Pith. sign in

REVIEW 5 major objections 5 minor 16 references

Learning Sensor Placement from Demonstration for UAV networks

T0 review · 5 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A convolutional network can learn a drone-placement utility function from expert demonstrations alone.

desk verdict A useful proof of concept for learning placement from demonstrations, but the learned utility is not verified as an argmax model, so treat the 75–91% correspondence numbers with caution. read the letter →

arxiv 1909.01636 v1 pith:MOWHXBWI submitted 2019-09-04 stat.AP

Assia Benbihi , Matthieu Geist , Cédric Pradalier This is my paper · ORCID
classification stat.AP
keywords learningfromdemonstrationutilityfunctionUAVplacementsensorconvolutionalneuralnetworksaliencymapregressionqualityofserviceartgalleryproblem
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

This paper argues that drone placement for civil missions can be automated by learning the expert's implicit utility function from demonstrations, rather than writing one by hand. The authors train a convolutional neural network (CNN) to map a 2D mission image to a 2D utility surface, and claim that the surface's maxima reproduce expert drone locations closely enough to match expert quality of service in simulated Wi-Fi coverage and building-monitoring missions. The central device is to blur the expert's discrete locations into Gaussian saliency maps, turning the hard non-differentiable problem of matching maxima into plain mean-squared-error regression. If the claim holds, each new mission no longer requires a bespoke utility function; the learned model supplies one from previous expert solutions.

What carries the argument

The load-bearing object is the Gaussian-saliency regression loss. Expert positions ζ_a are projected onto a 2D map as a sum of unnormalized Gaussians, using one sharp, high-amplitude Gaussian to punctuate exact positions and a second wider, low-amplitude Gaussian to create a learning signal in the areas between them. The CNN output is trained to match that map under an MSE loss, and at inference the predicted placements are the local maxima of the network's output. This transformation makes the entire argmax-and-assignment objective differentiable and is what allows the model to learn a utility function without a closed-form expression. The authors also rely on the submodular diminishing-returns structure of sensor placement to justify the loss's behavior as the number of proposed maxima grows.

What would settle it

Train the same model on demonstrations from an expert whose placements depend on a channel not present in the input map, such as a no-fly region stated only in the expert's instructions or crowd velocities from a time-lapse; if the learned placements regularly violate that hidden constraint or fail to match the expert's locations above chance on held-out maps, the claim that the CNN recovers the expert's true utility from static images is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a utility function generator U, implemented as an encoder–decoder CNN, can be trained purely from expert location demonstrations to imitate the expert's placement policy. For a mission rasterized as W_a with expert locations ζ_a, the training target R(ζ_a) is a 2D map with an unnormalized Gaussian peak at every expert position—typically at two scales, one sharp and one broad—and the loss is the squared Euclidean distance between R(ζ_a) and the network output U(W_a). Because this loss is differentiable, the network can be trained end-to-end without ever computing utility maxima or solving the assignment between learned and expert locations during training. The paper reports correspondence ratios of 75% (Wi-Fi strategy 1), 89% (Wi-Fi strategy 2), 91% (AGP monitoring), and 86% (fortress monitoring), with quality-of-service metrics comparable to or above those of the expert oracle and always higher than the correspondence ratio. The paper interprets the excess as evidence that when the network does not reproduce the expert's exact positions, its maxima are alternative valid solutions under the same implicit utility.

Load-bearing premise

The load-bearing premise is that the expert's placement strategy is a deterministic function of a single static 2D rasterized map of the mission, with no hidden information, temporal dynamics, or mixture of competing strategies.

Editorial extensions

If this is right

  • Automated drone deployment becomes feasible from past expert missions alone, without a closed-form utility function or task-specific hand-engineering.
  • Learned placements that differ from the expert's are still mission-valid: quality of service exceeds the correspondence ratio in every experiment, so the utility surface encodes alternative valid solutions.
  • The same pipeline transfers from Wi-Fi crowd-density maps to polygon building maps, indicating the method is not tied to a particular input representation.
  • The demonstration rasterization is not merely preprocessing: tuning the amplitude of the secondary Gaussian changes the correspondence ratio by 20–29%, so the saliency-map representation directly controls learning success.

Reading between the lines

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

  • If the learned utility function genuinely captures expert trade-offs, it could be reused as a prior or initialization for online adaptation in previously unseen environments, a step the paper does not test.
  • The Gaussian-saliency regression trick is generic: any placement or viewpoint-selection problem stated as sparse expert points on a raster could be trained the same way, provided the single-strategy assumption holds.
  • The static-map determinism assumption predicts a concrete failure mode: an expert who consults crowd velocities, 3D structure, or hidden constraints during placement will be only partially imitated; testing with such hidden-channel experts would bound the method's scope.
  • Because the training target is a smooth Gaussian map, the learned utility surface may be robust to small input perturbations; this could be quantified by perturbing input maps and measuring how much the predicted placements shift.
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.

Referee Report

5 major / 5 minor

Summary. The paper proposes a learning-from-demonstration approach to UAV/sensor placement. Given a rasterized 2D mission map, a CNN generates a 2D utility function whose local maxima are taken to be the learned sensor locations. The expert demonstrations are converted into a Gaussian-mixture saliency map, and the network is trained with a pixel-wise MSE loss between that map and the CNN output. The method is evaluated in a simulated Wi-Fi UAV network with two coverage strategies and in building-monitoring tasks based on the Art Gallery Problem and the Fortress Problem. The reported correspondence ratios between learned and expert placements range from 75% to 91%, precision from 72% to 82%, and task QoS is close to or equal to the expert's. The authors argue that additional learned maxima are alternative valid solutions and that the method generalizes across different input representations.

Significance. If the central claim holds, the paper offers a practical route to automating UAV placement without a hand-crafted closed-form utility function, which is genuinely relevant for applied network deployment. The problem is clearly formulated, the differentiable surrogate loss is an original and simple idea, and the use of task-level QoS metrics that are not part of the training loss provides some independent grounding for the imitation result. However, the paper is a proof of concept: there are no repeated runs, no error bars, no comparison against placement baselines, and no targeted validation of the claim that unmatched maxima are valid solutions. The strengths are the clean problem framing and the cross-domain demonstration; the main gap is that the experimental evidence is currently too thin to support the strength of the conclusions.

major comments (5)
  1. [II-B, Eq. (5)] The training loss is a pixel-wise MSE between the CNN output U(Wa) and the Gaussian mixture R(zeta_a). It never evaluates the argmax or solves the assignment between predicted and expert locations, so minimizing Eq. (5) does not in general control the set of local maxima of U(Wa). A slightly displaced or flattened bump can have small MSE while its peak is far from any expert location, and low-amplitude spurious maxima are only weakly penalized. The precision values in Tables IV and V (72-82%) show that such spurious maxima occur in practice. Because Eq. (1) defines the method entirely through the maxima of U(Wa), this surrogate loss is load-bearing: the paper should either relate Eq. (5) to the maximum-matching error theoretically, or add and evaluate a post-hoc maxima-selection/refinement step, or provide direct empirical evidence that the maxima of the learned map are trustworthy on new missions. As written, the model may be learning a smoothed version of the expert density rather than a utility function whose maxima can be reliably interpreted.
  2. [IV, Tables IV-V] The paper repeatedly explains unmatched learned maxima as 'other valid solutions', but this assertion is not tested. The QoS is computed for the whole learned placement set, not for the unmatched maxima separately. It is entirely possible that the matched maxima alone deliver all of the QoS while the unmatched maxima are useless or even harmful. This matters because the correspondence ratios are only 75-91%, and the 'alternative valid solutions' argument is the main defense for the gap. Please provide direct evidence, for example by computing QoS for the set of matched maxima only, for the set of unmatched maxima only, and for the full set, on the same test scenarios.
  3. [IV-C and Table II] The grid search over (alpha_2, sigma_2) is not clearly separated from the test data. The reported correspondence ratios appear to be the best values obtained by tuning these hyperparameters, and no separate validation set is mentioned. This can inflate the reported metrics and the 'alpha_2-rise' values. Please use a validation split or nested cross-validation for hyperparameter selection, report the test metrics for all grid points, and state explicitly how the final hyperparameters were chosen.
  4. [III-IV] No baseline methods are compared against. To support the claim that the learned utility function embeds expert knowledge and produces valid alternatives, the experiments should include simple heuristics such as placing sensors at density peaks, farthest-point sampling, random placement with the same number of sensors, and, for the monitoring tasks, the original AGP/FP solvers themselves. Without such comparisons, the QoS results mainly show that the learned placements are not catastrophic; they do not show that the learned policy is competitive with reasonable alternatives.
  5. [III-IV] All reported numbers are single-run values. CNN training is stochastic and the simulation scenarios are randomly generated, so the reader cannot assess the variability of the correspondence ratio, precision, or QoS. Please provide means and standard deviations over multiple seeds and, where applicable, over multiple generated test scenario sets.
minor comments (5)
  1. [Abstract] The phrase 'form demonstrations only' in the abstract should be 'from demonstrations only'.
  2. [I, Eq. (1)] The signature U: R^{h x w} -> R^2 -> R is confusing: U(Wa) is later treated as a function on pixel coordinates, but the notation suggests a map into R^2. Please write U(Wa) explicitly as a function on the pixel grid.
  3. [II-A, Eq. (1)] The expression argmax_{zeta in R^{N x 2}, N in N} U(Wa)(.) = zeta_a is not well-defined, since U(Wa) is defined on pixel coordinates rather than on sets of locations. Please reformulate, for example as 'the set of local maxima of U(Wa) equals zeta_a'.
  4. [II-B] The 'Modularity condition' bullet is hard to parse: please state the intended mathematical property explicitly and explain why the Gaussian-blurred MSE loss satisfies it.
  5. [IV-C, Figure 6] Figure 6 shows only the best and worst correspondence ratios for each task and the resulting alpha_2-rise. Please report the full grid results so the reader can judge the sensitivity of the method to alpha_2 and see whether the chosen values lie in a flat or a steep region.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned utility function is trained on demonstrations and tested on held-out missions, with an independent QoS metric.

full rationale

The paper's claimed derivation is a standard supervised-learning pipeline: a CNN U is trained to regress a Gaussian-smoothed target R(zeta_a) constructed from expert locations (Eq. 5), and the learned placements are the maxima of U(W_a). The objective and the primary metric both involve matching expert locations, but they are not identical by construction: the loss is a differentiable surrogate (MSE on saliency maps) that does not directly optimize the correspondence-ratio metric or the argmax condition, and the reported correspondence ratios are computed on held-out test missions (Table II). The outcome therefore depends on generalization, not on definition. The QoS evaluation is independent of the loss and provides external grounding, even though it uses the same simulator that generated the demonstrations. The discussion of the sigma/alpha parameters is hyperparameter tuning, not a fitted parameter renamed as a prediction. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The paper's explicit assumptions (single strategy, static rasterized input) are limitations relevant to correctness and generalization, but they do not make the derivation circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on hand-set (α1, σ1) and grid-searched (α2, σ2) loss parameters, on the domain assumptions that the expert utility is submodular, single-strategy, and inferable from a static 2D map, and on the unproven surrogate quality of the Gaussian MSE loss. No new physical entities are introduced.

free parameters (4)
  • α1 (narrow Gaussian amplitude) = 255
    Set by hand for the narrow Gaussian; amplitude chosen high enough to penalize misplaced maxima.
  • σ1 (narrow Gaussian variance) = 1
    Set by hand so that the saliency map has no residual local maxima around expert positions.
  • α2 (wide Gaussian amplitude) = per task, e.g., 50 (UAV1), 30 (UAV2), 40 (AGP/FP)
    Selected by grid search over [20,60]^2; controls tolerance to distant maxima and guides learning. Chosen per task to maximize correspondence ratio.
  • σ2 (wide Gaussian variance) = per task, e.g., 50 (UAV1), 30 (UAV2), 40 (AGP), 20 (FP)
    Selected jointly with α2 by grid search; affects the learning signal range.
assumptions (4)
  • domain assumption The expert's placement problem can be cast as maximizing a submodular function f over positions (diminishing returns).
    Section II-B: the loss design is motivated by submodularity, but the paper does not prove that the learned utility function or the loss satisfies the diminishing-return condition; it assumes the expert behaves submodularly.
  • domain assumption There is one solution strategy per task, though multiple solutions may exist.
    Section II-B: 'We assume that there may be several solutions for one task but only one solution strategy.' This is load-bearing because if the expert mixes strategies, a single utility function cannot be learned.
  • domain assumption A 2D rasterized map of the mission contains all information the expert uses to decide placements.
    The input to U is a single 2D image Wa (Section II-A); temporal dynamics and 3D structure are not represented. The paper assumes the density map captures the relevant state, though Figure 5 suggests temporal dynamics matter.
  • ad hoc to paper A Gaussian-blurred MSE loss is a faithful differentiable surrogate for matching utility maxima.
    Section II-B states the loss is defined to meet modularity and differentiability requirements, but the equivalence to the argmax/assignment loss is asserted, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Sensor Placement from Demonstration for UAV networks." pith.science (2026). https://pith.science/paper/MOWHXBWI

@misc{pith2026190901636,
  author       = {Pith},
  title        = {Pith review of: Learning Sensor Placement from Demonstration for UAV networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOWHXBWI}},
  note         = {Machine review of arXiv:1909.01636}
}
read the original abstract

This work demonstrates how to leverage previous network expert demonstrations of UAV deployment to automate the drones placement in civil applications. Optimal UAV placement is an NP-complete problem: it requires a closed-form utility function that defines the environment and the UAV constraints, it is not unique and must be defined for each new UAV mission. This complex and time-consuming process hinders the development of UAV-networks in civil applications. We propose a method that leverages previous network expert solutions of UAV-network deployment to learn the expert's untold utility function form demonstrations only. This is especially interesting as it may be difficult for the inspection expert to explicit his expertise into such a function as it is too complex. Once learned, our model generates a utility function which maxima match expert UAV locations. We test this method on a Wi-Fi UAV network application inside a crowd simulator and reach similar quality-of-service as the expert. We show that our method is not limited to this UAV application and can be extended to other missions such as building monitoring.

Figures

Figures reproduced from arXiv: 1909.01636 by the authors.

Figure 1
Figure 1. The crowd map is converted to a density map fed to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The CNN is trained by comparing its utility function [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Left: FP, the UAV cover the exterior perimeter. Right: [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Pedsim crowd simulation rendered in ROS. The grey [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 5
Figure 5. Figure 5: Average end-user throughput over time during a dy [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Correspondence ratio for several α2. Top: Wi-Fi UAV. Bottom: Monitoring UAV (AGP,FP). Setting α2 can boost the learning performance up to 29%. on how to improve because all the learning signal is restricted to the neighbourhood of ζa. This can significantly slow down t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages

  1. [1]

    Tensorflow: a system for large-scale machine learning

    A BADI , M., B ARHAM , P., C HEN , J., C HEN , Z., D AVIS, A., D EAN , J., D EVIN , M., G HEMAWAT, S., I RVING , G., I SARD , M., ET AL . Tensorflow: a system for large-scale machine learning. In OSDI (2016), vol. 16, pp. 265–283

    work page 2016

  2. [2]

    D., C HERNOVA , S., V ELOSO , M., AND BROWNING , B

    A RGALL , B. D., C HERNOVA , S., V ELOSO , M., AND BROWNING , B. A survey of robot learning from demonstration. Robotics and autonomous systems 57, 5 (2009), 469–483

    work page 2009

  3. [3]

    Segnet: A deep convolutional encoder-decoder architecture for image segmenta- tion

    B ADRINARAYANAN , V., K ENDALL , A., AND CIPOLLA , R. Segnet: A deep convolutional encoder-decoder architecture for image segmenta- tion. IEEE transactions on pattern analysis and machine intelligence 39, 12 (2017), 2481–2495

    work page 2017

  4. [4]

    State-of-the-art in visual attention modeling

    B ORJI , A., AND ITTI , L. State-of-the-art in visual attention modeling. IEEE transactions on pattern analysis and machine intelligence 35 , 1 (2013), 185–207

    work page 2013

  5. [5]

    DE CASTRO , P. M. M., D EVILLERS , O., H ERT, S., H OFFMANN , M., KETTNER , L., S CH ¨ONHERR , S., T IFREA , A., AND GIMENO , M. Geo- metric object generators. In CGAL User and Reference Manual, 4.10 ed. CGAL Editorial Board, 2017

    work page 2017

  6. [6]

    A guide to convolution arithmetic for deep learning

    D UMOULIN , V., AND VISIN , F. A guide to convolution arithmetic for deep learning. arXiv preprint arXiv:1603.07285 (2016)

    arXiv 2016

  7. [7]

    A randomized art-gallery algorithm for sensor placement

    G ONZ ´ALEZ -BANOS , H. A randomized art-gallery algorithm for sensor placement. In Proceedings of the seventeenth annual symposium on Computational geometry (2001), ACM, pp. 232–240

    work page 2001

  8. [8]

    C., AND GARCIA -H ARO, J

    G UILLEN -PEREZ , A., S ANCHEZ -IBORRA , R., C ANO , M.-D., SANCHEZ -AARNOUTSE , J. C., AND GARCIA -H ARO, J. Wifi networks on drones. In ITU Kaleidoscope: ICTs for a Sustainable World (ITU WT), 2016 (2016), IEEE, pp. 1–8

    work page 2016

Show all 16 references
  1. [9]

    Survey on unmanned aerial vehicle networks for civil applications: A communications view- point

    H AYAT, S., Y ANMAZ , E., AND MUZAFFAR , R. Survey on unmanned aerial vehicle networks for civil applications: A communications view- point. IEEE Communications Surveys & Tutorials 18 , 4 (2016), 2624– 2661

    2016

  2. [10]

    Adam: A method for stochastic optimization

    K INGMA , D., AND BA, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  3. [11]

    Near-optimal sensor placements in gaussian processes: Theory, efficient algorithms and empirical studies

    K RAUSE , A., S INGH , A., AND GUESTRIN , C. Near-optimal sensor placements in gaussian processes: Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research 9, Feb (2008), 235–284

    2008

  4. [12]

    Art gallery theorems and algorithms , vol

    O’ ROURKE , J. Art gallery theorems and algorithms , vol. 57. Oxford University Press Oxford, 1987

    1987

  5. [13]

    C., AND WYNN , H

    S HEWRY, M. C., AND WYNN , H. P. Maximum entropy sampling. Journal of applied statistics 14 , 2 (1987), 165–170

    1987

  6. [14]

    Very deep convolutional net- works for large-scale image recognition

    S IMONYAN , K., AND ZISSERMAN , A. Very deep convolutional net- works for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  7. [15]

    C., DE REZENDE , P

    T OZONI , D. C., DE REZENDE , P. J., AND DE SOUZA , C. C. A practical iterative algorithm for the art gallery problem using integer linear programming. Optimization Online (2013)

    2013

  8. [16]

    D., T AYLOR , G

    Z EILER , M. D., T AYLOR , G. W., AND FERGUS , R. Adaptive deconvo- lutional networks for mid and high level feature learning. In Computer Vision (ICCV), 2011 IEEE International Conference on (2011), IEEE, pp. 2018–2025

    2011

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.