{"id":"8d2fc5c7-9164-4305-95a9-5a79b2517f80","arxiv_id":"2501.09357","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A teaching-learning-based optimization variant with mutation and multi-subject learning plans paths for a three-drone triangular formation and beats GA, TLBO, and θ-PSO on one test case.","lead":"This paper adapts a classroom-inspired optimization algorithm, TLBO, to plan collision-free paths for a group of drones flying in a set formation. The authors add mutation and elite-selection tricks, and test the result in simulations and a small real flight.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The safety claim rests on obstacle-cost evaluation on the centroid path only; individual UAV offsets up to 3 m are never checked against obstacle cylinders.","rationale":"The reader's weakest-assumption analysis identifies exactly the same load-bearing premise: obstacle cost is evaluated on the centroid path, but safety must hold for the three offset UAV paths. I agree with that assessment. The paper's own formation model, Equation (4), makes the offsets explicit, and the experimental offsets are large enough (3 m horizontally) that a centroid-clearing path does not imply vehicle-clearing paths. If the proposed concrete test shows all individual paths clear every obstacle, the safety claim is restored; if not, the central claim of safe operation fails. I do not see a stronger objection: the FTLBO algorithm is plausible, the formation kinematics are standard, and the comparison, though based on limited detail, is not internally inconsistent. The missing fitness weights and single-scenario comparison are secondary weaknesses that also support a conditional verdict, but the centroid-only safety cost is the most decisive issue. Therefore the reader's CONDITIONAL verdict remains appropriate, and no further adjustment is needed.","tokens_in":6892,"tokens_out":2586,"duration_ms":33426,"concrete_test":"Take the FTLBO centroid path from Section 5.2 and generate the three UAV paths as P_n = P_F + Delta P_n for n = 1,2,3. For every obstacle cylinder k and every path segment j of each individual UAV path, compute d_{j,k} and compare it with R_k. If any d_{j,k} <= R_k, the safe-operation conclusion is false. Report the minimum clearance margin across all UAVs and obstacles, and, if possible, repeat this check for the real-UAV validation trajectories.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines the safe-operation fitness using Equations (8) and (9), where d_{j,k} is the distance from obstacle k to path segment j of the formation centroid path q. The actual vehicle paths, however, are P_n = P_F + Delta P_n, with Delta P_2 = [3,0,-1] m and Delta P_3 = [-3,0,-1] m in Section 5.1. A centroid segment can clear a cylinder while an offset UAV path intersects it, because Equation (8) assigns infinite cost only when the centroid path itself penetrates the cylinder and includes no term coupling Delta P_n to obstacle proximity. Consequently, the fitness values in Table 1 and the convergence comparison in Figure 8 do not establish that all three UAVs are collision-free. The real-UAV validation in Section 5.3 is qualitative and reports no measured clearance, so it does not close this gap. This is the weakest load-bearing premise for the paper's central claim of safe operation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses 3D path planning for a three-UAV group that must maintain a triangular formation while performing an inspection task. The formation is modeled by fixing each UAV's position as the group centroid plus a constant offset (Eq. (4)), and the planning problem is cast as the minimization of a weighted fitness function (Eq. (5)) comprising path length (Eq. (7)), obstacle-violation cost on the centroid path (Eqs. (8)-(9)), and an altitude/task cost (Eq. (10)). The authors propose an enhanced teaching-learning-based optimizer (FTLBO) with mutation, elite selection, and multi-subject learning, compare its converged fitness against GA, TLBO, and θ-PSO, and report a qualitative real-UAV validation. The central claim is that FTLBO generates safe, collision-free formation paths at lower cost than the comparison algorithms.","tokens_in":7059,"tokens_out":3651,"duration_ms":37881,"significance":"If fully established, the paper would provide an incremental but practical formation-path-planning baseline: the kinematic formation construction in Eq. (4) is sound, the optimization formulation is transparent, and the inclusion of a real-UAV flight test is a strength that goes beyond pure simulation. The paper also explicitly compares against three metaheuristic baselines and reports convergence statistics. However, the central safety claim is not supported by the obstacle-cost formulation because safety is evaluated only on the centroid path, not on the actual offset UAV paths, and several parameter values needed for reproducibility are omitted. Thus the contribution is a reasonable engineering study whose main claim requires additional verification before it can be accepted.","major_comments":[{"comment":"The obstacle-violation cost d_{j,k} is computed only for the centroid path q. In Section 5.1 the individual UAV positions are P_n = P_F + ΔP_n, with ΔP_2 = [3,0,-1] m and ΔP_3 = [-3,0,-1] m, so a centroid segment that clears a cylinder does not imply that the offset UAV paths are collision-free. Since Eq. (8) assigns infinite cost only when the centroid path itself penetrates a cylinder and contains no term depending on ΔP_n, the fitness values reported in Table 1 and the convergence comparison in Figure 8 do not establish safe operation for all three UAVs. The real-UAV validation in Section 5.3 is qualitative and reports no measured clearance distances, so it does not close this gap. The authors should either compute violation costs for each individual UAV path or equivalently inflate obstacle radii by the maximum formation offset, then rerun the optimization and the comparisons.","section":"Section 2.2, Eqs. (8)-(9)"},{"comment":"The weight coefficients α, β, and γ in the fitness function are never assigned numerical values anywhere in the text, in the experimental setup of Section 5.1, or in Table 1. Because the converged fitness value depends directly on these weights, the reported comparison between FTLBO, GA, TLBO, and θ-PSO is not reproducible, and no sensitivity analysis is provided. The authors should state the exact weight values used in the experiments and ideally report how the ranking of algorithms changes over a modest range of weightings.","section":"Section 2.2, Eq. (5)"},{"comment":"The altitude cost uses h_i^j, the altitude of UAV i at viewpoint j, but the optimization variable q is the centroid path. It is not explained how each UAV's altitude is derived from q and the fixed offsets ΔP_n. Without an explicit link, the task-cost term is either incompletely specified or depends on additional degrees of freedom that are not part of the optimized path. Please define h_i^j operationally in terms of q and ΔP_n, or revise the fitness formulation so that all terms are functions of the same decision variables.","section":"Section 2.2, Eq. (10)"}],"minor_comments":[{"comment":"The statement that the triangular formation is maintained follows directly from the construction: since ΔP_n are constant, Eq. (4) gives the same offsets at every waypoint by definition. This is not a limitation, but the plots in Figures 5-7 should be described as construction checks rather than independent evidence of formation-keeping behavior.","section":"Section 2.1, Eqs. (2)-(4)"},{"comment":"The 'Max cost' column is identical (248.28) for all four algorithms, which is suspicious and likely a formatting or reporting error. Please clarify what this column represents and why the maximum values coincide exactly.","section":"Table 1"},{"comment":"The pseudo-code contains several unclear or corrupted expressions, notably 'while rand < 1˘ξ/ξmax' on line 15 and the comment '/* using (12) */' in the teaching-phase mutation loop, where Eq. (13) appears to be the intended update. The update rule around Eq. (13) and the elite replacement in Eq. (17) should be stated unambiguously in pseudo-code.","section":"Figure 2 and Eq. (13)"},{"comment":"The typeset formula for the path-length cost is garbled; the norm of the segment vector is missing in the rendered text. The equation should be rewritten in a standard vector form so that the cost definition is unambiguous.","section":"Equation (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a modest engineering application of a known optimizer to a formation-path-planning problem. Its main scientific issue is not the optimizer but the safety evaluation: the definition of the safety cost on the centroid path makes the central 'safe operation' claim unverified. This is fixable within the manuscript's scope by reformulating the obstacle cost for individual UAV paths and rerunning the experiments, so I recommend major revision rather than rejection. I saw no signs of misconduct or citation manipulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this is a modest, honest paper, not a breakthrough. The new piece is small but legitimate: the authors take their earlier single-UAV TLBO path planner, add mutation, elite selection, and multi-subject learning, and plan a formation by optimizing a centroid path, then deriving each UAV's path as a fixed offset from that centroid. The formation model (Eq. 4) is sound by construction: if the centroid path is feasible under the fitness function, the triangle shape is preserved by definition. For a reader in the metaheuristic path-planning niche, that is a workable recipe.\n\nWhat it does well: the problem formulation is clean, the pseudo-code is detailed enough to reimplement the method, and the authors actually flew three 3DR Solos along the generated paths. Most papers in this genre stop at simulation. The convergence comparison against GA, TLBO, and θ-PSO on a single scenario shows FTLBO at the lowest cost, which is interesting if not statistically robust.\n\nThe soft spots are real, and one is load-bearing. The safety cost in Eqs. (8)-(9) computes distance from each obstacle to the centroid path segment only. The individual UAVs are offset by up to 3 m horizontally (ΔP2 = [3,0,-1], ΔP3 = [-3,0,-1]), so a centroid path that clears a cylinder does not imply the offset paths do. The paper's claim of \"safe operation\" for all UAVs is therefore not supported by the defined fitness. This is not a subtle math error; it's a mismatch between what the cost measures and what the text claims. Second, the fitness weights α, β, γ in Eq. (5) are never reported, which makes the exact objective indeterminate. Third, the real-UAV validation is qualitative: no measured clearance, no deviation, just \"the paths navigate the UAVs to avoid obstacles.\" Those quantitative gaps would all be fixable in revision.\n\nWho is this for? Anyone working on TLBO variants or formation path planning with metaheuristics who wants another data point. It is not going to change how you plan swarms. But it is not a waste of time either: the formulation is clear, the method is plausible, and the single-scenario comparison, while weak, is at least a fair attempt against three baselines.\n\nMy recommendation: this deserves a serious referee, not a desk reject. The safety-cost flaw is exactly what peer review should catch, and the paper is short enough that a careful reviewer can verify the derivation. Send it out, but expect the authors to need to add obstacle checks on the actual UAV paths and report their weights before I would trust the safety claim.","headline":"A plausible but incremental extension of the authors' own single-UAV TLBO planner to a three-UAV triangle formation, undermined by a safety cost that checks obstacles only on the centroid path, not on the individual UAV offsets.","tokens_in":7621,"tokens_out":1559,"would_cite":false,"duration_ms":48121,"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's central claim is that FTLBO, an enhanced TLBO with mutation, elite selection, and multi-subject learning, produces collision-free triangular paths for a three-UAV survey and beats GA, TLBO, and θ-PSO in cost.","keywords":["UAV swarm","formation path planning","teaching-learning-based optimization","FTLBO","collision avoidance","multi-objective fitness function","orchard monitoring","3D path planning"],"falsifier":"Recompute the obstacle-violation cost for each individual UAV path, not just the centroid path, in the same scene: if any waypoint's offset position lies within distance R_k of a cylinder center, or any segment between offset waypoints intersects a cylinder, the safety claim fails. A direct check is to plot the three offset paths from Figure 5 against the obstacle cylinders of Figure 3 and measure the minimum distances.","tokens_in":6660,"feed_emoji":"🚁","tokens_out":4000,"duration_ms":36380,"temperature":0.7,"pith_summary":"The paper tries to establish that a modified Teaching-Learning-Based Optimization (FTLBO) can solve the path-planning problem for a small UAV swarm that must keep a fixed triangular formation while surveying an area. It treats formation flight as an optimization problem: minimize a fitness function that rewards short paths, low collision risk with cylindrical obstacles, and staying within altitude limits. The authors add mutation, elite selection, and multi-subject learning to TLBO to avoid local optima. In a satellite-based orchard scenario with three UAVs, FTLBO converges to a cost of 94.70, lower than θ-PSO (100.63), TLBO (108.31), and GA (129.04), and the generated paths keep the desired triangle at every waypoint. A reader would care because the approach converts formation geometry into a single centroid path, making planning tractable, and claims feasibility on real UAVs.","feed_headline":"Formation TLBO finds cheaper UAV paths in tests","feed_subtitle":"A mutation- and elite-enhanced TLBO plans collision-free triangular paths for three drones in an orchard survey.","key_machinery":"The carrier of the argument is the FTLBO algorithm: the standard two-phase TLBO (teaching update S_new_i = S_old_i + w0(T − λA), and peer-learning update) augmented with (i) a mutation step using a chaotic logistic sequence z = 2×X_n − 1 with probability μ = 1 − ξ/ξ_max, (ii) an elite strategy that replaces the worst student when the mutated solution improves fitness, and (iii) a multi-subject learning update S_new_{i,j} = S_old_{i,j} + w_{i,j}|S_{i,j} − S_{k,j}|. The formation-to-path reduction is also central: UAV n's position is P_n = P_F + ΔP_n, with constraints d_n = r_F and equal neighbor distances, so planning the centroid path P_F automatically defines all individual paths.","core_discovery":"The central claim is that FTLBO generates valid, collision-free paths for three UAVs flying in a triangle, and that in the reported benchmark it converges to the lowest fitness value among GA, TLBO, and θ-PSO (94.70 after 107 iterations versus 100.63 for θ-PSO after 96). The method works by planning one path for the formation's centroid and deriving each UAV's route from fixed offsets ΔP1, ΔP2, ΔP3, so formation maintenance reduces to tracking the centroid. The fitness function sums path length, obstacle-violation cost (based on distance from path segments to cylinder centers), and altitude/task costs, weighted by coefficients α, β, and γ. The result is tested both in simulation and by uploading planned paths to three 3DR Solo quadcopters via Mission Planner; the authors report that the planned paths are smooth enough for autonomous flight and avoid obstacles.","pith_inferences":["The safety guarantee is only as strong as the assumption that centroid clearance implies offset-path clearance; with offsets up to 3 m, a rigorous planner should check per-UAV clearance or inflate obstacle radii.","The comparison is a single test scenario, so the claimed advantage over θ-PSO likely depends on scene difficulty, obstacle density, and waypoint count; testing on randomized maps would tell.","The multi-subject and chaotic-mutation mechanisms are generic enough to be applied to other metaheuristics such as PSO or GA, so the observed gain may come from the enhancements more than from TLBO itself."],"forward_implications":["In this scenario FTLBO reaches fitness 94.70 in 107 iterations, lower than θ-PSO (100.63), TLBO (108.31), and GA (129.04), so the enhancements do not hurt convergence in this benchmark.","The centroid-path formulation means the same planner can adapt to other formations by changing the ΔP_n offsets, with no change to the core algorithm.","Real-aircraft validation suggests the planned paths are smooth enough to be followed autonomously, indicating practical deployability for surveying tasks.","The fitness function can be reused with different obstacle models by replacing the cylinder-distance cost with other collision metrics."],"supporting_citations":[{"why":"Supplies the base TLBO algorithm with its teaching and learning phases.","marker":"[10]"},{"why":"Previous single-UAV TLBO path planner that this work extends to formation flight.","marker":"[13]"},{"why":"Provides the elite-selection strategy incorporated into FTLBO.","marker":"[15]"},{"why":"Provides the elitist multi-subject learning update used in FTLBO.","marker":"[16]"},{"why":"Source of the θ-PSO baseline and the real-time surface inspection system used in the comparison.","marker":"[18]"}],"fun_headline_variants":["Formation TLBO cuts UAV path cost in tests","Enhanced TLBO finds cheaper safe paths for drone swarm","Mutation-elite TLBO plans triangle paths for UAVs","TLBO variant outperforms GA and PSO on UAV paths","Three drones keep formation with smarter TLBO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that keeping the formation's centroid path away from obstacle cylinders is enough to keep all three UAVs safe, even though the UAVs are offset by up to 3 meters from that centroid.","fun_headline_variants_meta":{"raw":{"variants":["Formation TLBO cuts UAV path cost in tests","Enhanced TLBO finds cheaper safe paths for drone swarm","Mutation-elite TLBO plans triangle paths for UAVs","TLBO variant outperforms GA and PSO on UAV paths","Three drones keep formation with smarter TLBO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1365,"prompt_tokens":839,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":455,"tokens_out":526,"duration_ms":6266,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:05:54.487763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the obstacle-violation cost for each individual UAV path, not just the centroid path, in the same scene: if any waypoint's offset position lies within distance R_k of a cylinder center, or any segment between offset waypoints intersects a cylinder, the safety claim fails. A direct check is to plot the three offset paths from Figure 5 against the obstacle cylinders of Figure 3 and measure the minimum distances.","supporting_citations":[{"cited_title":"Rao, V .J","cited_arxiv_id":null,"evidence_quote":"Supplies the base TLBO algorithm with its teaching and learning phases."},{"cited_title":"Hoang, M.D","cited_arxiv_id":null,"evidence_quote":"Previous single-UAV TLBO path planner that this work extends to formation flight."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the elite-selection strategy incorporated into FTLBO."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the elitist multi-subject learning update used in FTLBO."},{"cited_title":"Hoang, M.D","cited_arxiv_id":null,"evidence_quote":"Source of the θ-PSO baseline and the real-time surface inspection system used in the comparison."}],"review_version":1}