SkySense: A Semi-Supervised Generative Framework for UAV Localization in ISAC Networks

Cixiao Zhang; Jie Yang; Shenghan Luo; Shi Jin; Wenjun Zhang; Yang Wang; Yin Xu

arxiv: 2606.04076 · v1 · pith:G3F3D24Xnew · submitted 2026-06-02 · 📡 eess.SP

SkySense: A Semi-Supervised Generative Framework for UAV Localization in ISAC Networks

Shenghan Luo , Yin Xu , Jie Yang , Yang Wang , Cixiao Zhang , Shi Jin , Wenjun Zhang This is my paper

Pith reviewed 2026-06-28 08:46 UTC · model grok-4.3

classification 📡 eess.SP

keywords UAV localizationCSI fingerprintingsemi-supervised learningconsistency modelISACgenerative modelmulti-BS fusionself-supervised encoder

0 comments

The pith

A generative consistency model maps self-supervised CSI features to UAV positions with 9.77 cm accuracy using only 1% labeled data.

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

The paper develops a semi-supervised generative framework to localize UAVs using channel state information in integrated sensing and communication networks. It addresses data scarcity and multipath ambiguity by first using a self-supervised encoder to learn latent representations from unlabeled CSI sequences based on trajectory correlations. A consistency model then generates coordinate predictions from these latents, fine-tuned with few labels, while a fusion step combines outputs from multiple base stations. This yields low-latency inference and 9.77 cm mean error under 3-BS fusion with 1% labels, beating fully supervised and discriminative semi-supervised approaches.

Core claim

The framework extracts robust spatial features from unlabeled CSI via a self-supervised encoder exploiting temporal correlations in flight trajectories, then employs a consistency model to map the latent space to physical coordinates while jointly fine-tuning with limited labels, achieving 9.77 cm mean localization error with 1% label fraction in 3-BS setups and low inference latency.

What carries the argument

The consistency model, a derivative of diffusion architectures, that models the conditional distribution from latent representations to coordinates and enables 1-2 step inference.

If this is right

Achieves mean localization error of 9.77 cm with only 1% labeled CSI data under 3-BS fusion.
Compresses inference to 1-2 steps avoiding latency of traditional diffusion models.
Outperforms existing fully supervised and semi-supervised discriminative baselines.
Supports lightweight distributed fusion across multiple base stations from multi-view geometry.

Where Pith is reading between the lines

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

The approach may apply to localizing other moving objects if their trajectories exhibit similar temporal correlations in CSI.
Reducing labeling requirements could accelerate deployment of ISAC-based localization in dynamic environments.
If the latent representations prove invariant across scenarios, the pre-trained encoder could transfer to new base station configurations without retraining.

Load-bearing premise

Temporal correlations in continuous UAV flight trajectories allow extraction of robust spatial features from unlabeled CSI sequences that the consistency model can accurately map to physical coordinates.

What would settle it

Measuring whether the mean localization error exceeds 20 cm on the real-world dataset when the self-supervised pretraining is removed or when label fraction is 1% without the generative component.

Figures

Figures reproduced from arXiv: 2606.04076 by Cixiao Zhang, Jie Yang, Shenghan Luo, Shi Jin, Wenjun Zhang, Yang Wang, Yin Xu.

**Figure 2.** Figure 2: The overall architecture of the proposed SkySense framework. The learning paradigm consists of two stages: Stage I extracts the underlying spatial [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Figure 4: Mean localization errors under varying label densities from 0.1% to [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 3.** Figure 3: CDF of 3D localization errors under the 1% label density setting. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 5.** Figure 5: The impact of inference steps on localization error in the single-BS [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

read the original abstract

Extreme data scarcity and inherent multipath spatial ambiguity severely limit existing deep learning-based channel state information (CSI) fingerprinting localization schemes for target unmanned aerial vehicles (UAVs). To overcome these challenges, we propose an end-to-end semi-supervised generative localization framework. First, by exploiting the temporal correlations inherent in continuous flight trajectories, a self-supervised encoder extracts robust spatial features from massive unlabeled CSI sequences to establish structured latent representations. Following this, we utilize a consistency model, a powerful derivative of diffusion architectures, as the core generative backbone to map the learned latent space to physical coordinates, jointly fine-tuning the pre-trained encoder with a strictly limited set of labeled CSI. This consistency formulation models the conditional distribution to resolve the mean collapse problem of discriminative models, while compressing the inference trajectory to 1-2 steps to avoid the latency bottleneck of traditional diffusion models. Furthermore, a lightweight distributed fusion mechanism is designed to aggregate spatial predictions across multiple base stations (BS) from a multi-view geometry perspective. Comprehensive evaluations on a real-world measurement dataset demonstrate that our framework achieves low latency and suppresses the mean localization error to 9.77 cm under a 3-BS fusion setup with only a 1\% label fraction, significantly outperforming existing fully supervised and semi-supervised discriminative baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SkySense puts together self-supervised temporal encoding and a consistency model for semi-supervised UAV CSI localization, reporting 9.77 cm error at 1% labels, but the result hinges on unshown details about whether the latents actually resolve spatial ambiguity.

read the letter

The main takeaway is that this paper builds a semi-supervised generative pipeline for UAV localization in ISAC networks: a self-supervised encoder exploits temporal correlations in flight trajectories to learn from unlabeled CSI, then a consistency model maps the latents to coordinates while fine-tuning on just 1% labels, with a lightweight multi-BS fusion step on top. It reports 9.77 cm mean error under 3-BS fusion on real measurements and claims to beat both fully supervised and semi-supervised discriminative baselines, with the generative backbone meant to avoid mean collapse and keep inference to 1-2 steps.

What the work does reasonably well is target a concrete pain point—label scarcity plus multipath ambiguity—in a practical setting, and the choice of consistency models for fast sampling plus the fusion mechanism from multi-view geometry makes sense for the domain. Evaluating on a real-world dataset rather than pure simulation adds some weight.

The soft spot is the one flagged in the stress-test: the performance numbers rest on the assumption that the self-supervised encoder extracts structured spatial features from the unlabeled sequences that the consistency model can then map reliably. The abstract gives no ablations, latent-space visualizations, or distribution checks to confirm the features are spatial rather than merely temporal, or that the short sampling trajectory preserves accuracy without introducing new ambiguity. If the full paper supplies those, the claim holds up better; if the outperformance traces mostly to the fusion or dataset quirks, the generative part may not be doing the heavy lifting advertised.

This is for people working on wireless sensing, ISAC, or UAV applications who need methods that work with very few labels. A reader in that niche would get a clear pipeline description and concrete numbers to compare against. It is coherent enough on its own terms to deserve peer review rather than a desk reject, though any referee would want to see the missing validation on the latent representations.

Referee Report

2 major / 1 minor

Summary. The paper proposes SkySense, a semi-supervised generative framework for UAV localization in ISAC networks. It uses a self-supervised encoder to extract robust spatial features from unlabeled CSI sequences by exploiting temporal correlations in UAV flight trajectories. A consistency model is then used as the generative backbone to map the latent representations to physical coordinates, fine-tuned with only 1% labeled data. A lightweight distributed fusion mechanism aggregates predictions from multiple base stations. Evaluations on a real-world dataset show a mean localization error of 9.77 cm with low latency under 3-BS fusion, outperforming fully supervised and semi-supervised discriminative baselines.

Significance. If the experimental results hold, this framework could significantly advance the field of CSI-based localization for UAVs by addressing data scarcity and multipath ambiguity through a combination of self-supervised learning and efficient generative modeling, enabling high-accuracy localization with minimal labeled data and low inference latency.

major comments (2)

[Abstract] Abstract: The central performance claim of suppressing the mean localization error to 9.77 cm with only 1% label fraction is presented without any details on the experimental setup, baseline methods, number of trials, or statistical significance tests, which is load-bearing for assessing whether the outperformance is valid.
[Abstract] Abstract: The key assumption that the self-supervised encoder extracts spatial features from unlabeled CSI sequences to resolve multipath spatial ambiguity is stated but not supported by any latent-space analysis, ablation studies on feature robustness, or distribution statistics in the provided text, undermining the claim that the consistency model reliably maps to coordinates.

minor comments (1)

The abstract could benefit from a brief mention of the dataset characteristics or the number of base stations used in the fusion setup for better context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the two major comments on the abstract below, indicating planned revisions to improve clarity while preserving the abstract's conciseness.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claim of suppressing the mean localization error to 9.77 cm with only 1% label fraction is presented without any details on the experimental setup, baseline methods, number of trials, or statistical significance tests, which is load-bearing for assessing whether the outperformance is valid.

Authors: We agree that the abstract would benefit from additional context on the key claim. The full manuscript details the real-world dataset, 3-BS fusion setup, baselines (fully supervised and semi-supervised discriminative models), averaging over multiple trials, and statistical reporting in Sections 4 and 5. In the revision, we will expand the abstract by one sentence to briefly note the experimental conditions, baselines, and that results include standard deviations, while keeping within length limits. This directly addresses the concern without altering the core claim. revision: yes
Referee: [Abstract] Abstract: The key assumption that the self-supervised encoder extracts spatial features from unlabeled CSI sequences to resolve multipath spatial ambiguity is stated but not supported by any latent-space analysis, ablation studies on feature robustness, or distribution statistics in the provided text, undermining the claim that the consistency model reliably maps to coordinates.

Authors: The manuscript body (Section 3.2 and 4.3) includes latent-space visualizations, ablation studies isolating the self-supervised encoder's role in handling multipath, and feature distribution statistics demonstrating robustness. The abstract summarizes these findings concisely. To strengthen the abstract, we will add a short clause referencing the supporting analyses in the full paper. If the referee's note applies strictly to the abstract text alone, we acknowledge the linkage could be clearer and will revise accordingly. revision: partial

Circularity Check

0 steps flagged

No derivation chain or self-referential reductions identified

full rationale

The provided abstract and description contain no equations, parameter fittings, or derivation steps that reduce the reported localization error or performance claims to inputs by construction. The framework is described in terms of methodological choices (self-supervised encoder exploiting temporal correlations, consistency model for mapping latent space, distributed fusion), with results presented as empirical outcomes from real-world dataset evaluation rather than derived predictions. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing elements. The central claims rest on experimental outperformance, making the presentation self-contained against external benchmarks with no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; full paper likely contains additional modeling choices and parameters not visible here. The core assumptions are extracted directly from the abstract text.

axioms (2)

domain assumption Temporal correlations inherent in continuous flight trajectories allow extraction of robust spatial features from unlabeled CSI sequences
Explicitly invoked in the abstract as the basis for the self-supervised encoder.
domain assumption A consistency model can map learned latent representations to physical coordinates while resolving mean collapse when fine-tuned on limited labels
Stated as the justification for choosing the consistency model over discriminative alternatives.

pith-pipeline@v0.9.1-grok · 5773 in / 1422 out tokens · 28155 ms · 2026-06-28T08:46:40.075512+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Secure Physical Layer Communications for Low-Altitude Economy Networking: A Survey,

L. Cai, J. Wang, R. Zhang, Y . Zhang, T. Jiang, D. Niyato, X. Wang, A. Jamalipour, and X. Shen, “Secure Physical Layer Communications for Low-Altitude Economy Networking: A Survey,”IEEE Communica- tions Surveys & Tutorials, vol. 28, pp. 2497–2530, 2026

2026
[2]

Toward Realization of Low-Altitude Economy Networks: Core Architecture, Integrated Technologies, and Future Directions,

Y . Wang, G. Sun, Z. Sun, J. Wang, J. Li, C. Zhao, J. Wu, S. Liang, M. Yin, P. Wang, D. Niyato, S. Sun, and D. In Kim, “Toward Realization of Low-Altitude Economy Networks: Core Architecture, Integrated Technologies, and Future Directions,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 5, pp. 2788–2820, 2025

2025
[3]

An Overview of Cellular ISAC for Low-Altitude UA V: New Opportunities and Challenges,

Y . Song, Y . Zeng, Y . Yang, Z. Ren, G. Cheng, X. Xu, J. Xu, S. Jin, and R. Zhang, “An Overview of Cellular ISAC for Low-Altitude UA V: New Opportunities and Challenges,”IEEE Communications Magazine, vol. 63, no. 12, pp. 88–95, 2025

2025
[4]

Passive UA V Detection Based on Channel Estimation and Temporal Variation Network,

J. Bai, Z. Zhang, Z. Xiao, H. Zhou, Y . Hei, X. Liu, T. Li, and L. Jiao, “Passive UA V Detection Based on Channel Estimation and Temporal Variation Network,”IEEE Transactions on Wireless Communications, vol. 25, pp. 14 351–14 366, 2026

2026
[5]

Generative AI for Integrated Sensing and Communication: Insights From the Physical Layer Perspective,

J. Wang, H. Du, D. Niyato, J. Kang, S. Cui, X. Shen, and P. Zhang, “Generative AI for Integrated Sensing and Communication: Insights From the Physical Layer Perspective,”IEEE Wireless Communications, vol. 31, no. 5, pp. 246–255, 2024

2024
[6]

Consistency mod- els,

Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency mod- els,” inProceedings of the 40th International Conference on Machine Learning. PMLR, 2023, pp. 32 211–32 252

2023
[7]

Improved Techniques for Training Consistency Models

Y . Song and P. Dhariwal, “Improved techniques for training consistency models,”arXiv preprint arXiv:2310.14189, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[8]

CSI-Based Fingerprinting for Indoor Localization: A Deep Learning Approach,

X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-Based Fingerprinting for Indoor Localization: A Deep Learning Approach,”IEEE Transactions on Vehicular Technology, vol. 66, no. 1, pp. 763–776, 2017

2017
[9]

ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information,

H. Chen, Y . Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information,”IEEE Access, vol. 5, pp. 18 066–18 074, 2017

2017
[10]

A CSI- Based Data-Driven Localization Framework Using Small-Scale Training Datasets in Single-Site MIMO Systems,

K. Vuckovic, S. Hosseini, F. Hejazi, and N. Rahnavard, “A CSI- Based Data-Driven Localization Framework Using Small-Scale Training Datasets in Single-Site MIMO Systems,”IEEE Transactions on Wireless Communications, vol. 23, no. 11, pp. 16 346–16 358, 2024

2024
[11]

CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,

G. Zhu, Y . Hu, W. Gao, W.-H. Wang, B. Wang, and K. J. R. Liu, “CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,” arXiv preprint arXiv:2505.21866, 2025

work page arXiv 2025
[12]

Self-Supervised Learning for User Localization,

A. Dash, J. Gu, G. Wang, and N. Ansari, “Self-Supervised Learning for User Localization,” in2024 International Conference on Computing, Networking and Communications (ICNC), 2024, pp. 886–890

2024
[13]

Measurement Testbed for Radar and Emitter Localization of UA V at 3.75 GHz,

J. Beuster, C. Andrich, M. D ¨obereiner, S. Schieler, M. Engelhardt, C. Schneider, and R. Thom ¨a, “Measurement Testbed for Radar and Emitter Localization of UA V at 3.75 GHz,” in2023 17th European Conference on Antennas and Propagation (EuCAP), 2023, pp. 1–5

2023

[1] [1]

Secure Physical Layer Communications for Low-Altitude Economy Networking: A Survey,

L. Cai, J. Wang, R. Zhang, Y . Zhang, T. Jiang, D. Niyato, X. Wang, A. Jamalipour, and X. Shen, “Secure Physical Layer Communications for Low-Altitude Economy Networking: A Survey,”IEEE Communica- tions Surveys & Tutorials, vol. 28, pp. 2497–2530, 2026

2026

[2] [2]

Toward Realization of Low-Altitude Economy Networks: Core Architecture, Integrated Technologies, and Future Directions,

Y . Wang, G. Sun, Z. Sun, J. Wang, J. Li, C. Zhao, J. Wu, S. Liang, M. Yin, P. Wang, D. Niyato, S. Sun, and D. In Kim, “Toward Realization of Low-Altitude Economy Networks: Core Architecture, Integrated Technologies, and Future Directions,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 5, pp. 2788–2820, 2025

2025

[3] [3]

An Overview of Cellular ISAC for Low-Altitude UA V: New Opportunities and Challenges,

Y . Song, Y . Zeng, Y . Yang, Z. Ren, G. Cheng, X. Xu, J. Xu, S. Jin, and R. Zhang, “An Overview of Cellular ISAC for Low-Altitude UA V: New Opportunities and Challenges,”IEEE Communications Magazine, vol. 63, no. 12, pp. 88–95, 2025

2025

[4] [4]

Passive UA V Detection Based on Channel Estimation and Temporal Variation Network,

J. Bai, Z. Zhang, Z. Xiao, H. Zhou, Y . Hei, X. Liu, T. Li, and L. Jiao, “Passive UA V Detection Based on Channel Estimation and Temporal Variation Network,”IEEE Transactions on Wireless Communications, vol. 25, pp. 14 351–14 366, 2026

2026

[5] [5]

Generative AI for Integrated Sensing and Communication: Insights From the Physical Layer Perspective,

J. Wang, H. Du, D. Niyato, J. Kang, S. Cui, X. Shen, and P. Zhang, “Generative AI for Integrated Sensing and Communication: Insights From the Physical Layer Perspective,”IEEE Wireless Communications, vol. 31, no. 5, pp. 246–255, 2024

2024

[6] [6]

Consistency mod- els,

Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency mod- els,” inProceedings of the 40th International Conference on Machine Learning. PMLR, 2023, pp. 32 211–32 252

2023

[7] [7]

Improved Techniques for Training Consistency Models

Y . Song and P. Dhariwal, “Improved techniques for training consistency models,”arXiv preprint arXiv:2310.14189, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[8] [8]

CSI-Based Fingerprinting for Indoor Localization: A Deep Learning Approach,

X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-Based Fingerprinting for Indoor Localization: A Deep Learning Approach,”IEEE Transactions on Vehicular Technology, vol. 66, no. 1, pp. 763–776, 2017

2017

[9] [9]

ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information,

H. Chen, Y . Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information,”IEEE Access, vol. 5, pp. 18 066–18 074, 2017

2017

[10] [10]

A CSI- Based Data-Driven Localization Framework Using Small-Scale Training Datasets in Single-Site MIMO Systems,

K. Vuckovic, S. Hosseini, F. Hejazi, and N. Rahnavard, “A CSI- Based Data-Driven Localization Framework Using Small-Scale Training Datasets in Single-Site MIMO Systems,”IEEE Transactions on Wireless Communications, vol. 23, no. 11, pp. 16 346–16 358, 2024

2024

[11] [11]

CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,

G. Zhu, Y . Hu, W. Gao, W.-H. Wang, B. Wang, and K. J. R. Liu, “CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,” arXiv preprint arXiv:2505.21866, 2025

work page arXiv 2025

[12] [12]

Self-Supervised Learning for User Localization,

A. Dash, J. Gu, G. Wang, and N. Ansari, “Self-Supervised Learning for User Localization,” in2024 International Conference on Computing, Networking and Communications (ICNC), 2024, pp. 886–890

2024

[13] [13]

Measurement Testbed for Radar and Emitter Localization of UA V at 3.75 GHz,

J. Beuster, C. Andrich, M. D ¨obereiner, S. Schieler, M. Engelhardt, C. Schneider, and R. Thom ¨a, “Measurement Testbed for Radar and Emitter Localization of UA V at 3.75 GHz,” in2023 17th European Conference on Antennas and Propagation (EuCAP), 2023, pp. 1–5

2023