    NeWRF: a deep learning framework for wireless radiation field reconstruc- tion and channel prediction. InProceedings of the 41st International Conference on Machine Learning (ICML’24). ICML, Vienna, Austria, 33147–33159