{"id":"21f6affe-5ea0-4f29-8f84-69217ead6be3","arxiv_id":"2412.02863","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A federated-learning LSTM pipeline lets a multi-UAV team recognize human commands and hand over control when one drone loses sight of the operator.","lead":"This paper combines LSTM action recognition with federated learning so that two drones can jointly learn and recognize hand gestures for proximal control. The authors report a real multi-drone demonstration, but the headline 96% real-robot accuracy does not match the 68.7% accuracy reported in the results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's >96% real-robot accuracy is contradicted by the paper's reported real-time accuracy of 68.67%, so the central claim is unsupported.","rationale":"The paper's strongest claim, as stated in the abstract, is that real-robot experiments achieve greater than 96% accuracy. The only number supporting that range is Table III's validation accuracy of 96-97% on dataset2, which is a model-validation metric, not a real-robot closed-loop metric. Section VI-B explicitly reports the real-time system accuracy as 0.6867. This is an internal contradiction in the manuscript's central quantitative claim, not a matter of disagreement with the field's consensus. The reader's weakest assumption correctly identified the transfer from validation accuracy to real-robot operation as the load-bearing premise, and the paper's own data show that transfer fails. The FL contribution is also not demonstrated against any non-FL baseline, and the overfitting disclaimer is logically insufficient, but the decisive issue remains the unsupported and contradicted >96% real-robot claim. Because the advertised result does not hold as stated, the existing REJECT verdict is reinforced; no verdict change is needed.","tokens_in":10329,"tokens_out":2719,"duration_ms":54327,"concrete_test":"Reconstruct the confusion matrix in Fig. 13 from the real-robot trials (or rerun the same protocol with logged ground-truth labels) and compute overall accuracy. If the value is approximately 0.6867, the abstract's '>96% real-robot accuracy' is contradicted; if a proper independent evaluation on real-robot data yields accuracy greater than 96%, the concern would be resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central quantitative claim—'experiments with real robots achieving an accuracy greater than 96%'—is contradicted by Section VI-B, which reports the real-time system accuracy as 0.6867. Table III reports 96-97% validation accuracy on dataset2, but that is not a real-robot closed-loop metric; the real-robot experiment yields 68.67% accuracy, 93% precision, and an F1 score of 0.83. The authors attribute the drop to distance, sensor quality, and lighting, but that explanation does not repair the mismatch because the abstract specifically claims real-robot accuracy. Additionally, the 'absence of overfitting' argument in Section VI-B is not a valid test: distinguishing a wrong prediction from a missed detection does not rule out overfitting to the training distribution. The federated-learning advantage is also untested against any centralized or non-FL baseline. The most load-bearing issue, however, is the self-contradictory accuracy claim: if the advertised result is real-robot accuracy greater than 96%, the paper's own reported 0.6867 accuracy falsifies it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a federated-learning (FedAvg) pipeline for LSTM-based action recognition from Mediapipe keypoints, used for proximal gesture control of a UAV, with a second UAV maintaining the human operator in view to avoid occlusions. The authors create a second gesture dataset, validate their model on two datasets, and report a real-robot demonstration with three clients and two drones. The abstract claims that experiments with real robots achieved accuracy greater than 96%, and the paper claims to be the first work using federated learning in HRI for proximal control of UAVs.","tokens_in":10563,"tokens_out":6526,"duration_ms":68773,"significance":"If the claims were supported, the paper would offer a practical demonstration of onboard federated LSTM action recognition for multi-UAV proximal control, with a public dataset and a real-robot occlusion-handover experiment. These are genuine strengths: the authors deployed two different UAV platforms, used standard FL tooling (Flower), report a confusion matrix, and provide a video link. However, because the headline real-robot accuracy is directly contradicted by the reported 68.67% real-time accuracy and because the FL component is not compared with any non-FL or centralized baseline, the significance of the contribution as stated is not established.","major_comments":[{"comment":"The abstract's claim that 'experiments with real robots achieving an accuracy greater than 96%' is directly contradicted by Section VI-B, which reports that 'in the real-time experiments, the Accuracy of our system was 0.6867.' Table III's 0.96/0.97 figures are model validation accuracies on dataset2, not closed-loop real-robot results. Since the paper's own real-robot experiment is the only direct evidence for the headline claim, the abstract-level result is unsupported and must be corrected.","section":"Abstract and Section VI-B"},{"comment":"The federated-learning contribution is not evaluated against any relevant baseline. Section VI-A states that each client used the same dataset, rather than the non-IID partitions that motivate FL in Section V, and the paper reports no comparison with centralized training or with local-only training. Consequently, the statements that FL enables distributed training and accelerates convergence are not supported by the experiments.","section":"Sections V and VI-A"},{"comment":"The 'absence of overfitting' argument is not a valid test: distinguishing a wrong action prediction from a missed detection addresses error modes, not the generalization gap between training and deployment. The paper does not report a held-out real-robot protocol or per-trial variability, so the drop from 96-97% validation accuracy to 68.67% real-time accuracy remains unexplained and is attributed only to environmental factors.","section":"Section VI-B"},{"comment":"The created dataset is described as having 33 unique actions and 6 unique static gestures after multiplying 11 actions and 2 gestures by three viewpoints, yet the architecture in Table I has a 13-unit softmax output and Section III-B describes commands for 13 classes. If viewpoint-specific labels are used, the output layer cannot represent them; if they are mapped back to 13 classes, the '33 unique actions' statement is misleading. The label mapping needs clarification.","section":"Section IV-2 and Table I"}],"minor_comments":[{"comment":"The captions alternate between 'Hover' and 'Have command' for what appears to be the same action; please standardize the action names.","section":"Fig. 12"},{"comment":"Reference [19] is cited for OpenPose but the arXiv identifier 2007.14390 actually belongs to the Flower framework, which is also cited as [24]; the OpenPose citation should be corrected.","section":"References"},{"comment":"The phrase 'In resume' should be 'In summary'.","section":"Section III-B"},{"comment":"The symbol R is used both for the rotation matrix and for the set of unit vectors in equation (3); please use distinct notation to avoid ambiguity.","section":"Equation (1)"}],"recommendation":"reject","confidential_remarks":"For the editor: the manuscript contains a useful hardware demonstration and a public dataset, but the headline accuracy figure is contradicted by the authors' own measurement in Section VI-B, and the federated-learning benefit is untested against any baseline. The novelty claim about being the first FL-based HRI proximal-control work would require a more systematic related-work search and experimental support before it can be evaluated. I therefore recommend rejection unless the authors substantially reframe the claims and add the missing comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one if you care about how federated learning gets bolted onto HRI. The punchline: the advertised headline number is wrong. The abstract says real-robot accuracy greater than 96%; Section VI-B reports real-time accuracy 0.6867. That is a direct contradiction. Table III's 96–97% is model validation on their dataset2, not closed-loop robot performance.\n\nWhat is actually here: a plausible integration of known components—Mediapipe pose, a two-layer LSTM, federated averaging via Flower, and a two-drone setup where the X500 and DJI cover each other's occlusion. The real-robot experiment shows the handover working: when the X500 loses the operator, the DJI picks up the command. They also release a dataset and code, which is reproducible evidence and worth something. The writing is clear, and the pipeline description is honest about the wait times and filtering.\n\nThe soft spots are real. The accuracy contradiction is load-bearing. The authors explain the drop by distance, sensor quality, and lighting, but that does not repair the mismatch with the abstract. There is no non-FL baseline, so we do not learn whether federated training actually helps or hurts. No error bars, no repeated trials beyond one sequence. And the 'absence of overfitting' argument—distinguishing a wrong prediction from no detection—does not test overfitting at all.\n\nAs is, rejection is the right call because the central claim fails. But the system works, the dataset is useful, and the occlusion handover is a nice demonstration. With a corrected abstract, honest reporting of validation versus real-robot accuracy, and a baseline comparison, this could become a decent short paper. The novelty is integration, not components, and the authors say as much.\n\nFor peer review: I would send it out. A referee can catch the contradiction and force the fix, and the dataset and repo make the work checkable. Expect major revision.","headline":"The abstract's >96% real-robot accuracy claim is contradicted by the paper's own 68.67% real-time result; the real contribution is a working FL-LSTM multi-UAV handover demo and dataset, but the paper overstates it.","tokens_in":11093,"tokens_out":3438,"would_cite":false,"duration_ms":35559,"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 aims to show that a federated-learning-trained LSTM classifier can give a UAV proximal gesture control, and that a second UAV prevents occlusion by taking over detection when the first loses sight of the operator.","keywords":["federated learning","human-robot interaction","UAV control","action recognition","LSTM","gesture recognition","occlusion","multi-robot systems"],"falsifier":"A controlled reproduction that runs the same trained model on the same drone platforms while varying operator distance, lighting, and camera quality in the ranges described in Section VI-B, and measures end-to-end command accuracy, would settle whether the offline accuracy transfers; if accuracy stays near 0.69 rather than above 0.96, the paper's headline accuracy claim is not supported.","tokens_in":10130,"feed_emoji":"🚁","tokens_out":4989,"duration_ms":48943,"temperature":0.7,"pith_summary":"The authors aim to establish that a lightweight LSTM action classifier, trained distributively through federated learning on multiple drones, can translate human gestures into drone commands, and that a multi-drone setup solves the occlusion problem by letting a secondary drone take over whenever the main drone's view is blocked. They claim this is the first use of federated learning in human-robot interaction for proximal UAV control. In support, they present offline validation accuracy of 96–97% on their own three-view dataset, and real-robot trials in which the view handover successfully executed a command sequence. The same trials, however, report a real-time accuracy of 0.6867 with precision 0.93 and F1 0.83, which the authors attribute to distance, sensor quality, and lighting conditions.","feed_headline":"Drones train together to read gestures and dodge blind spots","feed_subtitle":"A federated LSTM on two UAVs keeps a second camera on the operator when the first loses sight; real trials executed the command chain.","key_machinery":"The load-bearing mechanism is a two-layer LSTM followed by three dense layers (with dropout) that operates on a sliding window of 60 frames of 33 MediaPipe body keypoints (x, y, z coordinates), and is trained across clients by federated averaging. The occlusion countermeasure is a priority check: the secondary drone's commands are ignored whenever the primary drone detects the operator, so the second viewpoint acts as a fallback transmitter. This combination—the streaming keypoint window and the priority-checked viewpoint handover—carries the argument.","core_discovery":"The central claim is that an LSTM with two recurrent layers and three dense layers, trained with federated averaging across drone-mounted clients, classifies 13 gesture/action classes from MediaPipe body keypoints well enough to command a UAV, and that a second drone can take over command when the primary drone loses sight of the operator. The paper reports training and validation accuracy of 96–97% on its own dataset (Table III) and a real-time system accuracy of 0.6867, with precision 0.93 and F1 0.83 in the real-robot trials. The authors state that this is the first work applying federated learning to proximal control of UAVs in human-robot interaction.","pith_inferences":["If federated training converges in only three rounds with two clients, the value of FL here may be less about privacy or communication savings than about pooling training signal from multiple viewpoints; a direct comparison against a centrally trained model on the same data would separate these effects.","The gap between 96–97% offline validation accuracy and 0.6867 real-time accuracy suggests that robustness to lighting, distance, and sensor noise is the actual bottleneck; a controlled test that varies these conditions systematically could guide targeted data augmentation.","The 60-frame sliding window and the 10-second wait between commands restrict the system to discrete, slow command sequences; extending to continuous velocity commands would require a different action set and temporal decoding, a natural next step the paper does not address.","The novelty claim of first use of federated learning in proximal UAV control rests on the completeness of the cited literature; none of the gesture-control works referenced use FL, which supports the claim but does not prove it."],"forward_implications":["Gesture-based proximal control of UAVs can be trained without centralizing operator video data, since only model weights are shared across drones.","Multi-robot teams can maintain continuous command of a drone by viewpoint handover, because the secondary robot's detection covers the primary robot's blind spots.","The same federated architecture extends to more than two drones or cameras, as the authors note, so larger swarms could share a single action model.","A streaming action classifier with a 10.4-second command window can be deployed onboard with modest compute, as demonstrated on the X500 drone.","The work offers a comparison between a literature dataset and a new three-view dataset, providing a baseline for future gesture-control UAV research."],"supporting_citations":[{"why":"Supplies the literature UAV-Gesture dataset of 11 actions and 2 static gestures with body joints, which the authors test against and extend.","marker":"[17]"},{"why":"Provides the federated averaging algorithm that the whole distributed training scheme is built on.","marker":"[7]"},{"why":"Is the MRS UAV system used for flight control and state estimation of the primary drone in real-robot trials.","marker":"[23]"},{"why":"Flower is the FL framework whose federated average strategy implements Algorithm 1.","marker":"[24]"},{"why":"Is the prior multi-UAV full-body kNN action classifier that this work extends by adding federated learning.","marker":"[5]"},{"why":"OpenPose computes the body joints for the literature dataset, grounding the keypoint representation.","marker":"[19]"},{"why":"Is the prior application of federated learning to human activity recognition, the intellectual antecedent for FL in this domain.","marker":"[12]"}],"fun_headline_variants":["Federated drones learn gestures to avoid blind spots","Multiple UAVs share learning to keep eyes on operator","LSTM with federated training boosts UAV gesture control","Two drones collaborate to command from any angle","Distributed learning lets UAVs read gestures despite occlusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the high offline validation accuracy of 96–97% transfers to real-robot closed-loop operation, despite the paper reporting a real-time accuracy of 0.6867 in Section VI-B.","fun_headline_variants_meta":{"raw":{"variants":["Federated drones learn gestures to avoid blind spots","Multiple UAVs share learning to keep eyes on operator","LSTM with federated training boosts UAV gesture control","Two drones collaborate to command from any angle","Distributed learning lets UAVs read gestures despite occlusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1202,"prompt_tokens":900,"completion_tokens":302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":228}},"tokens_in":516,"tokens_out":302,"duration_ms":3592,"temperature":1.0,"reasoning_tokens":228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:00:19.376072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled reproduction that runs the same trained model on the same drone platforms while varying operator distance, lighting, and camera quality in the ranges described in Section VI-B, and measures end-to-end command accuracy, would settle whether the offline accuracy transfers; if accuracy stays near 0.69 rather than above 0.96, the paper's headline accuracy claim is not supported.","supporting_citations":[{"cited_title":"Uav-gesture: A dataset for uav control and gesture recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the literature UAV-Gesture dataset of 11 actions and 2 static gestures with body joints, which the authors test against and extend."},{"cited_title":"The MRS UA V System: Pushing the Frontiers of Repro- ducible Research, Real-world Deployment, and Education with Au- tonomous Unmanned Aerial Vehicles,","cited_arxiv_id":null,"evidence_quote":"Is the MRS UAV system used for flight control and state estimation of the primary drone in real-robot trials."},{"cited_title":"Controlling a swarm of unmanned aerial vehicles using full-body k-nearest neighbor based action classifier","cited_arxiv_id":null,"evidence_quote":"Is the prior multi-UAV full-body kNN action classifier that this work extends by adding federated learning."},{"cited_title":"Human activity recognition using federated learning,","cited_arxiv_id":null,"evidence_quote":"Is the prior application of federated learning to human activity recognition, the intellectual antecedent for FL in this domain."}],"review_version":1}