{"id":"19ca4a49-6a68-45eb-bde9-3b950bec17a9","arxiv_id":"1908.09236","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A centralized framework assigns goals and plans globally deconflicted, kinodynamically feasible paths for multiple UAVs, maintaining persistent coverage with graceful degradation when resources are insufficient.","lead":"This paper combines existing robot planning pieces, goal assignment, prioritized planning, and kinodynamic path planning, into one system that keeps multiple drones repeatedly covering an area whose coverage needs decay over time. The authors show the system works in simulation and on real drones, and that checking collisions across whole planned paths costs little extra compute.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world tracking error can break the time-stamped deconfliction invariant between planning cycles; the paper defers this to future work, so the real-world collision-free claim is unestablished.","rationale":"The reader's weakest assumption is exactly the one I consider load-bearing, so agreement_with_reader is 'agree'. The paper's contribution is an integration of goal assignment with globally deconflicting kinodynamic planning, and its headline evidence is that global deconfliction costs little and that coverage remains controlled in real flights. Both of those claims rest on the planner's committed plans being faithful predictors of future robot positions. The paper explicitly acknowledges that real UAVs deviate from planned paths and defers robust handling to future work. In a time-indexed deconfliction scheme, even a small timing error can eliminate a planned separation buffer; the reported stopped-time fraction of 7.12% makes this concrete. The missing piece is not a proof detail but an empirical quantity: actual inter-UAV separation and tracking error during the real-world runs. Without that, the 'guarantees collision-free UAV movement' claim is not supported for the real-world setting. I do not see a deeper internal inconsistency: the planning loop is coherent and the simulation claims are credible under perfect execution. Thus the reader's CONDITIONAL verdict is appropriate, and my stress test does not move it. Accepting the paper as is would require either adding tracking-error and separation data or softening the real-world guarantee to 'collision-free in simulation, collision-free in practice only when execution error is negligible'.","tokens_in":10916,"tokens_out":5595,"duration_ms":60858,"concrete_test":"Re-run the two-UAV small-map experiment (or replay its logged data, if retained) with differentially corrected GPS/RTK ground truth at at least 10 Hz. For each of the 8 runs, compute the minimum distance between actual UAV positions over time and the per-UAV tracking error (norm of actual minus committed-plan position, including timing offset) at each timestamp. Report the minimum pairwise distance and the maximum tracking error, especially during the stopped periods. If any pairwise distance is below dmin = 10 m, or if the tracking error exceeds the planned separation margin at a time when another committed plan is nearby, the deconfliction invariant is violated between planning cycles and the real-world collision-free claim fails; if all distances stay above dmin with margin, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's deconfliction invariant (Sec. 1, Sec. 4.2) treats each committed plan as a time-stamped trajectory that the UAV will execute exactly. The Goal Planner prunes transitions that collide 'in space or time' with other committed plans, so validity depends on each UAV staying on its committed plan on schedule. In the real world, the paper concedes that 'UAVs can deviate from their planned paths' and that 'accounting for large deviations from planned paths requires replanning and is part of future work' (Sec. 5). Between planning cycles, the Prioritized Planner only clips executed portions using the latest U_loc (Alg. 1, lines 4-6); it does not re-time or reconstruct the remaining committed plans to match the UAV's actual state. A lateral or temporal deviation, including the stopping maneuvers that occur up to 7.12% of mission time in Table 2, can therefore make the executed trajectories come closer than dmin = 10 m even when all committed plans are pairwise deconflicted. The real-world evaluation reports criticality timing but not tracking error or minimum achieved pairwise distance, so the central claim of collision-free real-world operation is not actually demonstrated. The architecture may be repairable by receding-horizon replanning, so this is a conditional concern rather than a rejection of the framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a centralized planning framework for persistent multi-UAV coverage. In each planning cycle, a Prioritized Planner iterates over the UAVs in round-robin order; a Goal Assigner selects the next cell via a multi-goal Dijkstra search with pseudo-edge costs encoding cell urgency, and a Goal Planner computes a kinodynamically feasible path on a state lattice using weighted A*. The planner maintains committed plans of bounded duration and enforces global deconfliction by pruning any candidate trajectory that intersects another UAV's committed plan in space or time. The framework is evaluated in simulation and on real UAVs, with dynamic UAV removal, and the authors claim graceful coverage (average criticality below one when enough UAVs are available) and a negligible computational price for global deconfliction compared with weaker local collision-checking schemes.","tokens_in":11167,"tokens_out":5855,"duration_ms":58239,"significance":"If the claims are correct, the paper offers a practical integration of goal assignment, kinodynamic planning, prioritized multi-robot planning, and inter-robot deconfliction for persistent coverage, and the real-world experiments with dynamic UAV removal are valuable. The architecture is clearly described (Alg. 1), the criticality metric is defined independently rather than fitted, and the authors make falsifiable claims about average criticality and planning times. However, the empirical base is thin in places, and the central real-world claim of collision-free operation is not directly demonstrated because tracking error between planning cycles is not measured and is explicitly deferred to future work. These issues are repairable, but they affect load-bearing statements in the paper.","major_comments":[{"comment":"The paper's global-deconfliction invariant (Sec. 1 and Sec. 4.2) is time-stamped: a newly planned trajectory is pruned only if it intersects another committed plan 'in space or time.' The real-world experiments, however, rely on the assumption that executed trajectories follow committed plans. The paper itself states that 'UAVs can deviate from their planned paths' and that 'accounting for large deviations from planned paths requires replanning and is part of future work' (Sec. 5). Alg. 1 (lines 4-6) only clips executed portions using the latest U_loc; it does not re-time or reconstruct the remaining committed plans from the actual UAV state. A lateral or temporal deviation, including the stopping maneuvers measured at up to 7.12% of mission time (Table 2), can therefore bring two executed trajectories closer than dmin = 10 m even when all committed plans are pairwise deconflicted. Because Sec. 5.3 does not report tracking error or minimum achieved pairwise distance, the claim that the framework provides collision-free real-world operation is not actually demonstrated. I ask the authors either to report per-run tracking-error and inter-UAV-distance statistics, or to implement a receding-horizon replan that re-times committed plans from the observed state, before claiming that the P2 hard constraint is satisfied in the real world.","section":"Sec. 5.3 / Sec. 4.2, Alg. 1"},{"comment":"The quantitative claims rest on very few executions. Table 1 reports a single simulated run per configuration, Fig. 6 explicitly states that each curve is a single mission execution with no confidence intervals, and Table 2 gives averages over eight real-world runs without variance or confidence intervals. The prose draws strong conclusions from these numbers—for example, that global deconfliction 'incurs a negligibly higher price' and that average criticality is kept below one. With a single run, an observed difference in goal-planning time or stopped time cannot be distinguished from run-to-run variance. I request multiple independent runs with reported variance (or at least min/max) for the headline metrics (average criticality, tGP, tstopped), and, for the real-world results, a per-run breakdown of criticality and stopping time.","section":"Sec. 5.2, Table 1, Table 2, Fig. 6"},{"comment":"Property P3 (Sec. 3) is stated as a per-cell guarantee: 'the age of each cell is smaller or equal to its lifetime.' The evaluation in Sec. 5.2, however, uses only the spatial average C̄_t (Eq. 1), and Fig. 5 (right) shows that a non-negligible fraction of cells expire even in runs that are otherwise described as successful (spikes reaching roughly 8% at the 5/10/15-minute lifetime boundaries). An average below one is compatible with systematic violations of P3 for particular zones, so the statement that the framework 'maintains persistent coverage' given sufficient UAVs should be qualified as maintaining low average criticality, not guaranteeing per-cell persistence. Either report the per-cell violation statistics or soften the wording accordingly.","section":"Sec. 3, Sec. 5.2, Eq. (1)"}],"minor_comments":[{"comment":"Typo: 'enviornment' should be 'environment.'","section":"Abstract"},{"comment":"The reference to 'Eq. (5.1)' should be 'Eq. (1)' (or 'Sec. 5.1').","section":"Sec. 5.2"},{"comment":"The caption states that tstopped is the maximum over all runs while the other columns are averages; please clarify this mixed reporting in the table caption or in the text.","section":"Table 2"},{"comment":"The caption refers to a black curve with confidence intervals for the simulated experiments but does not state how many simulation runs were used to form those intervals; please specify.","section":"Fig. 8(b)"},{"comment":"The description of the collision-checking schemes would be clearer if the meaning of 'Quarter, Half and Full Plan (static)' were defined in the text rather than only in the figure caption, and if the color/line style correspondence were stated explicitly in the caption.","section":"Sec. 5.2, Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The novelty claim that this is the first framework to answer goal assignment, kinodynamic planning, prioritized planning, and collision avoidance 'in unison' is hard to verify from the related-work section alone; I would encourage the editor to ask the authors for a more systematic comparison in revision. The tracking-error concern raised in the stress-test is real and is acknowledged in the manuscript itself, so it should be addressed head-on rather than deferred. The thin empirical base (single runs in Table 1 and Fig. 6) is acceptable for a conference paper but is marginal for a journal article; the framework itself appears sound and repairable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe take-home: this is a solid engineering integration paper, not a conceptual breakthrough. It combines goal assignment, prioritized planning, state-lattice kinodynamic planning, and committed-plan deconfliction into one loop and demonstrates it on real UAVs. That combination, in a real-world persistent-coverage setting, genuinely is new relative to the cited OP/VRP and coverage literature. The paper is honest about that: it claims 'first to answer all of these questions in unison,' and I think that claim holds.\n\nWhat it does well: the problem formulation is clean, especially the criticality measure (Eq. 1) and the committed-plan mechanism that bounds how far plans extend into the future. The empirical evaluation is appropriate in scope: simulation and real-world runs, including dynamic removal of a UAV. The comparison against weaker collision-checking schemes is the right kind of ablative study.\n\nThe soft spots are real but not damning. The first is execution uncertainty. The deconfliction invariant is defined on committed plans, not actual trajectories. The paper admits UAVs deviate from planned paths and that accounting for large deviations is future work (Sec. 5). That means the claim of collision-free real-world operation is not actually demonstrated; the experiments don't report tracking error or minimum achieved pairwise distance. This is a limitation, not a fatal flaw—the architecture likely survives by replanning faster or using a safety filter—but the paper should present it that way. The second soft spot is statistical: Fig. 6 shows single runs with no confidence intervals, Table 1 is single runs, and Table 2's 7.12% stopping time is a max without spread. For a system paper, this is minor but worth addressing.\n\nNo code or data is released, which makes independent verification of the real-world claims hard. I don't see this as a rejection reason, just a reproducibility gap.\n\nWho it's for: people building practical multi-robot coverage systems, and reviewers who want a concrete reference point for what a complete persistent-coverage stack looks like. It deserves serious peer review; the revision should tighten the claims around execution-level safety and add basic statistical reporting. I'd take it.","headline":"Solid engineering integration for persistent multi-UAV coverage; the real-world collision-free claim is not fully demonstrated due to unaccounted execution deviation, but the framework is honest, useful, and deserves review.","tokens_in":11720,"tokens_out":2114,"would_cite":true,"duration_ms":21189,"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":"This paper presents a centralized planning framework that maintains persistent multi-UAV coverage with globally deconflicted, kinodynamically feasible paths in real time.","keywords":["persistent coverage","multi-UAV planning","global deconfliction","prioritized planning","kinodynamic planning","coverage criticality","state-lattice planning","frontier-based goal assignment"],"falsifier":"In a real-world run, record each UAV's actual trajectory alongside its committed plan; if any two UAVs ever come within the dmin = 10 m threshold despite their committed plans being deconflicted, or if the measured deviation from the planned path exceeds the planner's assumed buffer, the central claim fails. A direct simulation stress test would deliberately inflate tracking noise until the first inter-UAV distance drops below dmin.","tokens_in":10720,"feed_emoji":"🚁","tokens_out":4491,"duration_ms":42919,"temperature":0.7,"pith_summary":"The paper sets out to solve persistent multi-UAV coverage—keeping a fleet repeatedly visiting an area whose coverage quality decays over time—as a single practical planning problem rather than a collection of disconnected subroutines. It claims to be the first framework that runs goal assignment, kinodynamic path planning, prioritized multi-robot planning, and full collision avoidance in one continuous loop, and that the resulting plans are globally deconflicted: no new committed plan ever intersects another in space and time. The authors argue that this matters because real deployments need answers to all four questions at once, and weaker local collision-checking schemes risk deadlocks or collisions. They support the claim with simulation and real-world experiments showing average cell criticality staying below one when enough UAVs are available, graceful decay when they are not, and negligible extra planning cost for global deconfliction.","feed_headline":"Drone fleet planner guarantees collision-free persistent coverage","feed_subtitle":"A centralized loop reassigns goals and replans paths in real time, keeping average coverage criticality below one.","key_machinery":"The load-bearing mechanism is the committed plan: the portion of a UAV's trajectory it is bound to execute, kept to a bounded duration tmax. The Prioritized Planner iterates over UAVs round-robin, the Goal Assigner proposes a next cell via a pseudo-goal graph whose edge costs encode remaining lifetime minus age, and the Goal Planner searches a state lattice with weighted A*, discarding every transition that collides in space or time with any existing committed plan. Appending only a partial new path holds the global deconfliction invariant—no new committed plan intersects any other committed plan—by construction, which is what converts a collection of myopic planning decisions into a collision-free persistent coverage loop.","core_discovery":"The central discovery is that a tractable prioritized planning loop, built on the notion of committed plans, can maintain the invariant of global deconfliction in real time for persistent coverage. Each UAV commits to a plan segment of bounded duration tmax; the Goal Assigner selects the next cell to cover using a pseudo-goal Dijkstra search that trades distance against cell urgency, and the Goal Planner computes a kinodynamically feasible path on a state lattice while pruning any transition that would intersect another UAV's committed plan in space or time. Only a portion of each newly computed path is appended to the committed plan, so the invariant holds by construction across planning cycles. The paper reports that with sufficient UAVs the average criticality—age divided by lifetime, averaged over the priority map—stays below one over mission durations of up to an hour, and that when resources are insufficient coverage decays slowly rather than collapsing. It also reports that this globally deconflicted scheme costs negligibly more in planning time than schemes that ignore collisions or treat other UAVs as static obstacles, while avoiding both collisions and excessive stopping.","pith_inferences":["Because the goal assignment and path planning are decoupled and greedy, the framework is myopic; a natural extension is to couple them or add learned macro-actions, which the paper itself identifies as future work.","The global deconfliction guarantee is only as strong as the UAVs' tracking accuracy; deliberately injecting larger wind disturbances or controller noise would reveal the deviation threshold at which the invariant breaks, a testable stress boundary not covered by the reported experiments.","The same pseudo-goal priority scheme could be transferred to heterogeneous fleets with different sensor radii or speeds, since the costs already average over covered cells and the planner is centralized.","A quantitative comparison of average criticality against an optimal or offline persistent coverage plan would show how much optimality is sacrificed for online global deconfliction."],"forward_implications":["A deployment can maintain persistent coverage with average criticality below one whenever the fleet is large enough for the map, and degrade gracefully when it is not.","Global deconfliction removes the need for reactive local collision avoidance in the planner, since conflicts are excluded by construction rather than repaired after the fact.","The committed-plan length tmax is a tunable trade-off: longer plans reduce replanning frequency but add redundant coverage; shorter plans increase reactivity but risk plan thrashing and stopping maneuvers.","Dynamic addition or removal of UAVs is supported without breaking the feasibility or deconfliction properties, at the cost of coverage quality.","Planning times stay within a 4-second budget in the tested maps, with total stoppage time at most a few percent of mission time, indicating real-time operation."],"supporting_citations":[{"why":"Supplies the prioritized planning method used to plan for multiple robots sequentially instead of in a joint state space.","marker":"[8]"},{"why":"Supplies the search-based multi-robot exploration approach adapted for online goal assignment with multiple objectives.","marker":"[6]"},{"why":"Supplies the search-based kinodynamic planning method for computing long, dynamically feasible maneuvers.","marker":"[16]"},{"why":"Supplies the state-lattice technique for generating the motion primitives that define the UAV action space.","marker":"[19]"},{"why":"Supplies the weighted A* heuristic search algorithm used by the Goal Planner.","marker":"[20]"},{"why":"Supplies the frontier-based exploration idea that the Goal Assigner adapts to trade off proximity and criticality.","marker":"[31]"},{"why":"Supplies the Dijkstra search used to find the optimal path to the pseudo-goal for goal assignment.","marker":"[7]"}],"fun_headline_variants":["Persistent drone coverage, collision-free via global deconfliction","Drone coverage planner keeps persistent missions collision-free","Global deconfliction costs little in planning time","Global deconfliction for multi-UAV persistent coverage","Real-time planning sustains persistent coverage without collisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole collision-free guarantee rests on each UAV executing its committed plan closely enough that the planned trajectories remain valid; if real-world tracking error exceeds the planner's assumed tolerance, the global deconfliction invariant can fail between planning cycles.","fun_headline_variants_meta":{"raw":{"variants":["Persistent drone coverage, collision-free via global deconfliction","Drone coverage planner keeps persistent missions collision-free","Global deconfliction costs little in planning time","Global deconfliction for multi-UAV persistent coverage","Real-time planning sustains persistent coverage without collisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001059,"raw_usage":{"total_tokens":4472,"prompt_tokens":1006,"completion_tokens":3466,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":3391}},"tokens_in":622,"tokens_out":3466,"duration_ms":25704,"temperature":1.0,"reasoning_tokens":3391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:59.289531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a real-world run, record each UAV's actual trajectory alongside its committed plan; if any two UAVs ever come within the dmin = 10 m threshold despite their committed plans being deconflicted, or if the measured deviation from the planned path exceeds the planner's assumed buffer, the central claim fails. A direct simulation stress test would deliberately inflate tracking noise until the first inter-UAV distance drops below dmin.","supporting_citations":[{"cited_title":"Algorithmica2(1-4), 477 (1987)","cited_arxiv_id":null,"evidence_quote":"Supplies the prioritized planning method used to plan for multiple robots sequentially instead of in a joint state space."},{"cited_title":"In: IROS, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the search-based multi-robot exploration approach adapted for online goal assignment with multiple objectives."},{"cited_title":"IJRR 28(8), 933–945 (2009)","cited_arxiv_id":null,"evidence_quote":"Supplies the search-based kinodynamic planning method for computing long, dynamically feasible maneuvers."},{"cited_title":"In: IROS, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the state-lattice technique for generating the motion primitives that define the UAV action space."},{"cited_title":"Artificial intelligence 1(3-4), 193–204 (1970)","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted A* heuristic search algorithm used by the Goal Planner."},{"cited_title":"In: Agents, vol","cited_arxiv_id":null,"evidence_quote":"Supplies the frontier-based exploration idea that the Goal Assigner adapts to trade off proximity and criticality."},{"cited_title":"Numerische mathematik 1(1), 269–271 (1959) 14","cited_arxiv_id":null,"evidence_quote":"Supplies the Dijkstra search used to find the optimal path to the pseudo-goal for goal assignment."}],"review_version":1}