Reinforcement Learning-Based Monocular Vision Approach for Autonomous UAV Landing

Tarik Houichime; Younes El Amrani

REVIEW 3 major objections 6 minor 15 references

Reinforcement Learning-Based Monocular Vision Approach for Autonomous UAV Landing

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A drone can be taught to land using only a front-facing camera and a color-shifting circular marker.

desk verdict A plausible marker idea buried in a paper that claims validation without providing any: equations are placeholders, experiments are future tense. read the letter →

arxiv 2505.06963 v1 pith:DIP3B7HN submitted 2025-05-11 cs.RO cs.AIcs.CV

Tarik Houichime , Younes EL Amrani This is my paper · ORCID

classification cs.ROcs.AIcs.CV

keywords UAVautonomouslandingmonocularvisionreinforcementlearningimage-basedvisualservoinglenticularlandmarksharedautonomydepthestimationpaddetection

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 proposes a way for a drone to land on a pad using only its front-facing monocular camera, with no GPS, downward camera, or depth sensor. The key idea is to place a specially designed lenticular circle, a 2D landmark printed in three colors, near the landing pad so that its apparent color and shape change with the drone's viewing position. The drone's controller uses reinforcement learning to approximate the mapping from those visual changes to altitude and depth, turning the landing task into an optimization problem that can be trained in simulation. The authors report simulation experiments in which a drone estimates altitude and depth accurately enough to land on both static and moving pads, and they frame the approach as a low-cost alternative for settings where conventional sensors are unavailable.

What carries the argument

The load-bearing object is the lenticular circle landmark: a 2D circle printed with three distinct colors, mounted at a fixed height and inclination near the landing pad. Because its apparent diameter and color distribution change with viewing geometry, the front camera's image of the circle encodes altitude and depth—the perceived diameter $D_1$ feeds altitude through $Q(D_1)$, and the viewing angle $\theta$ feeds depth through $Q'(\theta)$. Reinforcement learning is then used to approximate these estimation functions and to map visual features directly to control commands, while a self-supervised arbitration layer blends AI and human control inputs. The marker's distortion and color are the only source of position information, so the entire scheme stands or falls on whether those cues remain visible and unambiguous throughout the approach.

What would settle it

Compute the projection of the lenticular circle from all approach positions: if two different (altitude, depth) pairs can produce the same apparent diameter, color, and shape, the inverse mapping is not unique and the learned estimates cannot be reliable. A simpler field check is to fly the final approach with the marker outside the camera's field of view and see whether altitude and depth estimates remain within the reported 2–5 cm error bounds.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the visual distortion and color variation of a 2D lenticular circle, seen by a front-facing monocular camera, carry enough information to estimate both altitude and depth during approach. It states this as two learned functions, $\text{Altitude} = Q(D_1)$ and $\text{Depth} = Q'(\theta)$, where $D_1$ is the perceived diameter of the circle and $\theta$ the viewing angle, with reinforcement learning approximating the functions from visual input to control actions. In the paper's account, this lets the UAV dynamically estimate its position and altitude during descent, and the reported tables show altitude errors around 2–5 cm and lateral displacements around 1.5–4 cm on static tests, with somewhat larger tracking errors on moving platforms. The authors also claim that a self-supervised shared-autonomy layer lets a human pilot and an AI co-pilot arbitrate control inputs without conflict.

Load-bearing premise

The whole method rests on the assumption that the apparent color and shape of the lenticular circle always supply enough unambiguous visual information to estimate the drone's altitude and depth, even when the front camera loses sight of the marker during the final moments of descent.

Editorial extensions

If this is right

A UAV could perform approach and landing without downward-facing cameras, GPS, or depth sensors, lowering the cost and weight of landing systems.
Because the estimation functions are learned, the same framework could be retrained for different markers, landing environments, or aircraft dynamics by running further simulations.
The system is claimed to generalize to moving landing platforms, which would allow landings on vehicles or vessels at moderate speeds and rotation rates.
The shared-autonomy layer would let a human pilot retain control while the AI co-pilot corrects the approach, supporting mixed manual-autonomous operations.

Reading between the lines

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

The paper leaves implicit that the same appearance-based ranging idea could work with other printed patterns whose color or geometry shifts with viewing angle; if so, the method generalizes beyond lenticular circles to a family of low-cost visual beacons.
A consequence not stated in the paper is that the learned mapping's reliability can be checked before any flight by computing the projection of the marker: whenever two approach poses yield the same apparent diameter and color, the inverse mapping is ambiguous.
The reported error tables appear to be representative single trials rather than statistical aggregates, and the paper does not say how the reinforcement learning policy was trained or how many episodes were used; repeating the tests across lighting and weather conditions would be the natural next step.

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

The punchline: this is a method proposal disguised as a validated result. The core idea—using a front-facing monocular camera and a specially designed lenticular circle marker to estimate both altitude and depth, with reinforcement learning learning the mapping—is not stupid, but the paper presents no actual implementation, no optical model, and no real experiment. The reader's rejection is right.

What's new: the specific combination of a lenticular circle whose color changes with viewing angle and whose apparent diameter and shape encode range, used with a front camera rather than a downward one. I haven't seen that exact marker design in the UAV landing literature, though all the components (monocular servoing, RL for landing, custom markers) exist in the cited papers. The paper also frames the task as shared autonomy with a human pilot, but that part is purely conceptual.

What it does well: it identifies a real gap—landing with only a forward camera and no depth sensor—and honestly lists several limitations: lighting sensitivity, single-landmark fragility, and the blind area in the final approach. The writing is understandable, if repetitive.

The soft spots are large. Equations (1) and (2) just declare Altitude = Q(D1) and Depth = Q'(θ); there is no camera projection, no marker geometry, no uniqueness argument. Monocular circle pose estimation has known scale ambiguities, and the paper doesn't address them. The experiments section is future tense: "The tests are conducted" and "the UAV is tasked" describe a protocol, not results. Tables 1 and 2 give numbers with no trial counts, error bars, simulation environment, RL architecture, or training details. The paper itself acknowledges a blind area but gives no fallback. References are duplicated (e.g., [12]/[13], [2]/[15], [5]/[16], [7]/[17]) and sections are misnumbered, which suggests the manuscript wasn't carefully assembled.

So the central claim—"accurate and reliable landings"—is unsupported. The idea could eventually be made to work, but as written it is a proposal, not a result.

Who is this for? Maybe a researcher looking for new visual marker ideas for landing, but even they won't get enough detail to build on. I wouldn't cite it. I wouldn't bring it to reading group. A serious editor should desk reject it, or at most send it back with a demand for actual experiments and a derivation of the mapping. I recommend rejecting.

Referee Report

3 major / 6 minor

Summary. The paper proposes a reinforcement-learning-based method for autonomous UAV landing using only a front-facing monocular camera and a specially designed lenticular-circle landmark placed near the landing pad. The authors claim that perceived color and shape variations of the landmark encode altitude and depth, with the mappings Altitude = Q(D1) and Depth = Q'(θ) learned via reinforcement learning. The manuscript reports simulation results in two tables and concludes that the UAV can achieve accurate and reliable landings without downward-facing cameras or depth sensors.

Significance. The problem addressed is genuine: reducing sensor cost and avoiding the constraints of downward-facing cameras for UAV landing is of practical interest. The paper also identifies a real gap in shared-autonomy landing research. However, as presented, the contribution is a proposal rather than a validated system. The central estimation mappings are asserted without derivation or reproducible training details, and the reported numerical results are not anchored to any experimental protocol. The paper itself repeatedly flags unresolved limitations, including dependence on controlled lighting and the lack of real-world validation. If the proposed approach were properly specified and validated, it could be a useful incremental contribution to vision-based landing, but at present the central claims are unsupported.

major comments (3)

[II.2.1 (Equations 1 and 2)] Equations (1) and (2) assert Altitude = Q(D1) and Depth = Q′(θ) with no optical model, camera projection equations, landmark dimensions, color-encoding geometry, or calibration procedure. Monocular estimation of depth and altitude from a circular landmark is subject to known scale ambiguities and ellipse-to-pose multiplicity, yet the manuscript gives no uniqueness argument, sensitivity analysis, or error-propagation discussion. Because these mappings are the load-bearing components of the entire landing claim, the paper's central assertion that the UAV 'can accurately estimate altitude and depth' is not established.
[III (Tables 1 and 2)] Tables 1 and 2 report altitude errors, lateral displacements, tracking errors, and landing displacements without specifying the simulation environment, the reinforcement learning algorithm, the policy architecture, the reward function, the sensor noise model, the number of trials, or the statistical dispersion of the results. The point values are presented without error bars or trial counts, so the reader cannot assess whether the differences between cases are meaningful or whether the results are reproducible. The claim in Section 3.3 that 'Simulation experiments have shown that the UAV can accurately estimate altitude and depth' is therefore unverifiable.
[III.3 and IV (Future Work)] The manuscript's own limitation statements contradict the strength of its conclusions. Section 3.3 acknowledges dependency on controlled lighting and shadows, and Section IV states that 'the transition from simulation to real-world implementation is a crucial step for validating the proposed method's practicality and effectiveness.' This implies that no real-world experiments have been conducted, despite the abstract's claim that the method was 'assessed by simulations and experiments.' Additionally, Figure 2 identifies a blind area during the final approach when the marker leaves the front camera's field of view, and no mechanism is proposed to estimate altitude or depth in that phase. These are central gaps, not minor caveats.

minor comments (6)

[I. Introduction] The text contains a duplicated word: 'the UAV may may not always have a clear line of sight' should be corrected to 'may not.'
[References] References [12] and [13] are identical, as are references [5] and [16] and references [15] and [17]. The reference list should be deduplicated.
[II. Methodology] Section numbering is inconsistent: '2.1 Image-Based Visual Servoing' appears twice, and the reinforcement learning subsection is numbered 2.1 rather than following the sequence. The subsections should be renumbered.
[II.2.1 and II.2.2] The terms 'reinforcement learning' and 'self-supervised learning' are used interchangeably, and the relationship between them is never defined. The shared-autonomy framework is mentioned but not formalized with equations, inputs, or a training objective.
[III.1 and III.2] The experimental scenarios are described in future tense ('The tests are conducted,' 'the UAV is tasked'), which is inconsistent with the paper's claim that experiments were performed. The time frame should be made explicit.
[III.3] Tables 1 and 2 would benefit from captions that explicitly state the metric definitions and the units of every column, as well as the number of repeated trials per condition. The current presentation of single point values is insufficient for a scientific claim.

Circularity Check

1 steps flagged · score 6.0 of 10

Altitude/depth accuracy claims are validated by the same learned Q/Q′ functions, making the central evidence self-referential rather than an independent prediction.

fitted input called prediction [Section II (Reinforcement Learning for Optimal Approach), Eqs. (1)–(2), and Section 3.3 Discussion]
"The reinforcement learning algorithm is trained in a simulated environment, where it hones its proficiency in approximating the functions that map visual inputs to control actions. Altitude = Q(D1) (1) Depth = Q′(θ) (2) ... Simulation experiments have shown that the UAV can accurately estimate altitude and depth based on the visual distortion and color variations of the landmark."

Equations (1) and (2) do not specify an optical model; they simply denote the altitude and depth estimates as the outputs of learned functions Q and Q′. The text states that these functions are approximated by training in the simulated environment, and the only reported evidence for the central claim is the error tables from that same simulation. Since Q and Q′ were fit in that environment and no independent ground-truth benchmark, unseen test split, closed-form projection model, or external constraint is provided, the reported 'accurate estimation' is a restatement of how well the fitted functions reproduce the simulation states, not an independent prediction. The validation loop is load-bearing: the claimed accuracy is the fitted behavior of the estimator itself.

full rationale

The paper contains no self-citations and no imported uniqueness theorem, so the citation-based circularity patterns are absent. The only circularity-like feature is that the central accuracy claim is not supported by an independent model: the altitude/depth estimator is simply denoted by Q and Q′ in Eqs. (1)–(2) and learned in the same simulated environment used for validation, so the reported altitude/depth errors reduce to the fitted behavior of those functions. The paper also flags its own blind-area limitation in Figure 2 without supplying a mechanism to handle the final approach, and the experimental sections mix future-tense descriptions with tabulated numbers while omitting simulation details such as lighting, sensor noise, policy architecture, training procedure, and train/test splits. Those are correctness and reproducibility concerns rather than circularity by themselves, but they compound the self-referential validation. Because the load-bearing claim of 'accurate and reliable landings' rests on this validation loop, the circularity score is 6 rather than 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim depends entirely on two learned functions Q and Q' whose inputs are visual cues from a newly designed marker. The information content of the marker, the visibility of the marker throughout approach, the fidelity of the simulation, and the ability of RL to learn the mapping are all assumed without supporting evidence.

free parameters (2)

Learned altitude mapping Q(D1) = not specified
Equation (1) defines altitude as Q(D1); Q is learned via RL from simulation data, so any altitude estimate is a fitted output, not a derived quantity.
Learned depth mapping Q'(theta) = not specified
Equation (2) defines depth as Q'(theta); Q' is learned via RL from the same visual cues, so depth estimates are fitted values.

assumptions (4)

domain assumption The apparent color and shape of the 2D lenticular circle vary uniquely and monotonically with viewing angle and distance.
Section 2.2 states the landmark conveys altitude and depth through perceived color and form, but no model, calibration, or uniqueness proof is given. If two different positions produce similar visual cues, the mapping is ambiguous.
domain assumption The front-facing monocular camera can observe the lenticular circle throughout the approach, including the final phase.
Section 2.1 and Figure 2 acknowledge a blind area during final approach, yet the method relies on continuous visual feedback from the landmark. The paper does not describe how estimation continues when the marker is out of view.
domain assumption The simulated environment faithfully reproduces real-world visual and flight dynamics.
The RL agent is trained in a simulated environment, and the validation tables are said to come from simulations/experiments, but the simulator, its fidelity, and its validation are not described.
domain assumption Reinforcement learning can approximate Q and Q' accurately enough for landing from visual inputs.
The paper asserts RL 'hones its proficiency' but gives no algorithm, state/action space, reward design, or convergence evidence.

invented entities (1)

2D lenticular circle landmark
purpose: Encodes altitude and depth in perceived color and shape for monocular estimation
The marker's information content is asserted in Section 2.2 without measurement, optical model, or independent validation. No physical prototype or image data is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reinforcement Learning-Based Monocular Vision Approach for Autonomous UAV Landing." pith.science (2026). https://pith.science/paper/DIP3B7HN

@misc{pith2026250506963,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning-Based Monocular Vision Approach for Autonomous UAV Landing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIP3B7HN}},
  note         = {Machine review of arXiv:2505.06963}
}

read the original abstract

This paper introduces an innovative approach for the autonomous landing of Unmanned Aerial Vehicles (UAVs) using only a front-facing monocular camera, therefore obviating the requirement for depth estimation cameras. Drawing on the inherent human estimating process, the proposed method reframes the landing task as an optimization problem. The UAV employs variations in the visual characteristics of a specially designed lenticular circle on the landing pad, where the perceived color and form provide critical information for estimating both altitude and depth. Reinforcement learning algorithms are utilized to approximate the functions governing these estimations, enabling the UAV to ascertain ideal landing settings via training. This method's efficacy is assessed by simulations and experiments, showcasing its potential for robust and accurate autonomous landing without dependence on complex sensor setups. This research contributes to the advancement of cost-effective and efficient UAV landing solutions, paving the way for wider applicability across various fields.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[13]

Deep learning for vision -based micro aerial vehicle autonomous landing,

L. Yu et al., “Deep learning for vision -based micro aerial vehicle autonomous landing,” Int. J. Micro Air Veh. , vol. 10, no. 2, pp. 171 –185, Jun. 2018, doi: 10.1177/1756829318757470

work page doi:10.1177/1756829318757470 2018
[15]

Proactive Guidance for Accurate UAV Landing on a Dynamic Platform: A Visual –Inertial Approach,

C.-W. Chang et al., “Proactive Guidance for Accurate UAV Landing on a Dynamic Platform: A Visual –Inertial Approach,” Sensors, vol. 22, no. 1, Art. no. 1, Jan. 2022, doi: 10.3390/s22010404

work page doi:10.3390/s22010404 2022
[16]

Towards autonomous landing on a moving vessel through fiducial markers,

R. Polvara, S. Sharma, J. Wan, A. Manning, and R. Sutton, “Towards autonomous landing on a moving vessel through fiducial markers,” in 2017 European Conference on Mobile Robots (ECMR) , Sep. 2017, pp. 1 –6. doi: 10.1109/ECMR.2017.8098671

work page arXiv 2017
[17]

A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform,

A. Rodriguez -Ramos, C. Sampedro, H. Bavle, P. de la Puente, and P. Campoy, “A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform,” J. Intell. Robot. Syst. , vol. 93, no. 1, pp. 351 – 366, Feb. 2019, doi: 10.1007/s10846-018-0891-8

doi:10.1007/s10846-018-0891-8 2019
[1]

Learning to Assist Drone Landings,

K. Backman, D. Kulić, and H. Chung, “Learning to Assist Drone Landings,” IEEE Robot. Autom. Lett. , vol. 6, no. 2, pp. 3192 –3199, Apr. 2021, doi: 10.1109/LRA.2021.3062572

arXiv 2021
[3]

Autonomous Approach and Landing for a Low -Cost Quadrotor Using Monocular Cameras,

S. Dotenco, F. Gallwitz, and E. Angelopoulou, “Autonomous Approach and Landing for a Low -Cost Quadrotor Using Monocular Cameras,” in Computer Vision - ECCV 2014 Workshops, L. Agapito, M. M. Bronstein, and C. Rother, Eds., in Lecture Notes in Computer Science. Cham: Springer International Publishing, 2015, pp. 209 –

work page 2014
[4]

Systems and methods for autonomous landing using a three dimensional evidence grid,

E. R. Goossen and Y. Ma, “Systems and methods for autonomous landing using a three dimensional evidence grid,” US8996207B2, Mar. 31, 2015 Accessed: Sep. 05,

work page 2015
[6]

Autonomous Quadrotor Landing using Deep Reinforcement Learning,

R. Polvara et al., “Autonomous Quadrotor Landing using Deep Reinforcement Learning,” Feb. 27, 2018, arXiv: arXiv:1709.03339. doi: 10.48550/arXiv.1709.03339

arXiv doi:10.48550/arxiv.1709.03339 2018

Show all 15 references

[8]

Monocular Visual Autonomous Landing System for Quadcopter Drones Using Software in the Loop,

M. Saavedra -Ruiz, A. M. Pinto -Vargas, and V. Romero - Cano, “Monocular Visual Autonomous Landing System for Quadcopter Drones Using Software in the Loop,” IEEE Aerosp. Electron. Syst. Mag., vol. 37, no. 5, pp. 2–16, May 2022, doi: 10.1109/MAES.2021.3115208

2022
[9]

Autonomous landing of an Unmanned Aerial Vehicle on an autonomous marine vehicle,

T. K. Venugopalan, T. Taher, and G. Barbastathis, “Autonomous landing of an Unmanned Aerial Vehicle on an autonomous marine vehicle,” in 2012 Oceans, Oct. 2012, pp. 1–9. doi: 10.1109/OCEANS.2012.6404893

2012
[10]

Vision-based Autonomous Landing Control of a Multi-rotor Aerial Vehicle on a Moving Platform with Experimental Validations,

Z. Zhao et al., “Vision-based Autonomous Landing Control of a Multi-rotor Aerial Vehicle on a Moving Platform with Experimental Validations,” IFAC-Pap., vol. 55, no. 3, pp. 1–6, Jan. 2022, doi: 10.1016/j.ifacol.2022.05.001

2022 doi
[11]

Landing system for AR.Drone 2.0 using onboard camera and ROS,

T. Zhao and H. Jiang, “Landing system for AR.Drone 2.0 using onboard camera and ROS,” in 2016 IEEE Chinese Guidance, Navigation and Control Conference (CGNCC) , Aug. 2016, pp. 1098 –1102. doi: 10.1109/CGNCC.2016.7828941

2016
[14]

Reinforcement learning for shared autonomy drone landings,

K. Backman, D. Kulić, and H. Chung, “Reinforcement learning for shared autonomy drone landings,” Auton. Robots, vol. 47, no. 8, pp. 1419 –1438, Dec. 2023, doi: 10.1007/s10514-023-10143-3

2023 doi
[222]

doi: 10.1007/978-3-319-16178-5_14

doi
[2022]

Available: https://patents.google.com/patent/US8996207B2/en

[Online]. Available: https://patents.google.com/patent/US8996207B2/en

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [13]

Deep learning for vision -based micro aerial vehicle autonomous landing,

L. Yu et al., “Deep learning for vision -based micro aerial vehicle autonomous landing,” Int. J. Micro Air Veh. , vol. 10, no. 2, pp. 171 –185, Jun. 2018, doi: 10.1177/1756829318757470

work page doi:10.1177/1756829318757470 2018

[2] [15]

Proactive Guidance for Accurate UAV Landing on a Dynamic Platform: A Visual –Inertial Approach,

C.-W. Chang et al., “Proactive Guidance for Accurate UAV Landing on a Dynamic Platform: A Visual –Inertial Approach,” Sensors, vol. 22, no. 1, Art. no. 1, Jan. 2022, doi: 10.3390/s22010404

work page doi:10.3390/s22010404 2022

[3] [16]

Towards autonomous landing on a moving vessel through fiducial markers,

R. Polvara, S. Sharma, J. Wan, A. Manning, and R. Sutton, “Towards autonomous landing on a moving vessel through fiducial markers,” in 2017 European Conference on Mobile Robots (ECMR) , Sep. 2017, pp. 1 –6. doi: 10.1109/ECMR.2017.8098671

work page arXiv 2017

[4] [17]

A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform,

A. Rodriguez -Ramos, C. Sampedro, H. Bavle, P. de la Puente, and P. Campoy, “A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform,” J. Intell. Robot. Syst. , vol. 93, no. 1, pp. 351 – 366, Feb. 2019, doi: 10.1007/s10846-018-0891-8

doi:10.1007/s10846-018-0891-8 2019

[5] [1]

Learning to Assist Drone Landings,

K. Backman, D. Kulić, and H. Chung, “Learning to Assist Drone Landings,” IEEE Robot. Autom. Lett. , vol. 6, no. 2, pp. 3192 –3199, Apr. 2021, doi: 10.1109/LRA.2021.3062572

arXiv 2021

[6] [3]

Autonomous Approach and Landing for a Low -Cost Quadrotor Using Monocular Cameras,

S. Dotenco, F. Gallwitz, and E. Angelopoulou, “Autonomous Approach and Landing for a Low -Cost Quadrotor Using Monocular Cameras,” in Computer Vision - ECCV 2014 Workshops, L. Agapito, M. M. Bronstein, and C. Rother, Eds., in Lecture Notes in Computer Science. Cham: Springer International Publishing, 2015, pp. 209 –

work page 2014

[7] [4]

Systems and methods for autonomous landing using a three dimensional evidence grid,

E. R. Goossen and Y. Ma, “Systems and methods for autonomous landing using a three dimensional evidence grid,” US8996207B2, Mar. 31, 2015 Accessed: Sep. 05,

work page 2015

[8] [6]

Autonomous Quadrotor Landing using Deep Reinforcement Learning,

R. Polvara et al., “Autonomous Quadrotor Landing using Deep Reinforcement Learning,” Feb. 27, 2018, arXiv: arXiv:1709.03339. doi: 10.48550/arXiv.1709.03339

arXiv doi:10.48550/arxiv.1709.03339 2018