{"id":"21366964-b280-48df-ac57-cdbd07067dd4","arxiv_id":"2505.10770","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An integrated drone with YOLO deer detection, ant-colony coverage planning, and a charging dock is described; the detector's headline accuracy is contradicted by its own metrics and field validation is pending.","lead":"This paper describes a drone that uses a camera to spot deer in farm fields, follows energy-efficient routes to patrol, and lands on a charging pad to keep working. It was built for a 13-acre Minnesota vegetable farm and evaluated in simulation, with the charging and learning parts still under development.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 92% detection accuracy stated in the abstract is contradicted by the paper's own Table I and is unexplained; this alone undermines the central claim.","rationale":"The reader's verdict of REJECT is supported, but the single most load-bearing concern I find is not the cross-species generalization assumption identified as the reader's weakest assumption. That species-transfer issue is real, yet it depends on external validation that the paper does not supply. A more immediate and internal problem is that the abstract's headline detection accuracy of 92% is not merely unverified; it is difficult to reconcile with the paper's own Table I, where the best F1 is 0.69 and maximum recall is 0.86. If the authors meant a different accuracy metric, they do not define it, and no evidence for it appears anywhere in the results. Because detection is the first functional stage in the GUARD pipeline, a failure at this stage invalidates the integrated-system claim regardless of how well the path planner or charging dock eventually work. The coverage statistic of 15% improvement also excludes failed ACO trials, and the charging dock is explicitly under construction, so those subclaims are also weaker than the abstract implies. However, among the three components in the central claim, the detection-accuracy contradiction is the most load-bearing because it is internal to the paper's own reported data and cannot be resolved by future field tests alone. The concrete test I propose would settle the matter directly by reproducing the claimed number from the same checkpoint and test set used for Table I, or by forcing the authors to define the metric and provide the corresponding curve. Until that is done, the paper's central claim should not be accepted as stated.","tokens_in":9778,"tokens_out":4566,"duration_ms":45860,"concrete_test":"Obtain the exact model checkpoint and held-out test split used for Table I, rerun inference, and independently compute the reported 92% accuracy using the metric definition implied in Section V.A. Also compute mAP@0.5 and true-positive rate at threshold 0.34. If the 92% figure is not reproducible or is defined differently from the metrics in Table I, the detection claim in the abstract must be corrected to match the actual evaluation results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V.A states that the YOLOv5 detector achieved an average detection accuracy of 92% on the held-out test set, but Table I reports mAP@0.5=0.693, best F1=0.69 at threshold 0.34, maximum recall=0.86, and true-positive rate=0.69. No accuracy metric is defined, and the claimed 92% exceeds the reported maximum recall, so it cannot be a per-instance detection rate. Since deer detection is the input to deterrence decisions, the abstract's 'robust detection accuracy' claim is unsupported by the paper's own evaluation. This is load-bearing: if the detector performs at the reported ~69% true-positive rate rather than 92%, the integrated system's core promise fails. The cross-species issue (barasingha training data vs. the white-tailed deer target) is an additional external-validity concern, but the internal metric contradiction is sufficient to undermine the headline claim as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes GUARD, an integrated unmanned aerial vehicle (UAV) system intended to detect and deter deer on a small Minnesota vegetable farm. The system combines a YOLOv5-based vision module for real-time deer detection, an ant colony optimization (ACO) coverage path planner that minimizes energy expenditure, a proof-of-concept contact-based charging dock, and a reinforcement learning (RL) supervisor trained in an AirSim/Unreal Engine simulation. The authors report that the detector achieves 92% average detection accuracy on a held-out test set, that ACO coverage paths are more energy-efficient than a back-and-forth baseline (with the Conclusions stating a 15% improvement), and that the charging dock enables extended dusk-to-dawn operation. The paper includes results from the vision module, path planning simulations, and qualitative discussions of the charging dock and RL agent, but several hardware and control components are explicitly stated to be under construction or only simulated.","tokens_in":9960,"tokens_out":4594,"duration_ms":43788,"significance":"If the reported claims were fully substantiated, the system would represent a useful low-cost approach to autonomous wildlife deterrence in precision agriculture, combining perception, planning, and battery management for small farms. The paper has some strengths: the energy cost constants used in the planner are taken from prior literature rather than fitted to the results; the detector is evaluated on a held-out test set; and the system is motivated by a real farm partnership. However, the central claims are not supported by the evidence in the manuscript. The 92% detection accuracy is contradicted by the paper's own Table I, the coverage comparison excludes a substantial fraction of ACO trials, and the charging dock and RL supervisor are not yet implemented or field-validated. As a result, the significance of the contribution as presented cannot be assessed; the current manuscript primarily documents an in-progress prototype.","major_comments":[{"comment":"The claim of \"average detection accuracy of 92%\" is undefined and is inconsistent with the metrics reported in Table I. Table I lists mAP@0.5 = 0.693, best F1 = 0.69, maximum recall = 0.86, and true-positive rate = 0.69. Since the claimed 92% exceeds the reported maximum recall, it cannot be a per-instance detection rate, and no definition of \"accuracy\" is provided anywhere in the paper. This is a load-bearing issue because deer detection is the trigger for the entire deterrence pipeline; if the actual performance is closer to the reported true-positive rate of 0.69, the system's core promise fails. The abstract and conclusions repeat the 92% figure without qualification, so the headline claim is unsupported by the paper's own evaluation.","section":"Section V.A, Table I, Abstract, Conclusions"},{"comment":"The coverage path planning comparison is biased by the exclusion of invalid ACO trials. Section V.B states that out of 30 trials, AS generated valid solutions 19 times for the dual-drone problem and 28 times for the single-drone problem, while MMAS generated valid solutions 18 and 26 times, respectively. The statistics in Figure 11 are computed only over the valid solutions, whereas the back-and-forth baseline always produces a valid solution. This discards a substantial fraction of ACO runs that may represent real failure modes or higher-cost paths, and it inflates the apparent advantage of ACO. Additionally, the Conclusions claim that ACO covers fields \"15% more efficiently\" than the baseline, but no such figure is reported in Section V, and the figure does not provide error bars or significance tests. The 15% claim is therefore unsupported by the results as presented.","section":"Section V.B, Figure 11, Conclusions"},{"comment":"The system-level claims of \"extended dusk-to-dawn operation\" via an autonomous charging dock and of an \"integrated UAV prototype\" are not supported by the state of the components. Section IV.A states that the charging dock is a proof-of-concept: \"CAD models have been validated in Gazebo; plywood prototypes are under construction,\" and the authors \"intend to complete the circuit design\" and manufacture a prototype in the next phase. Section IV.D states that the RL policy is trained only in simulation and that \"integration and field validation of this agent are ongoing.\" Thus neither the autonomous charging nor the RL-based coordination has been demonstrated in the field, so the abstract and conclusions overstate the maturity of the system. This is load-bearing because the claimed contribution is the integrated system, not merely the individual simulated components.","section":"Sections IV.A, IV.D, Abstract, Conclusions"},{"comment":"The detector is fine-tuned on a dataset of barasingha deer (Rucervus duvaucelii) obtained from Roboflow Universe, while the target pest identified in Section I is the white-tailed deer on a Minnesota farm. The paper provides no cross-species evaluation and no field validation on the target species. Given that the authors themselves cite work (Crossling et al. [8]) showing that models trained on stock imagery generalize poorly to on-farm data, the external validity of the detection results for the actual deployment scenario is questionable. This is a significant gap because the deer species differ in appearance and habitat, and the detection results in Section V.A are the only quantitative evidence for the system's perception capability.","section":"Section IV.B, Section I"}],"minor_comments":[{"comment":"There are several typographical and grammatical errors, e.g., \"sytem\" in the Introduction and \"Some animals are may cause damage\" in Section II.A. These should be corrected.","section":"Section I, Section II.A"},{"comment":"The sentence \"an autonomous charging routine that handles return-to-dock and.\" appears to be incomplete; the final clause is missing its object.","section":"Section IV.A"},{"comment":"The waypoint spacing is described as \"arbitrarily set\" to 38 meters, and the authors acknowledge that \"more work has to be done to determine the ideal spacing.\" This free parameter affects the coverage results, and its arbitrariness should be discussed as a limitation in the evaluation.","section":"Section IV.C"},{"comment":"The caption states that one trial was conducted for the back-and-forth baseline \"since it always returns the same solution,\" but no variance or statistical significance is reported for the ACO means. It would be helpful to report standard deviations or confidence intervals over the valid trials.","section":"Figure 11, Section V.B"},{"comment":"The dataset URL in reference [24] appears to contain a typo (\"robomow.com\" instead of a Roboflow URL), and the dataset name is inconsistently spelled (\"Rucervus dasuceveli\" vs. \"Rucervus duvaucelii\").","section":"Reference [24]"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a workshop/competition report that has been submitted without substantial revision. The mismatch between the advertised 92% detection accuracy and the metrics in Table I is severe and is not a matter of presentation; it invalidates the primary quantitative claim. The coverage comparison's exclusion of failed ACO trials and the unsupported 15% improvement in the Conclusions are further indications that the reported results are not robust. The charging dock and RL controller are explicitly described as future work, yet the abstract and conclusions present them as completed components. These problems are load-bearing and cannot be fixed by local edits; the paper would need new experiments and a significant rewriting of the claims. For these reasons I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick read on GUARD. The headline claim — 'detect deer with 92% accuracy' — is the thing to check first. It doesn't hold up. Table I gives mAP@0.5=0.693, F1=0.69, recall 0.86, TPR 0.69. No accuracy metric is defined, and 92% exceeds max recall. The abstract and conclusions repeat it, so it's load-bearing. That's a genuine flaw, not a nitpick.\n\nWhat's actually here: a small-farm deer deterrent UAV, built for a specific 13-acre site. The integration is real: YOLOv5 on a Jetson Orin Nano, ACO coverage planner with an energy model, a contact charging dock (proof of concept), and an RL supervisor in AirSim. The paper is honest about what's not finished — the charging dock's electronics are under construction, the RL is simulated only, field trials are small. Credit for that. The coverage comparison is a legitimate piece of work: ACO vs. back-and-forth on a single/dual drone problem, with energy constants from prior literature. The reader notes the stats exclude invalid ACO trials, and that's a real soft spot — comparing mean of only successful trials to a baseline that always succeeds biases the comparison. They should report success rates and maybe cost of failure.\n\nThe cross-species issue is worth a sentence: training on barasingha deer and deploying on white-tailed deer with no field validation is a real external-validity gap. But the internal metric contradiction is enough to reject as written.\n\nCitation pattern looks fine; related work is relevant, including pigeon deterrence and docking stations. No red flags there. The circularity concern the reader raised doesn't land — the energy model comes from prior work, not fitted to their results.\n\nBottom line: this is a workshop-fair paper about system integration, not a principled new method. As written, the central accuracy claim is unsupported. But the components are sensible, the limitations are acknowledged, and the coverage study is a fair start. With corrected metrics and a clear statement that 92% refers to something else (or dropped), the paper could be a fine contribution. I'd send it to review — a serious referee would push for the fix, not a desk reject — but I wouldn't cite the 92% figure in anything.\n\nFor a reading group: maybe, if you want to discuss how an honest system paper can still undercut itself with an unexplained headline number.","headline":"The 92% detection claim is contradicted by the paper's own Table I and must be fixed, but the system integration and coverage comparison are worth a referee's time.","tokens_in":10487,"tokens_out":1883,"would_cite":false,"duration_ms":19143,"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":"An integrated drone can detect deer with 92% accuracy, patrol fields with 15% less energy than the standard back-and-forth route, and recharge itself for dusk-to-dawn operation.","keywords":["deer deterrence","UAV","YOLOv5","ant colony optimization","coverage path planning","autonomous charging","precision agriculture","agricultural robotics"],"falsifier":"Run the fine-tuned detector on a set of white-tailed deer images or on footage from the target farm's dusk patrol and compare its detections with ground-truth sightings from trail cameras or observers; if accuracy on the local species falls well below the 92% reported on the barasingha test set, the central detection claim does not transfer to the intended deployment.","tokens_in":9611,"feed_emoji":"🦌","tokens_out":9284,"duration_ms":85299,"temperature":0.7,"pith_summary":"This paper argues that a small, low-cost drone can take over the job of keeping deer out of a vegetable farm, replacing fences and scare devices that are too expensive or too easily ignored. The authors built an integrated prototype that combines a YOLOv5 detector for real-time deer spotting, an ant-colony route planner that minimizes energy, and a contact charging dock meant to keep the drone flying from dusk to dawn. On the paper's tests the detector reached 92% average accuracy on its held-out set, and the ant-colony routes used about 15% less energy than the standard back-and-forth coverage pattern. If those results hold in the field, the same modular stack could be pointed at other monitoring-and-response tasks in agriculture and beyond.","feed_headline":"Deer-spotting drone cuts farm patrol energy by 15 percent","feed_subtitle":"Self-charging UAV combines 92% deer detection with ant-colony routing to guard small farms around the clock.","key_machinery":"The mechanism that carries the argument is the tight coupling of three subsystems around an energy-aware waypoint loop. The detector is YOLOv5, a real-time object-detection network fine-tuned on annotated deer images; its output feeds a supervisory module that can interrupt the route. The coverage planner builds a grid of waypoints 38 meters apart, removes edges that collide with obstacles, and searches for a tour from the charging station through all valid waypoints and back using ant colony optimization, with an edge heuristic $\\eta_{hij}=1/(\\lambda d_{ij}+\\gamma\\theta_{hij})$ where $\\lambda=0.1164$ kJ/m and $\\gamma=0.0173$ kJ/deg weight straight-line distance and turning angle. The charging station uses passive 60-degree V-rails and spring-loaded pogo pins, with fiducial-marker landing to correct position before touchdown. What these pieces do together is convert a farm map, a battery gauge, and a camera feed into an autonomous patrol that can break off to haze deer and return to charge.","core_discovery":"The central claim is that the three hard problems in autonomous deer deterrence—seeing the deer, routing the patrol, and keeping the battery alive—can be solved on one inexpensive quadcopter and connected through a supervisory loop. The vision module is a fine-tuned YOLOv5 network that the paper reports as detecting deer with 92% average accuracy on its held-out test set (mAP@0.5 = 0.693, best F1 = 0.69 at a confidence threshold of 0.337). The coverage module models each flight leg by a weighted energy cost of straight-line distance and turn angle, then searches the waypoint tour with ant colony optimization; both the Ant System and Max-Min Ant System variants beat the conventional back-and-forth path, with a reported energy saving of roughly 15% and with valid tours found in the majority of trials. The charging module is a proof-of-concept dock whose V-rail and pogo-pin contacts, combined with fiducial-marker precision landing, is intended to return the drone to service without human battery swaps and thereby support dusk-to-dawn operation. The paper also describes a reinforcement-learning supervisor trained in a photorealistic simulator to coordinate detection, deterrence, and charging, with field integration still ongoing.","pith_inferences":["The 92% accuracy figure is a benchmark on the training dataset's own held-out images, so it should be read as detector capacity rather than as measured field performance on the farm.","The 15% energy saving comes from a cost model that counts only straight-line distance and turn angle; wind, climb, and obstacle-avoidance maneuvers could shrink or reverse that gap in real sorties.","Dusk-to-dawn operation is an architectural promise: the charging circuit and landing-pad prototype are still under construction, so the end-to-end recharge-and-resume loop has not yet been demonstrated.","If the simulation-trained supervisor transfers poorly to real conditions, the deployed system may regress to the static coverage planner, though the vision and routing gains would remain intact."],"forward_implications":["A single drone could patrol a 13-acre farm through the night without human battery swaps, provided the charging dock's electrical contact and alignment work as designed.","Because the ant-colony planner beat the back-and-forth baseline in both single- and dual-drone simulations, the same energy-saving routes should scale to larger fields and multiple drones.","The modular split between vision, planning, charging, and supervision means the reinforcement-learning policy can be improved or replaced without rebuilding the lower layers already validated in simulation and small-plot trials.","With only the detector retrained, the same stack could address other monitoring-and-response tasks such as wildlife surveys, poaching patrols, or wildfire spotting."],"supporting_citations":[{"why":"Supplies the motivation that low-altitude drones displace large mammals more reliably than ground-based hazing.","marker":"[5]"},{"why":"Provides the original YOLO architecture that the real-time deer detector is built on.","marker":"[6]"},{"why":"Provides the photorealistic simulator used to train the reinforcement-learning supervisor.","marker":"[9]"},{"why":"Documents the internals of YOLOv5, the specific detector variant fine-tuned for deer.","marker":"[17]"},{"why":"Supplies the precision-landing method used to align the drone on the charging dock.","marker":"[20]"},{"why":"Provides the annotated deer images used to fine-tune the detector on a domain-specific dataset.","marker":"[24]"},{"why":"Supplies the energy coefficients lambda and gamma used in the coverage cost model.","marker":"[25]"},{"why":"Defines the ant colony optimization algorithm used to search for energy-efficient tours.","marker":"[29]"},{"why":"Defines the Max-Min Ant System variant compared against the baseline.","marker":"[30]"},{"why":"Describes the back-and-forth coverage pattern used as the baseline that the 15% saving is measured against.","marker":"[33]"}],"fun_headline_variants":["GUARD drone: 92% deer detection, 15% energy cut","Self-charging deer drone saves 15% farm patrol energy","Deer detection drone tops 92% accuracy, cuts energy 15%","Robotic deer defender: self-charging, 92% accurate, 15% efficient","Drone deters deer with 92% detection, 15% less energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system's value on the target farm depends on a detector fine-tuned on barasingha deer images recognizing white-tailed deer in real Minnesota fields, and no cross-species or in-field validation is reported.","fun_headline_variants_meta":{"raw":{"variants":["GUARD drone: 92% deer detection, 15% energy cut","Self-charging deer drone saves 15% farm patrol energy","Deer detection drone tops 92% accuracy, cuts energy 15%","Robotic deer defender: self-charging, 92% accurate, 15% efficient","Drone deters deer with 92% detection, 15% less energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3296,"prompt_tokens":968,"completion_tokens":2328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2224}},"tokens_in":584,"tokens_out":2328,"duration_ms":14084,"temperature":1.0,"reasoning_tokens":2224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:03:26.587086+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fine-tuned detector on a set of white-tailed deer images or on footage from the target farm's dusk patrol and compare its detections with ground-truth sightings from trail cameras or observers; if accuracy on the local species falls well below the 92% reported on the barasingha test set, the central detection claim does not transfer to the intended deployment.","supporting_citations":[{"cited_title":"Drones outperform dogs for hazing bears: a compar- ison of carnivore aversive conditioning tools,","cited_arxiv_id":null,"evidence_quote":"Supplies the motivation that low-altitude drones displace large mammals more reliably than ground-based hazing."},{"cited_title":"Airsim: High-fidelity visual and physical simulation for autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Provides the photorealistic simulator used to train the reinforcement-learning supervisor."},{"cited_title":"Precision landing for low-maintenance remote operations with UA Vs,","cited_arxiv_id":null,"evidence_quote":"Supplies the precision-landing method used to align the drone on the charging dock."},{"cited_title":"Rucervus dasuceveli Ravindra Tripathi ZDCD0 Dataset (v1),","cited_arxiv_id":null,"evidence_quote":"Provides the annotated deer images used to fine-tune the detector on a domain-specific dataset."},{"cited_title":"UB-ANC planner: Energy efficient coverage path planning with multiple drones,","cited_arxiv_id":null,"evidence_quote":"Supplies the energy coefficients lambda and gamma used in the coverage cost model."},{"cited_title":"Ant colony optimization,","cited_arxiv_id":null,"evidence_quote":"Defines the ant colony optimization algorithm used to search for energy-efficient tours."},{"cited_title":"MAX-MIN ant system and local search for the traveling salesman problem,","cited_arxiv_id":null,"evidence_quote":"Defines the Max-Min Ant System variant compared against the baseline."},{"cited_title":"Concentrated coverage path planning algorithm of UA V formation for aerial photography,","cited_arxiv_id":null,"evidence_quote":"Describes the back-and-forth coverage pattern used as the baseline that the 15% saving is measured against."}],"review_version":1}