{"id":"a7d5c929-d26a-47df-8f97-fac4412062ce","arxiv_id":"2412.11717","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A simulation-trained reinforcement learning policy finds clustered weeds with much shorter UAV paths than row-by-row coverage, at the cost of missing some weeds.","lead":"The authors trained a drone-control program in simulation to search for weeds that grow in patches, instead of flying over every row of a field. On real field images, it flew 66% less distance than a standard row-by-row sweep while finding about 10% fewer weeds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 66% real-world path saving omits the high-altitude flight used to create the prior-knowledge map; including it reduces the saving to roughly 20-30%.","rationale":"The central claim is that a simulation-trained DQN policy can locate non-uniformly distributed weeds with minimal flight-path length, quantified in the abstract as a 66% shorter real-world path at 10% lower recall. For that number to support the claim, the path length attributed to the method must be fairly comparable with the row-by-row baseline. Section 2.4.5 creates the prior map by flying a high-altitude row-by-row survey with a 24x24 FoV over a 48x48 field; this requires roughly 96 m of flight before the DQN starts searching. The Results section compares only the DQN's 94 +/- 31 search steps to the 276-step baseline, omitting that survey cost. This is a concrete, checkable accounting error that materially weakens the headline real-world result. The reader's sim-to-real fidelity concern is also legitimate, but it is more diffuse; the omitted prior-collection cost is the most immediate load-bearing issue in the reported 66% figure. I do not see this as invalidating the method: with free prior knowledge from previous-season maps or satellite data, the search-only claim can stand, and the simulation experiments support the qualitative conclusion. The paper also releases code and discusses limitations honestly. However, the paper should either report end-to-end path length including prior acquisition or explicitly scope the 66% claim to search-path-only comparison. Since the reader already issued a CONDITIONAL verdict and this concern adds a condition rather than overturning the work, I recommend UNCHANGED.","tokens_in":20048,"tokens_out":10556,"duration_ms":106285,"concrete_test":"Recompute Experiment 5 with an end-to-end path-length accounting: use the same Fields2Cover planner to compute the high-altitude row-by-row survey path (path width 24, field size 48) that generates the prior-knowledge map, add it to the DQN's 94 +/- 31 search steps, and compare the total with the 276-step low-altitude row-by-row baseline. If the total saving remains near 66%, the headline stands; if it drops to 20-30%, the paper must rescope the claim to search-only path length or report the full mission comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Experiment 5's headline comparison is not end-to-end. In Section 2.4.5, the global prior-knowledge map is generated by 'a high-altitude row-by-row flight path with a field-of-view of 24x24 grid cells' over a 48x48 field. That survey requires about 2x48 = 96 m of straight-line coverage plus turn connectors, comparable to the DQN's own 94 +/- 31 search steps. The reported 66% saving credits the DQN with only the low-altitude search path while charging the row-by-row baseline with the full 276-step low-altitude coverage. If the prior-acquisition flight is included, the DQN's total path is roughly 190-214 steps, a 22-31% saving rather than 66%. The paper's Section 4.6 lists external sources of prior knowledge, so the search-only claim can survive if the prior is free, but the implemented Experiment 5 does not count its cost. The central quantitative claim is therefore overstated as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-Q-learning-based path planner for UAV weed localization. The agent receives a simulated or real detection map of the current field of view and an uncertain low-resolution prior map of the whole field, and outputs discrete flight actions (north, south, east, west, and optionally 'land'). The policy is trained entirely in simulation with Gaussian-mixture weed distributions and simulated perception and prior-knowledge errors. Five experiments evaluate the effect of weed distribution, detection-error level, prior-knowledge quality, stopping criteria, and transfer to four real-world orthomosaics. The central claim is that the learned search policy finds non-uniformly distributed weeds with a much shorter flight path than a row-by-row baseline; on real-world data the paper reports a 66% shorter path at a cost of a 10% lower percentage of found weeds.","tokens_in":20290,"tokens_out":4434,"duration_ms":40456,"significance":"If the claims hold, the paper makes a useful contribution to agricultural UAV search: it demonstrates a sim-to-real transfer of an RL search policy, provides an open-source simulation and training code, evaluates over 1000 episodes per condition with Welch t-tests, and honestly discusses practical limitations such as the inability to guarantee finding all weeds. The robustness analysis across detection-error levels and prior-knowledge qualities is a strength, as is the use of a learned stopping action. However, the headline real-world saving is overstated because the cost of acquiring the prior-knowledge map is omitted, and the simulation error model is not calibrated to the real detector; both issues affect the quantitative conclusions.","major_comments":[{"comment":"The reported 66% shorter real-world path does not account for the high-altitude row-by-row flight used to generate the prior-knowledge map. Section 2.4.5 states that the prior map is created by 'a high-altitude row-by-row flight path with a field-of-view of 24x24 grid cells' over a 48x48 field; that survey requires roughly 96 m of straight-line coverage plus turn connectors. The DQN search path is 94 ± 31 steps, so the end-to-end DQN cost is about 190–214 steps, while the row-by-row baseline is 276 steps. The resulting saving is approximately 22–31%, not 66%. Please report the end-to-end comparison including prior acquisition, or explicitly state that the prior is assumed to come for free from an external source. As written, the abstract and conclusion overstate the real-world benefit.","section":"§2.4.5, Experiment 5; §3.5; §5"},{"comment":"The default simulation parameters are internally inconsistent with the stated 'moderate' detection-error level. Table 1 lists rdt,fp = 0.05 and rdt,fn = 0.0001, whereas Table 3 defines the moderate level as rdt,fp = 0.0001 and rdt,fn = 0.05. The values are effectively swapped, and the false-positive rate differs by a factor of 500. Because experiments 1, 3, 4, and 5 use the default parameters, this inconsistency affects the interpretation of all results that are not part of Experiment 2. Please correct the table or clarify which parameter set is actually used.","section":"§2.1, Table 1, Table 3"},{"comment":"The simulation error model is not calibrated to the real YOLOv8-based detection system. The simulator assumes independent false positives, false negatives, and Gaussian positional offsets per detection, but Experiment 5 does not report the actual false-positive or false-negative rates of YOLOv8-nano on the four orthomosaics, nor does it examine whether real errors are spatially correlated (e.g., missed clusters or systematic localization bias). The transferability claim therefore rests on an unverified assumption that the simulated error statistics approximate the real detector. Please report the detector's confusion statistics on the real datasets and, if feasible, test the policy under spatially correlated error models.","section":"§2.1, §2.4.5, §4.6"}],"minor_comments":[{"comment":"The parameter name 'rpn,fn' appears to be a typo for 'rpk,fn'; also the values of rdt,fp and rdt,fn in Table 1 should be cross-checked against Table 3.","section":"Table 1"},{"comment":"The row-by-row baseline standard deviations at 100, 200, and 300 steps are inconsistent across tables (e.g., 0.37 ± 0.08 in Table 5 versus 0.37 ± 0.30 in Table 6), although the baseline is the same; please unify these values.","section":"Tables 5–7"},{"comment":"The caption for panel (a) says 'very low' but the detection-error level is 'very high'; please correct the caption.","section":"Figure 13"},{"comment":"The term 'UA V' contains an unintended space; please standardize to 'UAV'.","section":"Throughout"},{"comment":"For the 'no new weeds for N steps' stopping criterion, the sentence 'at least 2 weeds need to be detected' is ambiguous about whether the threshold counts the total number of detected weeds or the number of weeds in the most recent detection event; please clarify.","section":"§2.4.4"}],"recommendation":"major_revision","confidential_remarks":"The work is within the journal's scope and the central idea is sound, but the quantitative headline in Experiment 5 needs correction and the simulation-detector calibration needs to be addressed. The Table 1/Table 3 inconsistency is likely a typo, but it must be fixed because it affects the default experimental setup. I see no indication of authorship or novelty problems; the self-citation to van Essen et al. (2025) for the real-world datasets is transparent and appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper earns its keep as a careful RL search-policy study for patchy weed localization, but the marquee real-world number is overstated. The 66% shorter path in Experiment 5 counts only the low-altitude search flight; the prior-knowledge map was produced by a high-altitude row-by-row survey over the same field (~96 m of straight-line coverage plus turns, comparable to the DQN's own 94±31 steps). Include that and the saving drops to roughly 20-30%. Section 4.6 does list external sources for priors, so the search-only claim can survive if the prior is free, but Experiment 5 as implemented does not count its cost. That is a load-bearing caveat for the abstract.\n\nWhat is actually new: the abstract representation with uncertain prior knowledge, a learned landing action, and a zero-shot transfer test on real orthomosaics. The architecture is adapted from Theile et al., and the paper says so. The simulations are carefully done—1000 evaluation episodes, significance tests, multiple distributions, error levels, and stopping criteria. The code and simulator are on GitHub, and the self-citation for the real-world datasets is transparent. The real-data experiment is honest about its limitations (offline orthomosaics, four fields, no live drone).\n\nWhere it is soft, in proportion: the simulation's detection-error model is not calibrated against the actual YOLOv8 errors; the real-world evidence is pre-recorded rather than a flight test; and the 66% claim, as noted, is not end-to-end. None of these are fatal to the central simulation results—the policy does learn to exploit non-uniform distributions and is robust to moderate noise. The missing prior-survey cost is the one that changes the abstract's message, and it is fixable by reporting the total path (prior survey + search) or by using a free prior source.\n\nWho gets value: readers working on informative path planning for precision agriculture, or on sim-to-real transfer of RL search policies. It deserves a serious referee; a good reviewer will catch the prior-cost issue and ask for the error-model calibration. I would recommend engaging with it.","headline":"Solid simulation study with a guarded real-data transfer; the headline 66% saving omits the cost of the prior-knowledge survey, so read the real-world claim as ~20-30% end-to-end.","tokens_in":20862,"tokens_out":3657,"would_cite":false,"duration_ms":32559,"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 claims that a simulation-trained DQN search policy can find non-uniformly distributed weeds with far shorter UAV flight paths than row-by-row coverage.","keywords":["deep reinforcement learning","path planning","UAV weed detection","object search","prior knowledge","sim-to-real transfer","deep Q-learning","agricultural robotics"],"falsifier":"Run the transfer experiment on a field whose weed clusters are strongly elongated or row-aligned, outside the Gaussian-mixture ranges used in training, with a detector whose false positives are spatially correlated, for example concentrated in sun-glinted strips, and compare the learned policy's path length and fraction found against row-by-row; if the saving falls far below 66% or the agent lands before finding most weeds, the simulator-fidelity assumption is falsified.","tokens_in":19807,"feed_emoji":"🚁","tokens_out":6356,"duration_ms":57100,"temperature":0.7,"pith_summary":"This paper claims that a drone can find weeds scattered in patchy clusters far faster by learning to search than by flying row-by-row over the whole field. The authors train a deep-Q-learning agent in simulation, feeding it a low-resolution prior map of likely weed locations plus a high-resolution in-flight detection map, and let it choose the next flight direction and when to land. In simulation the learned policy found most weeds in roughly a third of the steps the row-by-row path needed, and it stayed efficient under noisy detections and imprecise prior knowledge. Transferred to four real-world field datasets without retraining, the policy flew a 66% shorter path than row-by-row while finding 10% fewer weeds. The practical point is that for patchy objects like weeds, a search policy can trade a small miss rate for a large saving in flight time.","feed_headline":"Learned drone search cuts weed-hunt flight path by 66 percent","feed_subtitle":"A deep-Q agent trained on simulated weed clusters beats row-by-row sweeping on real field images, missing 10% of weeds.","key_machinery":"The central object is the dual-map deep-Q-network state representation: a drone-centered global map (down-sampled prior knowledge of likely weed locations, already-detected weeds, and field boundaries) concatenated with a local map of the current field-of-view detection output and a scalar battery budget. The network uses two parallel convolutional feature extractors feeding fully connected layers that output Q-values for flying north, south, east, west, and, in experiments 4 and 5, a land action. This representation lets the same trained policy react to in-flight detections while steering toward clusters suggested by prior knowledge.","core_discovery":"The central discovery is that a simulation-trained deep-Q-network search policy, using a drone-centric dual-map state (coarse global prior, detailed local detections, and remaining battery), learns a search strategy that exploits the spatial clustering of weeds. When weeds are drawn from Gaussian mixtures, the agent finds more than 80% of them in 73 to 94 steps on average, versus 209 steps for row-by-row; on real-world orthomosaic data it lands after 94 ± 31 steps having found 81.2% ± 10.4% of weeds, versus 276 steps and 90.9% ± 5.8% for row-by-row, a 66% shorter path at a 10% lower find rate. The paper also claims the policy is robust to detection false positives, false negatives, and positional noise, needs only coarse prior knowledge, and learns a sound landing action that terminates the search, with only 2% premature landings.","pith_inferences":["The 66% saving is a point estimate on four real datasets; a practitioner should first check whether their field's weed cluster count, cluster spread, and detector error correlation fall inside the randomized training ranges, because the paper's own discussion says retraining may be needed otherwise.","The same dual-map search formulation should transfer to other patchy-object search tasks, such as disease hotspots, invasive plants, or scattered animal groups, where missing a small fraction of targets is acceptable; the learned land action is what makes the trade-off explicit.","A testable extension the authors did not run is giving the drone altitude actions so it can take a high-altitude overview as fresh prior knowledge, which they identify as future work and would relax the requirement for external prior maps."],"forward_implications":["On non-uniform weed distributions, the learned policy beats row-by-row coverage in path length for finding most weeds, but it does not guarantee complete coverage.","Detection errors up to moderate levels barely change performance; only a very high error level degrades results.","Coarse prior knowledge of moderate quality is enough; without any prior knowledge a row-by-row sweep is more efficient.","A learned land action gives a large path saving (74% shorter in simulation, 66% on real data) while sacrificing a modest fraction of found weeds.","The policy transfers to real-world imagery without retraining, suggesting the same simulation-trained policy can be deployed with a real onboard detector and flight controller."],"supporting_citations":[{"why":"Supplies the dual-map global/local network architecture that the state representation adapts.","marker":"Theile et al. (2021)"},{"why":"Supplies the deep Q-network algorithm, including the replay buffer and target network.","marker":"Mnih et al. (2015)"},{"why":"Supplies the four real-world orthomosaic datasets and the detection training data used in the transfer experiment.","marker":"van Essen et al. (2025)"},{"why":"Supplies the object detector used to build real local weed maps in Experiment 5.","marker":"Jocher et al. (2023)"},{"why":"Supplies the row-by-row coverage path planner used as the baseline.","marker":"Mier et al. (2023)"},{"why":"Provides prior evidence that a learned stopping action can terminate object search, which the land action builds on.","marker":"Yang et al. (2018)"},{"why":"Provides a comparison point for learned stopping actions and their failure modes.","marker":"Druon et al. (2020)"}],"fun_headline_variants":["AI drone cuts weed-hunt path by 66%","Drone RL learns weed spots, flies 66% less","Sim-trained drone policy beats row sweeping on weeds","Drone search learns weed clusters, trims flight by 66%","Reinforcement learning drone trims weed search path 66%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on the simulated weed fields and simulated detection errors matching real fields closely enough that a policy trained on them still makes good decisions in practice.","fun_headline_variants_meta":{"raw":{"variants":["AI drone cuts weed-hunt path by 66%","Drone RL learns weed spots, flies 66% less","Sim-trained drone policy beats row sweeping on weeds","Drone search learns weed clusters, trims flight by 66%","Reinforcement learning drone trims weed search path 66%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2307,"prompt_tokens":1034,"completion_tokens":1273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":1188}},"tokens_in":650,"tokens_out":1273,"duration_ms":10598,"temperature":1.0,"reasoning_tokens":1188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:38:42.416394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the transfer experiment on a field whose weed clusters are strongly elongated or row-aligned, outside the Gaussian-mixture ranges used in training, with a detector whose false positives are spatially correlated, for example concentrated in sun-glinted strips, and compare the learned policy's path length and fraction found against row-by-row; if the saving falls far below 66% or the agent lands before finding most weeds, the simulator-fidelity assumption is falsified.","supporting_citations":[{"cited_title":", author Bayerlein, H","cited_arxiv_id":null,"evidence_quote":"Supplies the dual-map global/local network architecture that the state representation adapts."},{"cited_title":", author van Henten , E","cited_arxiv_id":null,"evidence_quote":"Supplies the four real-world orthomosaic datasets and the detection training data used in the transfer experiment."},{"cited_title":", author Qiu, J","cited_arxiv_id":null,"evidence_quote":"Supplies the object detector used to build real local weed maps in Experiment 5."}],"review_version":1}