{"id":"164f270a-38e5-465a-89e0-0696fcf351bf","arxiv_id":"2607.17771","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A polar-coordinate differential evolution algorithm plans UAV search paths and reports higher model-based detection probabilities and lower runtimes than five generic metaheuristics in simulations.","lead":"This paper proposes a polar-coordinate differential evolution algorithm (PDE) for planning drone search paths that maximize the model-based chance of detecting a moving target. The method is computationally simple and produces flyable paths, but the 'state-of-the-art' claim rests on comparisons with generic metaheuristics and one favorable field run.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 21's J is not the true model detection probability: the belief is never conditioned on no-detection events, so Table 4 may not support the headline claim.","rationale":"The central claim is that PDE outperforms in detection probability. The metric used to support this is J, Eq. 21. For that metric to be meaningful, the Bayesian recursion in §2.3 and the Gaussian belief in §2.1 must be consistent. They are not: after no-detection events the posterior is not Gaussian, yet the paper keeps σ_k constant and never describes the non-Gaussian update. This makes J not the true detection probability of the planned path. The missing-baseline issue is also real, but it would only affect the breadth of the comparison; the belief-update flaw affects the correctness of the reported detection probability itself. The proposed grid-based histogram test directly computes the exact J under the paper's own model and would show whether the discrepancy matters. If the discrepancy is small, the concern is resolved and the CONDITIONAL verdict can stand; if large, the abstract's claim should be revised or withdrawn. The paper's source code availability is a point in its favor, as it makes the test feasible.","tokens_in":15376,"tokens_out":9532,"duration_ms":111259,"concrete_test":"Implement a grid-based histogram filter for the six scenarios: initialize with the Gaussian belief; at each waypoint shift the grid according to Eq. (13); multiply the belief by 1 − p(D_k | ξ) inside the camera footprint (computed from Eqs. 7–10); renormalize; compute r_k from Eq. 17 and J from Eqs. 18–21. Do this for the PDE paths shown in Figure 4 and for the comparison algorithms' best paths (or re-run each optimizer with this exact fitness). If the exact J values differ from Table 4 by more than 10%, or if the algorithm ranking changes, the reported detection-probability advantage is an artifact of the open-loop belief. The code is available, so this test is straightforward.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 defines a recursive Bayesian filter: prediction (Eqs. 12–14) and update (Eq. 15) condition the target-location belief on observations ζ_{1:k}. Section 2.4 then defines the cumulative detection probability J (Eq. 21) via r_k = p(\\bar{D}_k|ζ_{1:k−1}) (Eq. 17), which is the probability of no detection at step k given all previous observations. For J to equal the actual detection probability of the planned path, p(ξ^t_k|ζ_{1:k−1}) in Eq. 17 must be the posterior after conditioning on all previous no-detections. But Section 2.1 fixes the belief as a single Gaussian with constant covariance σ_k (Eq. 1), and Algorithm 1 contains no operation that would produce the non-Gaussian posterior that results from multiplying the predicted belief by (1 − detection likelihood) at each visited footprint (Eq. 15 with ζ_k = \\bar{D}_k). The constant-σ Gaussian is thus the open-loop prior, not the conditional belief. Consequently r_k and J overstate the probability of detection, especially for paths that revisit searched regions, and the Table 4 comparison—all algorithms evaluated on this same J—does not establish the abstract's claim about detection probability even within the paper's own target model. This is an internal inconsistency, separate from the acknowledged fragility of the target-motion model in Section 4.4.2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a path-planning method, polar coordinate-based differential evolution (PDE), for a UAV searching for a moving target. The target is modeled as a deterministic Markov process with an initial Gaussian belief (Eq. 1); the camera footprint is derived geometrically; the detection likelihood is obtained from the NVESD model (Eq. 7). A recursive Bayesian prediction-update formulation is written down (Eqs. 12–15), and the cumulative detection probability J (Eq. 21) is used as the fitness function. The paper then encodes candidate paths in polar coordinates (step length and yaw angle), applies DE mutation/crossover/selection directly to this representation, and reports simulation comparisons with standard DE, GA, PSO, GWO, and GSK on six scenarios, plus one real-world experiment with a 3DR Solo. The abstract claims that PDE outperforms state-of-the-art algorithms in detection probability and execution time while remaining practical, and the source code is publicly released.","tokens_in":15591,"tokens_out":4695,"duration_ms":64558,"significance":"If the claims were fully supported, the contribution would be a practically useful encoding for evolutionary search planning: polar coordinates make UAV kinematic constraints simple to enforce and reduce the trigonometric overhead of evaluating footprints. The real experiment and the public code are genuine assets, and the paper is clearly written. However, the headline detection-probability claim rests on an internally inconsistent treatment of the Bayesian belief update: the fitness J is computed from an open-loop prior, not from the posterior conditioned on the candidate path's own no-detection events. In addition, the 'state-of-the-art' comparison omits the search-specific algorithms the paper itself cites (motion-encoded PSO, ACO, Bayesian optimization, greedy), and the execution-time advantage is not supported by statistical tests. With a corrected belief update and a more relevant baseline comparison, the central algorithmic idea could still be a valid contribution.","major_comments":[{"comment":"The derivation defines the Bayesian update in Eq. (15), but the implemented fitness J never applies this update for the no-detection observations that the candidate path itself would produce. In Eqs. (17)–(21), r_k is evaluated using p(ξ^t_k|ζ_{1:k−1}) from the prediction step, yet Algorithm 1 contains no operation that conditions the belief on the visited footprints and the corresponding 'no detection' events. The belief used in the fitness evaluation remains the constant-variance Gaussian from Eq. (1), propagated open-loop. Consequently J is not the detection probability of the planned path under the paper's own target and sensor model; it is a surrogate that is generally optimistic, especially for paths that revisit previously searched areas. Table 4 therefore does not establish the abstract's claim about detection probability. The authors should either implement the correct Bayesian","section":"§2.3–2.4, Eq. (21), Algorithm 1"},{"comment":"The comparison is against generic metaheuristics (DE, GA, PSO, GWO, GSK), not against the search-specific planners that the paper itself identifies as the state of the art in Table 1: motion-encoded PSO [12], ACO [13], Bayesian optimization [18], and greedy search [16]. Since the claimed contribution is to moving-target search rather than to generic DE, Table 4 at most shows that PDE beats these generic baselines on the (open-loop) surrogate J. A comparison with at least one dedicated search planner is necessary to support the 'state-of-the-art' wording.","section":"§4.3, Tables 4–5"},{"comment":"The paper states that PDE is fastest in 5/6 scenarios, but the reported times do not clearly support that statement. For example, in Scenario 1 PDE is 246±16 s vs GA 249±33 s; in Scenario 3 PDE is 516±8 s vs GA 505±46 s; in Scenario 5 PDE is 907±85 s vs GA 956±57 s — differences are within or comparable to the reported standard deviations. No statistical test (e.g., paired t-test or Wilcoxon) is given. The execution-time advantage should be either supported with an error-bar analysis or restated as 'comparable to' the other Cartesian-coordinate algorithms.","section":"§4.3, Table 5"},{"comment":"The scaling factor F is sampled from [0.1, 0.4], which the authors acknowledge is smaller than the standard DE range, and no sensitivity study is reported for this choice, nor for p_Cr and N_p. Since the PDE's advantage over standard DE could depend on this specific hand-tuning, the robustness of the conclusions to the parameter setting should be demonstrated, or at least a small sensitivity analysis should be added.","section":"§4.1, parameters"}],"minor_comments":[{"comment":"The mapping in Eq. (30) is denoted f: U → ξ, but f was already used for the fitness function in Eqs. (21) and (24). This double use of f is confusing; use a different symbol for the coordinate transformation.","section":"§3.2, Eq. (30)"},{"comment":"The table entries have inconsistent formatting, e.g., '0.0867±003' in Scenario 6 and '0.3215±0.011' in Scenario 4; standard errors or standard deviations should be labeled explicitly and formatted consistently.","section":"§4.3, Table 4"},{"comment":"There is a duplicated sentence fragment: 'while the heading angle is used directly in the footprint evaluation without requiring additional coordinate conversion. while the heading angle is used directly...' The sentence should be rewritten.","section":"§4.3, text after Table 5"},{"comment":"The sentence 'Note that the detection event happened in this experiment is due to the alignment of our setup with the assumptions used in theory... hence there is no guarantee that the UAV will find the target' is an honest limitation, but it also underscores that the field test validates path feasibility rather than detection performance. This caveat should be moved to the experimental-design section.","section":"§4.4.2"},{"comment":"The belief is described as a Gaussian with constant σ_k, but the target is also modeled as following a deterministic known transition (Eqs. 13–14). It would help to state explicitly whether σ_k is meant only to represent initial localization uncertainty rather than motion uncertainty, since under a deterministic transition a Gaussian prior remains Gaussian with the same covariance.","section":"§2.1, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The core issue is not the polar-coordinate encoding, which is plausible and well motivated, but the mismatch between the Bayesian theory in Section 2.3 and the objective actually optimized in Algorithm 1. If the authors fix this by implementing a proper no-detection update, the reported numerical values will change and the comparison may change, so the paper cannot be accepted without this revision. The omission of search-specific baselines also weakens the claimed significance. The field experiment is a strength but should be framed as a feasibility demonstration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know before reading: the polar-coordinate encoding for DE is the real contribution here. It directly enforces UAV step-length and yaw limits, keeps candidate paths feasible, and cuts trigonometric overhead. The code is public, the simulation comparison is honestly reported as 10-run averages, and there is no sign of parameter fitting to game the results. The real flight demonstrates that a generated path can be tracked, which is a useful sanity check. For those reasons the paper is worth a look.\n\nThe soft spot is load-bearing. The stress-test note is correct on reading. Section 2 defines a recursive Bayesian filter with an update step that conditions on observations, but Algorithm 1 never applies that update. The fitness J in Eq. 21 is computed using the open-loop predictive density from the initial Gaussian and the deterministic target transition. The belief is never multiplied by the no-detection likelihood along the candidate path. So r_k and R_k are not the true no-detection probabilities, and J overstates the probability of detection, especially for paths that revisit searched areas. Comparing algorithms on this J does not support the abstract's claim about detection probability. This is an internal inconsistency between the stated model and the implementation, not just an external limitation of the target-motion assumption.\n\nAlso missing is the most relevant baseline: the same group's motion-encoded PSO [12]. Comparing against generic DE/GA/PSO/GWO/GSK is fine, but it does not establish \"state-of-the-art\" for this problem class. The single real trial is explicitly acknowledged as aligned with the assumptions, so it is a feasibility demonstration, not detection-rate evidence. The hand-tuned scaling factor F without sensitivity analysis is a minor issue.\n\nWhat holds up: the polar encoding is clean, the path-feasibility argument is sound, and there are no invented entities or hidden circularity. For a researcher working on metaheuristic encodings for UAV search, this paper is a useful data point.\n\nMy recommendation: send it to peer review. The core idea is plausible, the code is available, and the main flaw is fixable by conditioning the belief on no-detection events or by running Monte Carlo detection simulations. The missing baseline is easy to add. A competent referee could turn this into a solid paper after major revision.","headline":"The polar-coordinate DE encoding is a useful, practical trick, but the paper's headline claim about detection probability is undercut by an open-loop belief that is never conditioned on no-detection observations.","tokens_in":16187,"tokens_out":2577,"would_cite":false,"duration_ms":33239,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes encoding UAV search paths in polar coordinates — step length and yaw angle — in differential evolution, and claims this raises cumulative detection probability, cuts execution time, and keeps paths flyable.","keywords":["UAV target search","polar coordinates","differential evolution","Bayesian search","moving target","NVESD detection model","path planning","cumulative detection probability"],"falsifier":"Run repeated real or simulated searches where the target's initial position is offset from the belief-map center by, say, one sigma, and where its heading changes at random intervals; compare realized detection frequency with the J value the planner predicts. If realized detection is near zero while predicted J is high, the central claim collapses. A cheaper check is to benchmark PDE against a polar-encoded version of any other metaheuristic; if the speed advantage disappears, the benefit comes from the encoding alone, not from the algorithm.","tokens_in":15152,"feed_emoji":"🚁","tokens_out":7433,"duration_ms":79050,"temperature":0.7,"pith_summary":"The paper tries to establish that encoding a UAV's search path as a sequence of polar coordinates — a step length and a yaw angle per segment — rather than as Cartesian waypoint coordinates lets a differential evolution (DE) planner maximize the cumulative probability of detecting a moving target more effectively and more quickly than standard metaheuristics. The authors derive a Bayesian fitness function J that combines a Gaussian belief map of the target's location, a deterministic Markov transition model for its motion, the geometry of the camera footprint, and the NVESD model of human target detection from imagery. The proposed polar DE (PDE) applies UAV kinematic limits as simple bounds on the step length and yaw angle, narrowing the search to flyable paths and avoiding repeated trigonometric conversions during fitness evaluation. Across six simulated scenarios, PDE reports the highest detection probability in five and the shortest execution time in five compared with standard DE, genetic algorithm, particle swarm optimization, grey wolf optimization, and gaining-sharing knowledge; a real quadcopter flight confirms the planned path can be followed. If this holds, the practical consequence is that a modest change in search representation can improve both hit probability and runtime for vision-based UAV search and rescue.","feed_headline":"Polar-coordinate evolution improves UAV detection in 5/6 scenarios","feed_subtitle":"Encoding UAV paths as step lengths and yaw angles cuts trig overhead and keeps paths flyable, raising hit probability.","key_machinery":"The central mechanism is the polar-coordinate path encoding and its direct mapping to UAV kinematics. A path is stored as step lengths bounded by speed limits and yaw angles bounded by turn limits, then converted to Cartesian via the recurrence above. This has two effects: kinematic and physical constraints are enforced by clamping scalars rather than by solving constraint equations, and the camera footprint calculation receives the yaw angle directly, eliminating atan2 and square-root conversions needed when starting from raw x,y coordinates. The objective being maximized is the cumulative first-detection probability J, built from the Markov predict-update Bayesian recursion and the NVESD d","core_discovery":"The central claim is that the polar-coordinate representation of the flight path is the key improvement: each candidate path is encoded as a sequence of step lengths and yaw angles, which maps to Cartesian waypoints through a recurrence using cosine and sine of the yaw angle. The mutation and crossover of differential evolution then operate directly on these feasible control parameters, and UAV speed and yaw limits become variable bounds instead of post-hoc constraint checks. The same fitness function J, defined as the sum over time of the probability of first detecting the target (computed by Bayesian prediction-update with the NVESD detection likelihood), is used. The paper claims PDE outp","pith_inferences":["Untested extension: the reported speedup likely transfers to other metaheuristics. If the real cost is repeated atan2 and sqrt in Cartesian representation, a polar-encoded PSO or GWO should show similar runtime gains, a benchmark the paper does not run.","The detection-probability advantage may be sensitive to target model mismatch. Since the field experiment succeeded only because the target started at the last known location and moved along the predicted heading, a natural test is to run PDE with perturbed initial positions or turning targets and compare realized detection frequency against the model's predicted J.","The constant sigma_k assumption in the belief update is a deliberate simplification; adapting sigma_k over time would extend the model to stochastic motion and might change the relative ranking of planners, which is a testable variant of the algorithm.","A quick falsifying experiment: use a polar-encoded version of the same DE with a different initial population or a simpler search pattern (e.g., lawnmower) and compare J; this would separate the contribution of polar encoding from the contribution of the DE evolutionary search itself."],"forward_implications":["Any Cartesian-space evolutionary search for UAV paths can be re-expressed in polar coordinates; the comparison shows standard DE in polar form substantially outperforms Cartesian DE, so the representation itself is a transferable improvement.","Because speed and yaw limits are encoded as simple bounds, the optimizer's output is physically feasible without a separate path-smoothing or feasibility step, as the real flight path overlap demonstrates.","The derived fitness J provides a concrete, comparable objective for moving-target search that couples sensor geometry, target dynamics, and human observer performance; future planners can optimize this same objective with different solvers.","Detection probability degrades as map size grows (from about 0.53 to 0.10 across the six scenarios), indicating that a single UAV's detection ceiling is limited; the paper's proposed direction of multi-UAV coordination is a direct corollary of these results."],"fun_headline_variants":["Polar-coordinate evolution improves UAV search detection","UAV search gets a boost from polar-coordinate DE","Polar-coordinate encoding yields flyable, efficient UAV paths","New PDE algorithm finds targets faster on UAVs","Polar-coordinate DE: superior UAV search in tests"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the premise that the target moves as a deterministic Markov process with a known transition and that the Gaussian belief map with constant sigma_k captures all residual uncertainty — if the target does not start where the belief says or does not follow the assumed heading, the maximized fitness J is not the true detection probability, and the paper concedes there is no guarantee the UAV will find the target.","fun_headline_variants_meta":{"raw":{"variants":["Polar-coordinate evolution improves UAV search detection","UAV search gets a boost from polar-coordinate DE","Polar-coordinate encoding yields flyable, efficient UAV paths","New PDE algorithm finds targets faster on UAVs","Polar-coordinate DE: superior UAV search in tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000725,"raw_usage":{"total_tokens":3093,"prompt_tokens":755,"completion_tokens":2338,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2260}},"tokens_in":499,"tokens_out":2338,"duration_ms":19704,"temperature":1.0,"reasoning_tokens":2260,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:03:11.495514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run repeated real or simulated searches where the target's initial position is offset from the belief-map center by, say, one sigma, and where its heading changes at random intervals; compare realized detection frequency with the J value the planner predicts. If realized detection is near zero while predicted J is high, the central claim collapses. A cheaper check is to benchmark PDE against a polar-encoded version of any other metaheuristic; if the speed advantage disappears, the benefit comes from the encoding alone, not from the algorithm.","supporting_citations":[],"review_version":1}