HybridSim: A Physics-Learning Hybrid Digital Twin for mmWave Human Sensing

Hongfei Xue; Kok Chung Chua; Peng Li; Pu Wang; Tianyu Liu; Toa Chean Khim; Weitao Xiong

REVIEW 4 major objections 5 minor 36 references

HybridSim: A Physics-Learning Hybrid Digital Twin for mmWave Human Sensing

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read HybridSim claims that splitting mmWave radar simulation into a physics-based direct path and a learned multipath path yields synthetic training data that transfers to real human activity recognition far better than existing simulators.

desk verdict Decoupled direct/indirect mmWave synthesis is a genuinely useful idea, but the headline HAR number is inflated by a same-identity training protocol. read the letter →

arxiv 2607.15806 v1 pith:AFZ55TQO submitted 2026-07-17 cs.CV

Weitao Xiong , Tianyu Liu , Peng Li , Kok Chung Chua , Toa Chean Khim , Pu Wang , Hongfei Xue This is my paper

classification cs.CV

keywords mmWaveradarsimulationhumanactivityrecognitionRange-Dopplerheatmapmicro-Doppler3DGaussiansplattinginverserenderingBRDFsim-to-realtransfer

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

HybridSim sets out to show that a practical mmWave radar simulator can generate realistic training data for human activity recognition from animated 3D human meshes and a simple description of a fixed room. Its core bet is that radar propagation should be split into two parts: a physically modeled direct path (single-bounce reflections, with amplitude learned and phase computed from actual delays) and an indirect path (all multi-bounce reflections) approximated by a learned surrogate using 3D Gaussian splatting and virtual receivers on the walls. The paper reports that this decoupling preserves fine-grained micro-Doppler details in the synthesized Range-Doppler heatmaps, and that a classifier trained only on the simulator's output reaches about 92 percent accuracy on real radar measurements, a large jump over a leading physical augmenter. If correct, this would make site-specific radar training data cheap to produce at about one second per frame, without expensive ray tracing.

What carries the argument

The load-bearing mechanism is the decoupled two-path signal expression: the direct return is a coherent sum of single-bounce scatterers with a learned BRDF amplitude and a physically computed phase from propagation delay; the indirect return is a coherent sum of three-segment paths through a proxy virtual receiver on each room boundary, with the wall reflection and scatterer response collapsed into a learned directional amplitude coming from 3D Gaussian splatting. The proxy-receiver geometry is the trick that makes multi-bounce effects differentiable and cheap: instead of tracing rays, the network learns spatial multipath texture as splat parameters, while the delay term still enforces that

What would settle it

Record real radar heatmaps in a fixed room, train HybridSim, then move a large reflective panel to a new position while keeping all learned splat parameters frozen; if the synthetic Range-Doppler interference pattern does not shift in the way the real measurements shift, the learned multipath surrogate is memorizing site-specific textures rather than modeling the physics.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the key to high-fidelity synthetic mmWave signals is explicitly disentangling the direct surface reflection from environmental multipath. The direct path models the human body and walls with a microfacet BRDF whose amplitude parameters are learned but whose phase follows from physical propagation delay; the indirect path replaces explicit multi-bounce ray tracing with a differentiable three-segment model that routes energy from each scatterer to a proxy virtual receiver on each wall and then to the physical receiver, with 3D Gaussian splatting supplying learned directional amplitudes. The paper validates the design by training a downstream activity-cla

Load-bearing premise

Everything rests on the assumption that a three-segment path through a proxy receiver on each wall, with the wall's reflection and the scatterer's response squeezed into one learned amplitude, is enough to reproduce the true multi-bounce interference a real room produces; if the real effect depends on reflection phase, polarization, or frequency-dependent absorption that this surrogate omits, the claimed realism for a new scene is not supported.

Editorial extensions

If this is right

A classifier trained purely on HybridSim-generated heatmaps can reach about 92 percent accuracy on real radar measurements in the fixed-room setting, against about 54 percent for the leading physical augmenter, suggesting the synthetic data carries the motion semantics needed for sim-to-real transfer.
Because synthesis runs in under a second per frame, generating large labeled radar datasets for a site from animated meshes becomes practical on a single GPU.
The explicit decoupling means the direct path alone carries most of the kinematic information, while the indirect path adds site-specific multipath; removing the indirect path drops accuracy by about three percentage points, so both components contribute.
Learnable background noise is a necessary component: without noise injection accuracy falls below 73 percent, and with fixed noise it rises to about 90 percent, indicating that matching the noise floor is part of the sim-to-real gain.
The same architecture adapts to a second room layout by fine-tuning, with about 88.7 percent downstream accuracy, showing site-specific retraining is feasible.

Reading between the lines

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

A natural next test is whether the proxy-receiver surrogate transfers to rooms with non-planar clutter; the paper's proxy walls absorb unseen objects into learned splat parameters, so shifting furniture without retraining would reveal whether that absorption is a feature or a limitation.
The large gap between global image metrics (roughly a 10 percent PSNR gain) and downstream accuracy (roughly a 38-point gain) suggests that Range-Doppler similarity metrics undervalue micro-Doppler content; we infer that future simulators should be benchmarked on downstream tasks first, not just on heatmap matching.
We infer the direct-path model is target-agnostic: any articulated mesh, human or otherwise, could be plugged in without changing the equations, so the same decoupled design may generalize to animal or robot motion sensing.
The learnable-noise result implies that synthetic radar training sets should be generated with a noise model calibrated to the target sensor; we infer that matching the background distribution can matter more than adding additional clean simulation frames.

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

The genuinely new thing here is the decoupling itself: modeling the direct human reflection with an inverse-rendering microfacet BRDF and the indirect environmental multipath with 3D Gaussian Splatting plus proxy wall receivers. That split is physically sensible, the equations (Eqs. 1–3) are clearly stated, and the phase is geometry-derived rather than fitted, which is a real strength. The ablation in Table 3 supports the core claim that decoupling helps, and the GCN-as-training-regularizer detail is a thoughtful touch. I believe the authors are onto something for site-specific radar simulation where full ray tracing is too expensive.

The soft spot is load-bearing and it is in the evaluation protocol, not the modeling. The downstream HAR experiment trains the classifier on synthetic data generated from an unseen subject's SMPL meshes and tests on that same subject's real radar recordings. But those meshes come from mmMesh, which reconstructs them from the very same radar recordings used as test ground truth. So the training set contains the test identity and essentially the same motions. This is a same-subject cross-modal transfer test, not a held-out sim-to-real evaluation. The 92.07% versus 54.22% comparison therefore cannot be attributed to simulator fidelity; it may largely reflect identity leakage. The supplementary second subject at 97.02% repeats the same design, so it does not fix the problem. The intra-subject synthesis evaluation is also on the training subject's remaining frames, and no error bars are reported anywhere.

The indirect-path proxy receiver approximation (collapsing wall reflection and scatterer response into a single learned directional term) is a softer concern. It ignores phase and polarization effects, but the kinematic Doppler trend is still enforced through the geometric delay, so the surrogate can be useful even if physically incomplete. The more immediate issue is that the headline quantitative evidence is not an honest held-out test. The paper itself concedes per-scene fine-tuning is needed, which limits generality, but that is stated clearly.

Who is this for? People working on RF simulation, neural rendering, and synthetic data for radar perception. The decoupling idea deserves to be tried by others. But the evaluation needs major revision before the claims can be trusted: a truly held-out subject whose meshes are never used to generate classifier training data, proper error bars, and ideally code or data release. I would send it to peer review with this as the central required change, not desk-reject it. The core architecture is interesting enough to warrant the referee time.

Referee Report

4 major / 5 minor

Summary. HybridSim proposes a hybrid physics-learning simulator that synthesizes mmWave radar Range-Doppler heatmaps from dynamic human meshes in a fixed indoor scene. The method decouples the signal into a direct single-bounce path, modeled by inverse rendering with a microfacet BRDF, and an indirect multi-bounce path, modeled by 3D Gaussian Splatting with proxy virtual receivers on room boundaries. The framework is optimized end-to-end against real RD heatmaps using a dB-domain MSE loss. Experiments report improved synthesis metrics over mmGPE and RF-Genesis, and downstream human activity recognition accuracy of 92.07% when a classifier is trained on HybridSim-generated data and tested on real measurements. The paper claims that the decoupling preserves kinematic micro-Doppler details and that the simulator is a practical site-specific data augmentation tool.

Significance. If the claims hold, HybridSim offers a computationally efficient alternative to full ray tracing for site-specific mmWave radar simulation, with a clearly stated direct/indirect decomposition and supporting ablations. The direct-path phase is geometry-derived (Eqs. 1-2), the dB-domain loss is well motivated, and the reported per-frame runtime under 1 second is attractive. However, the headline downstream result is undermined by the evaluation protocol, and the indirect-path surrogate is validated only indirectly. The paper does not release code, and the main quantitative claims lack error bars, which limits confidence. The core idea is plausible and worth pursuing, but the evidence as presented is not sufficient to support the stated sim-to-real generalization claims.

major comments (4)

[Sec. 4.5, Table 2] The HAR evaluation is not a held-out sim-to-real test. The classifier is trained on synthetic data 'generated for an unseen subject' and tested on 'corresponding physical ground truth measurements.' Since mmMesh provides SMPL meshes reconstructed from the same radar recordings that serve as test ground truth, the synthetic training data are generated from the same subject's meshes and the same motions as the real test data. This is same-subject cross-modal transfer, not a test of simulator generalization. The 92.07% vs. 54.22% comparison therefore conflates simulator fidelity with identity/motion leakage. A valid protocol requires a held-out identity whose meshes are never used to synthesize classifier training data (e.g., train simulator on subject A, generate training data from subject B, test on subject C). Without this, the central quantitative claim is unsupported.
[Sec. 3.3, Eq. (3)] The proxy-receiver surrogate collapses the wall reflection and scatterer response into a single learned directional amplitude SH_{p,k}, while the phase is entirely determined by the geometric three-segment delay τ_{p,k}. Because the loss (Eq. 4) operates only on RD amplitude heatmaps, any reflection-dependent phase offset, polarization, or frequency-dependent absorption cannot be constrained by the optimization. The paper provides no direct evidence that this amplitude-only surrogate reproduces true multi-bounce interference. A concrete test would be to compare Eq. (3) against explicit two-bounce ray tracing in a scene where ground-truth multipath is known, or to measure phase coherence in a controlled environment. As it stands, the claim that the surrogate is 'physically defensible' is an assumption, not a demonstrated property.
[Sec. 4.3, Supp. Sec. 9] The main evaluation uses a single primary training subject and a single unseen test subject, with no error bars or repeated trials. The supplementary second-subject result (97.02% HAR) uses the same flawed protocol as Sec. 4.5, so it does not mitigate the circularity concern. For a paper whose central claim is cross-subject generalization and sim-to-real transfer, reporting variance over multiple subjects and repeated optimization runs is necessary. At minimum, the authors should state the number of independent runs and report mean ± std for Tables 1-4.
[Sec. 4.8] The 'site-specific second-scene adaptation' experiment uses a self-collected dataset with a single male subject and no comparison to the baselines used in the main paper. It reports 22.01 dB PSNR and 88.68% HAR accuracy, but without a baseline comparison or a ray-tracing reference in the same scene, it is difficult to attribute these numbers to the method's generality rather than to scene-specific fitting. This experiment should be expanded or repositioned as preliminary evidence, not as a full validation of site-specific adaptability.

minor comments (5)

[Fig. 2] The label 'mmWave Broad:' appears incomplete or typoed; likely should be 'mmWave Broadband' or similar. Also, the legend for 'Masked Feature' and 'Reconstruct Feature' is visually ambiguous.
[Sec. 3.4, Eq. (4)] The notation '20log10(·)' lacks a space and could be typeset as '20 log10(·)' for clarity. Additionally, the use of |F_RD(Sim)| implicitly assumes that the RD FFT is applied to the complex signal; this should be stated explicitly at first use.
[Table 1] The column header 'Unseen Frames (Intra-Subject)' is slightly confusing; it means held-out frames of the training subject, not 'unseen frames' in a semantic sense. Consider renaming to 'Held-out Frames'.
[Sec. 4.7, Fig. 10] The qualitative decomposition is helpful, but the caption mentions 'Action labels A1–A8 follow the action definitions in Sec. 10' while the action definitions are in Sec. 11 of the supplementary material. Fix the cross-reference.
[References] Reference [1] is formatted as an ACM book with a 2023 date, which may be incorrect for the SCAPE paper (originally 2005). Please verify all reference dates and venues.

Circularity Check

1 steps flagged · score 6.0 of 10

Partial circularity: the headline downstream HAR gain (92.07%) is forced by training the classifier on synthetic data generated from the same 'unseen' subject's meshes as the real test recordings, making it a same-identity mesh-to-radar transfer score rather than a held-out sim-to-real evaluation.

other [Sec. 4.5 (Downstream Task: Human Activity Recognition), Supplementary Sec. 9, Sec. 4.1 (dataset), Sec. 2.2 (SMPL input)]
"We implement our framework using the mmMesh dataset [22]... Under a sim-to-real evaluation protocol, we train the classifier exclusively on synthetic data generated for an unseen subject across eight action categories, and then test it directly on corresponding physical ground truth measurements. ... We evaluate one additional unseen male subject in the original scene. HybridSim achieves 22.37 dB PSNR, 0.262 SSIM, 0.092 LPIPS, and 97.02% accuracy / 0.9702 F1-score HAR performance."

The synthetic training data are produced by running HybridSim on the unseen subject's SMPL meshes. Since the paper uses the mmMesh dataset, whose meshes are reconstructed from the same mmWave radar recordings that supply the 'corresponding' real test heatmaps, the classifier's synthetic training set and the real test set share the same identity and action instances. Thus the reported gain (92.07% and 97.02%) is a same-subject mesh-to-radar transfer measurement, not a held-out-identity sim-to-real prediction. The target identity is effectively an input to the generation of the training data, so the result is forced by construction rather than by simulator generalization.

full rationale

The underlying signal equations (Eqs. 1-3) are not circular: the phase/delay terms are computed from mesh geometry and the propagation path, and the amplitudes (BRDF field, 3DGS SH coefficients, noise rate) are parameters optimized against ground-truth RD heatmaps. Optimizing a simulator and then evaluating on unseen poses of the training subject or on a different subject (Table 1) is a normal inverse-rendering/generalization test. The circularity is concentrated in the downstream sim-to-real HAR protocol of Sec. 4.5: the classifier is trained on HybridSim output generated 'for an unseen subject' and tested on 'corresponding physical ground truth measurements.' Because HybridSim's input is the subject's SMPL mesh, and the mmMesh dataset provides meshes reconstructed from the same mmWave recordings used as test GT, the synthetic training set already contains the test identity and test motions. The reported 92.07% (and 97.02% for the supplementary second subject) is therefore a same-identity mesh-to-radar transfer score, not a held-out-identity generalization result; it is partly forced by feeding the target subject into the training pipeline. The main synthesis-fidelity comparison (Table 1) and the internal ablations still provide independent evidence, but the headline quantitative claim of 'consistent gains on downstream radar-based human sensing tasks' is overstated as a sim-to-real augmentation result. No self-citation chain or imported uniqueness theorem is load-bearing; the issue is evaluation/protocol leakage, not equation-level circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The central contribution rests on learned material and multipath parameters fitted to the evaluation environment, plus a surrogate indirect-path model. The phase/Doppler backbone is physically derived, but the amplitude and noise components are fitted to the same mmMesh data used in evaluation.

free parameters (5)

Per-vertex BRDF material parameters (roughness α, complex refractive index η/k, metallic blending factor, polarization-m = not reported individually
Learned by the BRDF decoder during end-to-end training against real RD heatmaps (Sec. 3.2); these directly determine direct-path amplitude L_p.
Per-Gaussian 3DGS parameters (rotation, spherical harmonics, opacity, scaling, offset) = not reported
Learned to fit site-specific multipath for the indirect path (Sec. 3.3); they encode the surrogate reflection model.
Learnable noise-rate parameter = not reported
Scales injected zero-mean complex Gaussian noise to match background statistics; optimized as part of training (Sec. 3.4, Sec. 4.9).
Neural network weights (tri-plane encoder, GCN, BRDF decoder, 3DGS decoder) = not reported
All trainable parameters optimized by the RD-heatmap MSE loss.
Fixed noise rate r_fix = 0.025
Hand-chosen from empirical dataset statistics for the control experiment (Suppl. Sec. 12); not used in the full model but used for baseline comparisons.

assumptions (6)

domain assumption The SMPL mesh surface, with per-vertex material fields, is a sufficient geometric representation for radar scattering from a human body.
Used throughout Sec. 3.1-3.2 to attach BRDF parameters to mesh vertices and apply linear blend skinning.
domain assumption A microfacet BRDF with Trowbridge-Reitz NDF, Smith shadowing, and complex Fresnel equations models mmWave surface reflection amplitude.
Sec. 3.2 / Suppl. Sec. 7; adapted from optical rendering and prior radar inverse rendering [10,19] without validation against controlled material measurements.
ad hoc to paper The three-segment proxy-receiver sum (Eq. 3) is a valid, optimization-friendly surrogate for multi-bounce ray tracing.
Sec. 3.3; no derivation or error bound connects the surrogate to true multi-bounce propagation; its validity rests on the empirical ablations only.
domain assumption Hidden Point Removal from the transceiver position determines line-of-sight visibility for radar scattering paths.
Sec. 3.1 / Suppl. Sec. 6; a single-viewpoint convex-hull test is an approximation of round-trip occlusion for a non-monostatic array.
domain assumption Radar absolute phase can be discarded; amplitude-only loss on Range-Doppler heatmaps is sufficient to learn coherent multipath and micro-Doppler.
Sec. 3.4; justified by sensitivity of phase to sub-millimeter errors, but this also prevents learning reflection-dependent phase offsets.
ad hoc to paper Injected learnable complex Gaussian noise can represent the real sensor/hardware noise floor.
Sec. 3.4 and Sec. 4.9; no physical noise model, only a statistical fit.

invented entities (1)

Proxy virtual receivers placed on room boundaries
purpose: Convert multi-bounce reflections into differentiable three-segment paths (TX -> scatterer -> wall receiver -> RX) for indirect-path simulation.
They are a computational construct introduced in Sec. 3.3; they have no physical existence and no falsifiable prediction outside the fitted RD heatmaps.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HybridSim: A Physics-Learning Hybrid Digital Twin for mmWave Human Sensing." pith.science (2026). https://pith.science/paper/AFZ55TQO

@misc{pith2026260715806,
  author       = {Pith},
  title        = {Pith review of: HybridSim: A Physics-Learning Hybrid Digital Twin for mmWave Human Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFZ55TQO}},
  note         = {Machine review of arXiv:2607.15806}
}

read the original abstract

High-fidelity simulation of mmWave radar signals for dynamic human motion is valuable for developing radar-based human sensing models; yet collecting accurately labeled measurements for a specific deployment site remains expensive. We present HybridSim, a physics-learning hybrid simulator that synthesizes mmWave radar signals from dynamic human meshes under a fixed indoor room configuration, explicitly decoupling propagation into two components. To parameterize the human subject, we use a tri-plane representation to extract human features and a Graph Convolutional Network to stabilize optimization and mitigate gradient instability. The direct signal path is modeled via an inverse-rendering formulation with a microfacet BRDF to capture primary surface reflections. In parallel, the indirect path is approximated by combining 3D Gaussian Splatting with a virtual-receiver geometry to fit and reproduce site-specific multipath interference patterns, achieving substantially lower computational cost than explicit full ray tracing. Experiments in a fixed-room setting show improved agreement with a physically based reference and consistent gains on downstream radar-based human sensing tasks when using HybridSim for site-specific data augmentation.

Figures

Figures reproduced from arXiv: 2607.15806 by the authors.

**Figure 2.** Pipeline of our framework. We extract human features from a canonical pose using a tri-plane representation and obtain scene features through positional encoding. To model the direct path, the human features are masked and reconstructed via a Graph Convolutional Network to generate spatially regularized embeddings. These enhanced features are processed by a BRDF Decoder to predict material properties including rough… view at source ↗

**Figure 3.** Illustration of the indirect path simulation using surface-aligned [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Qualitative comparison of RD heatmaps. The red boxes provide zoomed-in views of the micro-Doppler signatures generated by articulated human limb movements. Compared to existing baseline simulators (mmGPE and RF-Genesis), HybridSim synthesizes signals that exhibit signi…

**Figure 5.** Figure 5: Visual comparison of Range-Doppler heatmaps generated under [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 7.** Figure 7: Visual comparison of ground truth RD heatmaps in linear amplitude [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

**Figure 8.** Figure 8: Visual symptom of unstable training without GCN. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

**Figure 9.** Figure 9: Visual analysis of the effect of GCN-enabled inference. [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]

**Figure 10.** Figure 10: Qualitative decomposition of HybridSim in RD space. For the same motion frame or short sequence, we show the direct-path signal, the indirect-path signal, the learned noise term, the final HybridSim output, and the corresponding ground truth. The direct path explains …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 3 canonical work pages

[1]

Association for Computing Machinery, New York, NY, USA, 1 edn

Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: SCAPE: Shape Completion and Animation of People. Association for Computing Machinery, New York, NY, USA, 1 edn. (2023),https://doi.org/10.1145/3596711.3596797

arXiv 2023
[2]

In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., de Mello, S., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient Geometry- aware 3D Generative Adversarial Networks . In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16102–16112. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2...

arXiv 2022
[3]

Chen, X., Ding, J., Zheng, K., Fang, X., Zhang, X., Lu, C.X., Li, Z.: Invertwin: Solving inverse problems via differentiable radio frequency digital twin (2025), https://arxiv.org/abs/2508.14204

arXiv 2025
[4]

In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, X., Feng, Z., Qian, K., Zhang, X.: Radio frequency ray tracing with neural object representation for enhanced rf modeling. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21339–21348 (2025). https://doi.org/10.1109/CVPR52734.2025.01988

arXiv 2025
[5]

In: Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems

Chen, X., Feng, Z., Sun, K., Qian, K., Zhang, X.: 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. p. 464–477. SenSys ’24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3666025.3699351

arXiv 2024
[6]

In: Proceedings of the 21st ACM Conference on Embedded Networked Sensor Systems

Chen, X., Zhang, X.: Rf genesis: Zero-shot generalization of mmwave sensing through simulation-based data synthesis and generative diffusion models. In: Proceedings of the 21st ACM Conference on Embedded Networked Sensor Systems. p. 28–42. SenSys ’23, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3625687.3625798

arXiv 2024
[7]

Chen, X., Zhang, X., Xia, Q., Fang, X., Lu, C.X., Li, Z.: Differentiable radio frequency ray tracing for millimeter-wave sensing (2023),https://arxiv.org/abs/ 2311.13182

arXiv 2023
[8]

https: //www.seas.upenn.edu/~richeek/blog/2024/mmwave- material- extraction/ (2024), accessed: 2026-03-02

Das, R.: Building differentiable simulators for mmwave radar signals. https: //www.seas.upenn.edu/~richeek/blog/2024/mmwave- material- extraction/ (2024), accessed: 2026-03-02

2024

Show all 36 references

[9]

Heitz, E.: Understanding the masking-shadowing function in microfacet-based brdfs (2014),https://api.semanticscholar.org/CorpusID:123674006

2014
[10]

IEEE Journal of Microwaves5(2), 356–372 (2025).https: //doi.org/10.1109/JMW.2025.3535077

Hofmann, N., Wirth, V., Bräunig, J., Ullmann, I., Vossiek, M., Weyrich, T., Stamminger, M.: Inverse rendering of near-field mmwave mimo radar for ma- terial reconstruction. IEEE Journal of Microwaves5(2), 356–372 (2025).https: //doi.org/10.1109/JMW.2025.3535077

2025
[11]

In: 2024 IEEE International Conference on Big Data (BigData)

Joshi, V., Xu, S., Cao, Q., Zhu, Y., Wang, P., Xue, H.: Towards robust mmwave- basedhumanactivityrecognitionusinglargesimulateddatasetformodelpretraining. In: 2024 IEEE International Conference on Big Data (BigData). pp. 7332–7337 (2024).https://doi.org/10.1109/BigData62323.20...

2024
[12]

ACM Trans

Kerbl, B., Kopanas, G., Leimkuehler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4) (Jul 2023).https: //doi.org/10.1145/3592433

2023 doi
[13]

In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human Gaussian Splats . In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 505–515. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2024).https://doi.org/10.1109/CVPR52733....

2024
[14]

ACM Trans

Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: a skinned multi-person linear model. ACM Trans. Graph.34(6) (Nov 2015).https://doi. org/10.1145/2816795.2818013

2015
[15]

Mehra, R., Tripathi, P., Sheffer, A., Mitra, N.J.: Technical section: Visibility of noisy point cloud data. Comput. Graph.34(3), 219–230 (Jun 2010).https: //doi.org/10.1016/j.cag.2010.03.002

2010 doi
[16]

Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM65(1), 99–106 (Dec 2021).https://doi.org/10.1145/3503250

2021 doi
[17]

In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive Body Capture: 3D Hands, Face, and Body From a Single Image . In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10967–10977. IEEE Computer S...

2019
[18]

IEEE Journal of Microwaves1(4), 962–974 (2021).https://doi.org/10

Schüßler, C., Hoffmann, M., Bräunig, J., Ullmann, I., Ebelt, R., Vossiek, M.: A realistic radar ray tracing simulator for large mimo-arrays in automotive environ- ments. IEEE Journal of Microwaves1(4), 962–974 (2021).https://doi.org/10. 1109/JMW.2021.3104722

2021
[19]

In: Proceedings of the 18th Eurographics Conference on Rendering Techniques

Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet models for refraction through rough surfaces. In: Proceedings of the 18th Eurographics Conference on Rendering Techniques. p. 195–206. EGSR’07, Eurographics Association, Goslar, DEU (2007),https://dl.acm.org/doi/1...

2007
[21]

ACM MobiCom ’23, Association for Computing Machinery, New York, NY, USA (2023).https://doi.org/10.1145/3570361.3613302

Xue, H., Cao, Q., Miao, C., Ju, Y., Hu, H., Zhang, A., Su, L.: Towards generalized mmwave-basedhumanposeestimationthroughsignalaugmentation.In:Proceedings of the 29th Annual International Conference on Mobile Computing and Networking. ACM MobiCom ’23, Association for Computing...

2023
[22]

In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services

Xue, H., Ju, Y., Miao, C., Wang, Y., Wang, S., Zhang, A., Su, L.: mmmesh: towards 3d real-time dynamic human mesh construction using millimeter-wave. In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services. p. 269–282. MobiSys ...

2021 doi
[23]

Yang, K., Dong, G., Ji, S., Du, W., Srivastava, M.: Gsrf: Complex-valued 3d gaussian splatting for efficient radio-frequency data synthesis (2025),https://arxiv.org/ abs/2502.01826

2025
[24]

IEEE Access3, 1089–1100 (2015).https://doi.org/10.1109/ ACCESS.2015.2453991

Yun, Z., Iskander, M.F.: Ray tracing for radio propagation modeling: Principles and applications. IEEE Access3, 1089–1100 (2015).https://doi.org/10.1109/ ACCESS.2015.2453991

2015
[25]

Neural Process

Zhang, H., Lu, G., Zhan, M., Zhang, B.: Semi-supervised classification of graph convolutional networks with laplacian rank constraints. Neural Process. Lett.54(4), 2645–2656 (Aug 2022).https://doi.org/10.1007/s11063-020-10404-7

2022 doi
[26]

IEEE Trans- actions on Wireless Communications25, 10419–10433 (2026).https://doi.org/ 10.1109/TWC.2026.3652154 18 W

Zhang, L., Sun, H., Berweger, S., Gentile, C., Qingyang Hu, R.: Rf-3dgs: Wireless channel modeling with radio radiance field and 3d gaussian splatting. IEEE Trans- actions on Wireless Communications25, 10419–10433 (2026).https://doi.org/ 10.1109/TWC.2026.3652154 18 W. Xiong et al

2026
[27]

In: 2024 International Conference on Computing, Networking and Communications (ICNC)

Zhang, L., Sun, H., Sun, J., Hu, R.Q.: Wisegrt: Dataset for site-specific indoor radio propagation modeling with 3d segmentation and differentiable ray-tracing: (invited paper). In: 2024 International Conference on Computing, Networking and Communications (ICNC). pp. 744–748 (...

2024
[28]

In: Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems

Zhang, X., Li, Z., Zhang, J.: Synthesized millimeter-waves for human motion sensing. In: Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems. p. 377–390. SenSys ’22, Association for Computing Machinery, New York, NY, USA (2023).https://doi.org/10.1145/3...

2023
[29]

In: Proceedings of the 29th Annual International Conference on Mobile Computing and Networking

Zhao, X., An, Z., Pan, Q., Yang, L.: Nerf2: Neural radio-frequency radiance fields. In: Proceedings of the 29th Annual International Conference on Mobile Computing and Networking. ACM MobiCom ’23, Association for Computing Machinery, New York, NY, USA (2023).https://doi.org/10...

2023
[30]

Torso rotations: The subject stands in place and continuously twists their upper body
[31]

Clockwise walking: The subject walks in a continuous circular trajectory in a clockwise direction
[32]

Counter-clockwise walking: The subject walks in a continuous circular trajectory in a counter-clockwise direction
[33]

Arm swing: The subject stands still while randomly swinging his/her arms horizontally, upward, or downward
[34]

HybridSim 25

Walking back and forth: The subject walks straight toward the radar and then walks backward, keeping the arms relatively still. HybridSim 25
[35]

Walking back and forth with arm swing: The subject walks straight toward the radar and backward while actively and naturally swinging their arms
[36]

Walking in the place: The subject performs a marching motion without spatial displacement
[37]

Lunges: The subject keeps performing a lunge pose, alternating between using his/her left and right leg. This diverse action space ensures that the generated mmWave signals are rigor- ously tested against a wide spectrum of kinematic complexities, ranging from dominant torso r...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Association for Computing Machinery, New York, NY, USA, 1 edn

Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: SCAPE: Shape Completion and Animation of People. Association for Computing Machinery, New York, NY, USA, 1 edn. (2023),https://doi.org/10.1145/3596711.3596797

arXiv 2023

[2] [2]

In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., de Mello, S., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient Geometry- aware 3D Generative Adversarial Networks . In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16102–16112. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2...

arXiv 2022

[3] [3]

Chen, X., Ding, J., Zheng, K., Fang, X., Zhang, X., Lu, C.X., Li, Z.: Invertwin: Solving inverse problems via differentiable radio frequency digital twin (2025), https://arxiv.org/abs/2508.14204

arXiv 2025

[4] [4]

In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, X., Feng, Z., Qian, K., Zhang, X.: Radio frequency ray tracing with neural object representation for enhanced rf modeling. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21339–21348 (2025). https://doi.org/10.1109/CVPR52734.2025.01988

arXiv 2025

[5] [5]

In: Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems

Chen, X., Feng, Z., Sun, K., Qian, K., Zhang, X.: 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. p. 464–477. SenSys ’24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3666025.3699351

arXiv 2024

[6] [6]

In: Proceedings of the 21st ACM Conference on Embedded Networked Sensor Systems

Chen, X., Zhang, X.: Rf genesis: Zero-shot generalization of mmwave sensing through simulation-based data synthesis and generative diffusion models. In: Proceedings of the 21st ACM Conference on Embedded Networked Sensor Systems. p. 28–42. SenSys ’23, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3625687.3625798

arXiv 2024

[7] [7]

Chen, X., Zhang, X., Xia, Q., Fang, X., Lu, C.X., Li, Z.: Differentiable radio frequency ray tracing for millimeter-wave sensing (2023),https://arxiv.org/abs/ 2311.13182

arXiv 2023

[8] [8]

https: //www.seas.upenn.edu/~richeek/blog/2024/mmwave- material- extraction/ (2024), accessed: 2026-03-02

Das, R.: Building differentiable simulators for mmwave radar signals. https: //www.seas.upenn.edu/~richeek/blog/2024/mmwave- material- extraction/ (2024), accessed: 2026-03-02

2024