Localizing Backscatters by a Single Robot With Zero Start-up Cost

Shengkai Zhang; Sheyang Tang; Shi Jin; Tao Jiang; Wei Wang

REVIEW 4 major objections 6 minor 16 references

Localizing Backscatters by a Single Robot With Zero Start-up Cost

T0 review · 4 major / 6 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 single robot with commodity WiFi and an inertial sensor can localize both itself and low-power backscatter tags at once, with no map, no landmarks, and no site survey.

desk verdict Rover is a genuinely new system—first landmark-free single-robot backscatter localization—but the IMU odometry derivation is under-specified and the experiments are too thin to support the headline accuracy. read the letter →

arxiv 1908.03297 v1 pith:GEIPS426 submitted 2019-08-09 cs.RO cs.NI

Shengkai Zhang , Wei Wang , Sheyang Tang , Shi Jin , Tao Jiang This is my paper · ORCID

classification cs.ROcs.NI

keywords backscatterlocalizationsimultaneousandmappingangleofarrivalinertialmeasurementunitcommodityWiFiindoorslidingwindowoptimizationlow-powerIoTtags

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

The paper sets out to show that backscatter localization need not require any start-up investment: no pre-measured landmarks, no floor plan, and no multiple fixed access points. It presents Rover, a robot-borne system that estimates the robot's own trajectory and the positions of nearby backscatter tags at the same time, using only WiFi angle-of-arrival measurements and inertial readings. The point of the claim is that low-power IoT tags could be localized power-on-and-go, removing a major deployment cost for backscatter in smart homes, warehouses, and object tracking. Experiments report 39.3 cm mean trajectory error for the robot and 74.6 cm for tags in line-of-sight deployments.

What carries the argument

The load-bearing mechanism is a sliding-window graph SLAM (simultaneous localization and mapping) in which hidden states are robot positions and tag positions, and edges are two constraint types: AoA constraints, written as a cross-product zero-residual between the measured direction to a tag and the vector from robot to tag, and odometry constraints from IMU preintegration between consecutive robot states. Because the rotations R0_k are treated as known via short-term gyroscope integration and the robot is assumed to move on a horizontal plane, the whole optimization becomes a linear least-squares problem, solved in a sliding window so the robot can keep navigating in real time.

What would settle it

Take motion-capture or ruler-measured ground truth and run Rover on a floor tilted by a known angle, or add a fixed artificial bias to the gyroscope in software; the reconstructed robot path and tag positions should rotate and stretch by a predictable amount, revealing whether the horizontal-plane IMU assumption is the one carrying the metre-scale accuracy.

Watch

Extended reading notes

Core claim

The central claim is that a single moving robot, with no prior knowledge of the room, can simultaneously solve for its own path and the positions of low-power backscatter tags by fusing two kinds of measurements in a joint optimization: drift-free WiFi angles of arrival to each tag and metric-scale inertial odometry from an IMU. The angle measurements alone yield only shapes, not sizes; the IMU supplies metres, and the joint optimization lets the WiFi angles hold the inertial drift in check while the inertial motion defines scale and stabilizes the geometry. The paper argues this is the first such system to work with commodity WiFi, a single robot, and zero start-up cost, and the experiments are offered as evidence: 39.3 cm robot trajectory error and 74.6 cm tag localization error, with tags entering and leaving the window as the robot moves.

Load-bearing premise

The system's scale and direction come from the robot's inertial sensors, which are assumed to sit on a flat horizontal floor with gravity cleanly removed; if those sensors drift or the floor tilts, the reconstructed positions rotate or stretch and the advertised accuracy is not guaranteed.

Editorial extensions

If this is right

Backscatter tags can be deployed without surveying landmarks or building a map: switch on the robot and move, and the system builds the geometry as it goes.
One mobile robot replaces the fixed multi-access-point infrastructure that earlier backscatter localization required; the robot's motion creates the multi-view geometry.
Tags beyond the robot's immediate communication range are still localizable if the robot later moves near them, since tags can enter and leave the sliding window.
The same architecture gives the robot a self-localization service in spaces without GPS, since trajectory and tags are solved together.
Real-time operation is part of the claim: the sliding window keeps computation bounded so the robot can navigate while localizing.

Reading between the lines

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

If the inertial sensor is the only source of metric scale, then the same framework should tolerate replacing the IMU with any other metric odometry, such as wheel encoders, visual-inertial odometry, or ultrasonic ranging; this is an extension the paper does not test.
The 3D extension the paper calls trivial likely needs more than a larger antenna array: a single planar array gives only azimuth, so height estimation would have to come from the IMU's vertical channel or from additional measurements, which is a nontrivial robustness question.
A stress test that follows directly from the stated weakness: run the robot on a slightly inclined floor or inject a known gyroscope bias, and the reported tag positions should rotate and rescale by a predictable amount; if they do not, the system is less dependent on the horizontal-plane assumption than the formulation suggests.
The same AoA-odometry fusion could be applied to other low-power signals, such as Bluetooth beacons, which the paper names as future work; if it transfers, anchor-free localization would be a generic service rather than a backscatter-specific one.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Rover is worth reading: it's the first backscatter localization system I know of that runs on a single robot with commodity WiFi and no map or landmarks, using the robot's IMU to supply metric scale and AoA to correct drift. The frequency-sweeping trick for separating multiple tags is practical, the sliding-window graph formulation is standard but appropriate, and they built a real prototype with plausible results. No circularity found: the accuracy numbers come from direct measurement and the covariances are calibrated a priori.

The weaknesses are mostly empirical, but they matter. One trajectory, no baselines (IMU-only or AoA-only), no error bars, no independent ground truth—they use the commanded trajectory as truth, which can diverge from where the robot actually went—and no code or data. The headline 74.6 cm is the LOS-only mean; the NLOS tag's 145.9 cm error is excluded without saying so in the abstract. These are fixable but currently make the claimed accuracy unverifiable.

More seriously, the odometry factor in Section III-B is under-specified. Eqs. (10)–(12) never define the frames for R0_k, ν_k, and g clearly. Under the conventional reading—R0_k body-to-world, g world-frame—the gravity term −R0_k g Δt²/2 in Eq. (10) is suspicious; the correct term would be −g Δt²/2 (or +, depending on sign convention). Eq. (12) is algebraically consistent with Eq. (10), so either both are right in a nonstandard convention or both inherit the same error. Rotation estimation is deferred to reference [15], there are no IMU bias states, and because AoA constraints are scale-invariant, a global scale error from an uncalibrated accelerometer would not be corrected by any other measurement. That cuts at the zero-start-up claim, which rests on the IMU providing the metric scale. A referee needs a clean derivation and a bias/scale sensitivity analysis.

The core idea is credible and the system is genuinely new, but the paper as written doesn't fully prove it. It deserves serious peer review—major revision at least, and release of code/data would help. I'd bring it to a reading group if you work in RF sensing or SLAM.

Referee Report

4 major / 6 minor

Summary. The paper presents Rover, a system that uses a robot equipped with commodity WiFi and an inertial measurement unit (IMU) to simultaneously estimate the robot's trajectory and the positions of multiple battery-free backscatter tags, without requiring a prior map or known landmarks. The approach estimates angle-of-arrival (AoA) of backscattered WiFi signals with a three-antenna array and fuses these AoA measurements with IMU odometry in a sliding-window, graph-based SLAM formulation. The reported evaluation is a single run in a 9x5 m meeting room with four tags, reporting a mean robot localization error of 39.3 cm and a mean tag error of 74.6 cm (the latter averaged over the three LOS tags only).

Significance. If the claims hold, Rover would be the first system to localize backscatter tags with a single mobile robot and commodity WiFi without any site survey or landmark setup, and the zero-start-up aspect is attractive for IoT deployment. The paper includes a working prototype on off-the-shelf hardware and addresses practical issues such as multi-tag interference and real-time processing. However, the significance is tempered by the load-bearing dependence on an incompletely derived and unvalidated IMU factor, and by an experimental evaluation that is too limited to substantiate the headline accuracy and zero-start-up claims.

major comments (4)

[§III-B, Eqs. (10)-(12)] The odometry factor as written is frame-inconsistent. With g = [0,0,9.8]^T declared to be the vertical world-frame gravity, Eq. (10) multiplies g by R0_k, which produces a body-frame vector in a world-frame update; Eq. (11) applies R_{k+1}^k to the world-frame velocity ν_k, which is not the rotation linking world and local frames; and Eq. (12) subtracts ν_k Δt (a world-frame quantity) from body-frame terms multiplied by R_k^0. Because this factor is the sole source of metric scale and absolute orientation in the system, the derivation must be corrected and a consistent frame convention adopted.
[§III-B, Eqs. (4)-(12)] The state vector S in Eq. (4) contains only robot and tag positions, but the odometry constraints in Eqs. (10)-(12) depend on velocities ν_k that are not part of S. The paper states only that ν_k 'can be updated by Eq. (11)', which does not explain how the velocities used in the linear system remain consistent with the positions being optimized. Velocities should either be included in the joint optimization or the paper should specify a marginalization procedure; otherwise the least-squares problem is not defined over all unknowns.
[§III-B and §IV] The central claim of 'zero start-up cost' is not established. The state S contains no gyroscope or accelerometer bias states, and the rotation R0_k is obtained by 'solving a linear system' with details deferred to reference [15]. At startup, an uncalibrated gyro bias or accelerometer scale error will systematically corrupt every odometry edge, and since AoA constraints are scale-invariant, no other measurement can correct a global scale error. The paper should either include bias states in the optimization or report experiments with a cold-start, uncalibrated IMU and a sensitivity analysis of the results to IMU bias and scale.
[§IV-C and Abstract] The experimental support is too thin for the headline claims. There is one trajectory, no repeated trials or error bars, and no comparison baselines such as IMU-only or AoA-only positioning, so the benefit of the fusion cannot be assessed. Moreover, the abstract's '74.6 cm for the tags' is the mean over the three LOS tags only; the NLOS tag has an error of 145.9 cm and this exclusion is not stated in the abstract or introduction. The paper should also include a sensitivity analysis of the claimed 'insensitive' initial distance guess and of the tag-position initialization at (0,0).

minor comments (6)

[§III-B, Eq. (9)] The definition of R_k^t uses 'i' in the summation without a clear relation to the time index t; please rewrite with a consistent time index to avoid confusion.
[§III-B, Eqs. (7)-(12)] The notations R0_k, R_k^0, and R_{k+1}^k are used without a single explicitly stated convention; please define the frame mapping once (e.g., R_a^b maps frame a to frame b) and use it consistently throughout.
[§IV-C] The sentence 'Initially, the tags' locations are set to be (0,0)' should be clarified: if this is the optimizer's initial guess, the influence of this guess on convergence and final error should be reported.
[Figure 7(b)] The system deployment figure should show ground-truth tag positions and estimated tag positions as well as the trajectory; without ground truth the figure cannot support the accuracy claims.
[§II-B] The statement that ToF 'cannot correctly infer the traveling distance' is a limitation of ToF resolution; please explain how the joint AoA-ToF estimator identifies the direct path when multiple virtual paths may have similar ToFs.
[Footnote 2 and §III-B] The assumption that the robot moves on a horizontal plane is load-bearing for the gravity-subtraction step; list it explicitly as a limitation and state what happens if the floor is tilted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Rover's accuracy claims are measured end-to-end, and its SLAM constraints are standard residual formulations rather than fitted inputs renamed as predictions.

full rationale

The paper's derivation chain is self-contained with respect to circularity. The claimed accuracies (39.3 cm for the robot, 74.6 cm for tags) are end-to-end experimental measurements against ground truth, not outputs of a fitted model reproducing its own inputs. The AoA constraints in Eq. (8) rewrite the geometric relation as a zero residual via a cross product; this is a standard residual construction for direction observations, and the quantity being minimized is a measurement mismatch, not the target variable defined by the fit. The odometry constraints in Eq. (12) are obtained by algebraic rearrangement of the propagation model in Eq. (10), so they are not circular. The initial distance guess used in the AoA covariance is explicitly stated to be insensitive and is refined during optimization, which is an iterative weighting scheme rather than a prediction derived from the fit. The paper cites prior work for AoA estimation [6], IMU preintegration [14], and rotation initialization [15], but these are external references and not self-citations by the present authors, and none of them is invoked to force the paper's central conclusion. The potential frame inconsistency in the IMU odometry equations (10)-(12) identified in review is a correctness risk, not a circularity, and the reviewer's own score was 0.0. No load-bearing step reduces to its own inputs by definition.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard AoA/ToF super-resolution, planar 2D motion with known gravity, short-term gyroscope integration for rotations, and channel-based tag separation. No new physical entities are introduced. The only fitted values are sensor noise covariances and an initial distance guess used for residual weighting, neither of which is used to define the target result.

free parameters (2)

AoA observation noise covariance = pre-measured on hardware; values not reported
Weights the AoA residual terms in Eq. (6). It is a calibration quantity, not a target-of-the-fit, but if it is mis-measured the balance between AoA and IMU terms changes. The paper does not report the values.
Initial tag-robot distance d_j^i = unspecified 'reasonable guess'
Used in Eq. (8) to set the covariance weight for each AoA observation. The authors state that it is refined automatically and that the initial guess is insensitive, but no sensitivity study is provided.

assumptions (4)

domain assumption The received CSI is accurately modeled as a sum of virtual paths whose direct path is identifiable as the smallest-ToF AoA, as in Eq. (3).
Borrowed from SpotFi and WiTag [6,13]. If the direct path is not resolvable, the AoA observations fed into SLAM are biased. The paper only validates this through the AoA CDF in Section IV-B.
domain assumption The robot moves in a horizontal plane, gravity is known, and accelerometer readings after subtracting gravity give a reliable metric scale.
Stated in Section III-B: 'Since the robot only moves in a room (assuming a horizontal plane), it is safe to obtain the accelerations that account for motions by subtracting the gravity.' Floor tilt or unmodeled accelerometer bias breaks the scale.
domain assumption Rotations R0_k from short-term gyroscope integration are accurate enough to treat the position model as linear in Eqns. (10) to (12).
The paper omits the rotation-estimation details and refers to [15]. Orientation errors are not jointly optimized in the graph, so unmodeled gyro bias would propagate into both robot and tag positions.
domain assumption Each tag's frequency shift maps it to a separate WiFi channel with no residual sideband interference, and the receiver sweeps channels fast enough to stay synchronized with all tags.
From Section II-A, based on Eqns. (1) and (2). Scalability is limited by the number of non-overlapping WiFi channels and assumes the tag transistor toggling works as described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Localizing Backscatters by a Single Robot With Zero Start-up Cost." pith.science (2026). https://pith.science/paper/GEIPS426

@misc{pith2026190803297,
  author       = {Pith},
  title        = {Pith review of: Localizing Backscatters by a Single Robot With Zero Start-up Cost},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEIPS426}},
  note         = {Machine review of arXiv:1908.03297}
}

read the original abstract

Recent years have witnessed the rapid proliferation of low-power backscatter technologies that realize the ubiquitous and long-term connectivity to empower smart cities and smart homes. Localizing such low-power backscatter tags is crucial for IoT-based smart services. However, current backscatter localization systems require prior knowledge of the site, either a map or landmarks with known positions, increasing the deployment cost. To empower universal localization service, this paper presents Rover, an indoor localization system that simultaneously localizes multiple backscatter tags with zero start-up cost using a robot equipped with inertial sensors. Rover runs in a joint optimization framework, fusing WiFi-based positioning measurements with inertial measurements to simultaneously estimate the locations of both the robot and the connected tags. Our design addresses practical issues such as the interference among multiple tags and the real-time processing for solving the SLAM problem. We prototype Rover using off-the-shelf WiFi chips and customized backscatter tags. Our experiments show that Rover achieves localization accuracies of 39.3 cm for the robot and 74.6 cm for the tags.

Figures

Figures reproduced from arXiv: 1908.03297 by the authors.

**Figure 2.** (a) shows the interference caused by the backscatter [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The received signal traverses two physical paths whe [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Localization principle: triangulation with the rob [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 6.** Figure 6: The experimental platform. The left shows the receiv [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: (a) The accuracy of AoA estimation. (b) We use the NUC t [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

**Figure 8.** Figure 8: SLAM performance in the meeting room. C. Simultaneous Localization [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[15]

Initi alization- free monocular visual-inertial state estimation with appl ication to au- tonomous mavs,

S. Shen, Y . Mulgaonkar, N. Michael, and V . Kumar, “Initi alization- free monocular visual-inertial state estimation with appl ication to au- tonomous mavs,” in Experimental robotics. Springer, 2016, pp. 211– 227

work page 2016
[1]

Wi- Fi backscatter: Internet connectivity for RF-powered devi ces,

B. Kellogg, A. Parks, S. Gollakota, J. R. Smith, and D. Wet herall, “Wi- Fi backscatter: Internet connectivity for RF-powered devi ces,” in ACM SIGCOMM COMP . COM., vol. 44, no. 4, 2014, pp. 607–618

work page 2014
[2]

PLoRa: a passive long-range data network from ambient LoRa transmissions,

Y . Peng, L. Shangguan, Y . Hu, Y . Qian, X. Lin, X. Chen, D. Fa ng, and K. Jamieson, “PLoRa: a passive long-range data network from ambient LoRa transmissions,” in ACM Proc. SIGCOMM , 2018

work page 2018
[3]

Backscatter communications over ambient ofdm signals using null subcarriers,

M. A. ElMossallamy, Z. Han, M. Pan, R. J¨ antti, K. G. Seddi k, and G. Y . Li, “Backscatter communications over ambient ofdm signals using null subcarriers,” in IEEE Proc. GLOBECOM , 2018

work page 2018
[4]

Netscatter: Enab ling large- scale backscatter networks,

M. Hessar, A. Najaﬁ, and S. Gollakota, “Netscatter: Enab ling large- scale backscatter networks,” in USENIX Proc. NSDI , 2019

work page 2019
[5]

A game theoreti c approach for backscatter-aided relay communications in hybrid radi o networks,

J. Li, J. Xu, S. Gong, C. Li, and D. Niyato, “A game theoreti c approach for backscatter-aided relay communications in hybrid radi o networks,” in IEEE Proc. GLOBECOM , 2018

work page 2018
[6]

Localizing low-power backscatter tags using commodity wiﬁ,

M. Kotaru, P . Zhang, and S. Katti, “Localizing low-power backscatter tags using commodity wiﬁ,” in ACM Proc. CoNEXT , 2017

work page 2017
[7]

Drone relays for battery-fr ee networks,

Y . Ma, N. Selby, and F. Adib, “Drone relays for battery-fr ee networks,” in ACM Proc. SIGCOMM , 2017

work page 2017

Show all 16 references

[8]

Autonomous aerial navigation using monocular visual-ine rtial fusion,

Y . Lin, F. Gao, T. Qin, W. Gao, T. Liu, W. Wu, Z. Y ang, and S. S hen, “Autonomous aerial navigation using monocular visual-ine rtial fusion,” J. Field Robot. , vol. 35, no. 1, pp. 23–51, 2018

2018
[9]

Pervasive ﬂoorplan generati on based on only inertial sensing: feasibility, design, and implement ation,

Y . He, J. Liang, and Y . Liu, “Pervasive ﬂoorplan generati on based on only inertial sensing: feasibility, design, and implement ation,” IEEE JSAC, vol. 35, no. 5, pp. 1132–1140, 2017

2017
[10]

Tool rel ease: Gath- ering 802.11n traces with channel state information,

D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool rel ease: Gath- ering 802.11n traces with channel state information,” ACM SIGCOMM Comput. Commun. Rev. , 2011

2011
[11]

Hitchhik e: Practical backscatter using commodity wiﬁ,

P . Zhang, D. Bharadia, K. Joshi, and S. Katti, “Hitchhik e: Practical backscatter using commodity wiﬁ,” in ACM Proc. SenSys , 2016

2016
[12]

Decimeter-level l ocalization with a single wiﬁ access point,

D. V asisht, S. Kumar, and D. Katabi, “Decimeter-level l ocalization with a single wiﬁ access point,” in USENIX Proc. NSDI , 2016

2016
[13]

Spotﬁ: D ecimeter level localization using wiﬁ,

M. Kotaru, K. Joshi, D. Bharadia, and S. Katti, “Spotﬁ: D ecimeter level localization using wiﬁ,” in ACM Proc. SIGCOMM , 2015

2015
[14]

IMU preintegration on m ani- fold for e ﬃcient visual-inertial maximum-a-posteriori estimation,

D. F. S. D. Forster C, Carlone L, “IMU preintegration on m ani- fold for e ﬃcient visual-inertial maximum-a-posteriori estimation, ” in Proc. Robot. Sci. Syst. , 2015

2015
[16]

Visual-inertial-aided na vigation for high- dynamic motion in built environments without initial condi tions,

T. Lupton and S. Sukkarieh, “Visual-inertial-aided na vigation for high- dynamic motion in built environments without initial condi tions,” IEEE Trans. Robot., vol. 28, no. 1, pp. 61–76, 2012

2012

Pith tools

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

[1] [15]

Initi alization- free monocular visual-inertial state estimation with appl ication to au- tonomous mavs,

S. Shen, Y . Mulgaonkar, N. Michael, and V . Kumar, “Initi alization- free monocular visual-inertial state estimation with appl ication to au- tonomous mavs,” in Experimental robotics. Springer, 2016, pp. 211– 227

work page 2016

[2] [1]

Wi- Fi backscatter: Internet connectivity for RF-powered devi ces,

B. Kellogg, A. Parks, S. Gollakota, J. R. Smith, and D. Wet herall, “Wi- Fi backscatter: Internet connectivity for RF-powered devi ces,” in ACM SIGCOMM COMP . COM., vol. 44, no. 4, 2014, pp. 607–618

work page 2014

[3] [2]

PLoRa: a passive long-range data network from ambient LoRa transmissions,

Y . Peng, L. Shangguan, Y . Hu, Y . Qian, X. Lin, X. Chen, D. Fa ng, and K. Jamieson, “PLoRa: a passive long-range data network from ambient LoRa transmissions,” in ACM Proc. SIGCOMM , 2018

work page 2018

[4] [3]

Backscatter communications over ambient ofdm signals using null subcarriers,

M. A. ElMossallamy, Z. Han, M. Pan, R. J¨ antti, K. G. Seddi k, and G. Y . Li, “Backscatter communications over ambient ofdm signals using null subcarriers,” in IEEE Proc. GLOBECOM , 2018

work page 2018

[5] [4]

Netscatter: Enab ling large- scale backscatter networks,

M. Hessar, A. Najaﬁ, and S. Gollakota, “Netscatter: Enab ling large- scale backscatter networks,” in USENIX Proc. NSDI , 2019

work page 2019

[6] [5]

A game theoreti c approach for backscatter-aided relay communications in hybrid radi o networks,

J. Li, J. Xu, S. Gong, C. Li, and D. Niyato, “A game theoreti c approach for backscatter-aided relay communications in hybrid radi o networks,” in IEEE Proc. GLOBECOM , 2018

work page 2018

[7] [6]

Localizing low-power backscatter tags using commodity wiﬁ,

M. Kotaru, P . Zhang, and S. Katti, “Localizing low-power backscatter tags using commodity wiﬁ,” in ACM Proc. CoNEXT , 2017

work page 2017

[8] [7]

Drone relays for battery-fr ee networks,

Y . Ma, N. Selby, and F. Adib, “Drone relays for battery-fr ee networks,” in ACM Proc. SIGCOMM , 2017

work page 2017