{"id":"23f5fb5a-e2b2-47ba-a088-5d4db284753d","arxiv_id":"2505.06963","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proposes, but does not validate, a reinforcement learning based method for monocular UAV landing using a color-coded lenticular circle landmark.","lead":"This paper proposes a UAV landing method that uses only a front-facing monocular camera and a color-coded circular landmark on the landing pad, with reinforcement learning to estimate altitude and depth. The authors claim simulations show accurate landings, but the manuscript contains no implementation details or verifiable experimental data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central landing claim rests on asserted estimation functions (Eqs. 1–2) and future-tense experiment sections; no reproducible evidence or camera model is provided, so the claimed accuracy is unverifiable.","rationale":"The reader's weakest assumption is that the lenticular circle's color and shape provide sufficient, unambiguous visual information throughout the approach, including the blind area. My analysis agrees: this is precisely the load-bearing condition for Eqs. (1)–(2), and the paper neither derives these functions from a camera model nor supplies the experiments needed to validate them. My independent check—should such a derivation be attempted—would expose whether the claimed accuracy is physically plausible. The reader also flags that the experiments are described in future tense, which is confirmed in Section III; the tables therefore cannot be treated as evidence. I find no basis to alter the reader's REJECT verdict; if anything, the lack of any algorithm description or reproducible code makes the result even less verifiable. This is not a disagreement with the novelty or potential of the idea; it is a statement that the central claim lacks the technical support required for acceptance.","tokens_in":7497,"tokens_out":2278,"duration_ms":26548,"concrete_test":"Build a pinhole-camera simulation of the proposed setup: a circular landmark with three color sectors and known radius, viewed by a front-mounted camera at the approach positions described in Sec. 3.1 (5, 10, 15 m; 0°, 15°, 30°). Render the landmark, extract the apparent ellipse parameters and color-boundary angles, and recover altitude and depth from these visual features using the unstated Q, Q′. Then compare recovered values against ground truth and check invertibility, including poses where the marker is partially or fully out of view. If the recovery errors exceed Table 1's values, or if the mapping is non-invertible in the final approach phase, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that a front-facing monocular camera plus a lenticular-circle landmark yields accurate altitude and depth estimates for autonomous landing—depends entirely on the validity of the mapping Altitude = Q(D1) and Depth = Q′(θ) in Eqs. (1)–(2). These equations are asserted without specifying the optical model, the landmark's physical dimensions, the color-encoding scheme, or the camera projection equations. Monocular pose estimation from a circle has known ambiguities (scale/depth coupling, ellipse-to-pose multiplicity), yet no uniqueness or sensitivity argument is given. The experimental evidence is also not anchored: Section III describes tests in future tense ('The tests are conducted', 'the UAV is tasked'), and Tables 1–2 present numerical errors without any description of the simulation environment, the RL policy architecture, the training procedure, or the sensor noise model. The paper itself acknowledges a blind area in Figure 2 during the final approach but offers no mechanism to estimate altitude or depth when the marker leaves the front camera's field of view. Consequently, the claim that the UAV 'can accurately estimate altitude and depth' (Sec. 3.3) and 'achieve precise and reliable landings' (Sec. IV) is not supported by any reproducible experiment or closed-form derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7769,"tokens_out":2054,"duration_ms":21055,"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":[{"comment":"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.","section":"II.2.1 (Equations 1 and 2)"},{"comment":"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.","section":"III (Tables 1 and 2)"},{"comment":"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.","section":"III.3 and IV (Future Work)"}],"minor_comments":[{"comment":"The text contains a duplicated word: 'the UAV may may not always have a clear line of sight' should be corrected to 'may not.'","section":"I. Introduction"},{"comment":"References [12] and [13] are identical, as are references [5] and [16] and references [15] and [17]. The reference list should be deduplicated.","section":"References"},{"comment":"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.","section":"II. Methodology"},{"comment":"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.","section":"II.2.1 and II.2.2"},{"comment":"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.","section":"III.1 and III.2"},{"comment":"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.","section":"III.3"}],"recommendation":"reject","confidential_remarks":"The manuscript is best understood as a method proposal, yet it is framed as a validated experimental study. The core estimation equations are placeholders, and the reported tables are not backed by any reproducible protocol. The authors' own future-work section concedes that real-world testing remains to be done, which conflicts with the abstract and conclusion. The paper would need a complete rewrite with a formal optical model, a real RL training description, and credible experiments before it could be considered for publication. The excessive self-citation overlap in the reference list also needs editorial attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat'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.\n\nWhat 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.\n\nThe 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.\n\nSo 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.\n\nWho 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.","headline":"A plausible marker idea buried in a paper that claims validation without providing any: equations are placeholders, experiments are future tense.","tokens_in":8258,"tokens_out":2246,"would_cite":false,"duration_ms":21921,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A drone can be taught to land using only a front-facing camera and a color-shifting circular marker.","keywords":["UAV autonomous landing","monocular vision","reinforcement learning","image-based visual servoing","lenticular landmark","shared autonomy","depth estimation","landing pad detection"],"falsifier":"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.","tokens_in":7313,"feed_emoji":"🛬","tokens_out":7246,"duration_ms":68959,"temperature":0.7,"pith_summary":"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.","feed_headline":"One printed circle could replace depth sensors for drone landings","feed_subtitle":"No GPS, no downward camera, no depth sensor—just a front camera reading a color-shifting marker.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Establishes that a low-cost quadrotor can approach and land using monocular cameras, which is the feasibility baseline the proposed method extends.","marker":"[3]"},{"why":"Shows a monocular vision landing system for quadcopters validated in software in the loop, supporting the paper's claim that camera-only landing is realistic.","marker":"[8]"},{"why":"Demonstrates a deep reinforcement learning strategy for landing on a moving platform, the direct precedent for using reinforcement learning to learn landing policies.","marker":"[7]"},{"why":"Provides a reinforcement-learning shared-autonomy landing method whose human–AI arbitration idea the paper adapts.","marker":"[14]"},{"why":"Supplies a visual-inertial baseline for landings on dynamic platforms, against which the proposed dynamic-condition performance is framed.","marker":"[15]"},{"why":"Shows fiducial-marker-based landing on a moving vessel, supporting the use of a visual landmark for precision approach.","marker":"[16]"},{"why":"Shows deep reinforcement learning applied to autonomous quadrotor landing, a background for the choice of learning algorithm.","marker":"[6]"},{"why":"Provides evidence that deep learning can drive vision-based micro aerial vehicle landing, supporting the learning-from-visual-input approach.","marker":"[12]"}],"fun_headline_variants":["A single printed ring guides drone landings without depth sensors","Drone lands using only a monocular camera and a printed pattern","Reinforcement learning reads a circle pattern to land drones safely","No depth sensor needed: RL interprets a lenticular circle for landing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["A single printed ring guides drone landings without depth sensors","Drone lands using only a monocular camera and a printed pattern","Reinforcement learning reads a circle pattern to land drones safely","No depth sensor needed: RL interprets a lenticular circle for landing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000641,"raw_usage":{"total_tokens":2919,"prompt_tokens":880,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1967}},"tokens_in":496,"tokens_out":2039,"duration_ms":15715,"temperature":1.0,"reasoning_tokens":1967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:27:49.678942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Autonomous Approach and Landing for a Low -Cost Quadrotor Using Monocular Cameras,","cited_arxiv_id":null,"evidence_quote":"Establishes that a low-cost quadrotor can approach and land using monocular cameras, which is the feasibility baseline the proposed method extends."},{"cited_title":"Reinforcement learning for shared autonomy drone landings,","cited_arxiv_id":null,"evidence_quote":"Provides a reinforcement-learning shared-autonomy landing method whose human–AI arbitration idea the paper adapts."},{"cited_title":"Proactive Guidance for Accurate UAV Landing on a Dynamic Platform: A Visual –Inertial Approach,","cited_arxiv_id":null,"evidence_quote":"Supplies a visual-inertial baseline for landings on dynamic platforms, against which the proposed dynamic-condition performance is framed."},{"cited_title":"Towards autonomous landing on a moving vessel through fiducial markers,","cited_arxiv_id":null,"evidence_quote":"Shows fiducial-marker-based landing on a moving vessel, supporting the use of a visual landmark for precision approach."}],"review_version":1}