{"id":"c76fcfef-16bb-46e4-b3ba-2dd5647febdd","arxiv_id":"2411.14403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Social-GAN-style LSTM generator predicts UAS landing trajectories and beats a Gaussian Mixture Regression baseline on real drone data, though the advantage disappears on simulated data beyond four steps.","lead":"This paper applies a GAN-based neural network, borrowed from pedestrian trajectory prediction, to forecast UAS landing trajectories, and reports lower average displacement error than a Gaussian Mixture Regression baseline on a new real drone-landing dataset. The real dataset of over 2,600 manually piloted Tello drone trajectories is a potential resource for vertiport conflict detection studies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-data GAN advantage rests on a single table with overlapping uncertainty intervals and no statistical test; paired evidence is needed before the central claim is secure.","rationale":"The reader's weakest assumption concerns the modeling premise in Section III.A, that ten past positions alone contain enough information to predict the next ten positions. That is a real concern about the method's ceiling, but it is not where the central claim is most immediately vulnerable. Even if the input representation were sufficient, the paper's own quantitative evidence for the headline claim is incomplete: Table 1 gives only pointwise mean ADE and an ambiguous uncertainty interval, with no paired comparison or significance testing. The simulated-dataset results explicitly contradict the abstract's 'various datasets' phrasing, which makes the real-dataset result the only pillar for the claimed superiority. I therefore focus on the statistical and methodological validity of that pillar. My recommended verdict remains CONDITIONAL, matching the reader, because the issue is a missing-evidence problem rather than a demonstrated false result; the concern can be settled by releasing per-trajectory data and test statistics. I do not see an internal contradiction that forces outright rejection, and the real-dataset finding is plausible as reported.","tokens_in":7733,"tokens_out":4269,"duration_ms":43729,"concrete_test":"Obtain the per-trajectory ADE values for the 100 real-dataset test trajectories under both GAN and GMR, then run a paired permutation test or Wilcoxon signed-rank test at each of the ten horizons and compute bootstrap 95% confidence intervals for the mean difference. If all upper confidence bounds remain below zero and the paired p-values are below 0.05, the concern is resolved. In parallel, re-run the GMR baseline with the same input/output windowing, select the number of Gaussian components K by cross-validation, and report the exact hyperparameters; if GAN still outperforms GMR on the real dataset at every horizon under this controlled comparison, the central empirical claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim can be reduced to Table 1 (Section IV.A): GAN's mean ADE is lower than GMR's at all ten horizons on the real dataset. But the table reports only mean plus/minus a spread over 100 evaluation trajectories, without stating whether the spread is standard deviation or standard error, and without any paired significance test or confidence interval. At horizons 9 and 10 the reported intervals overlap substantially, so the observed superiority is not self-evidently significant. Because the simulated-dataset rows show GAN worse than GMR after point 4, the real-dataset column is effectively the only quantitative support for the abstract's claim that the model beats GMR 'in various datasets.' In addition, the GMR baseline configuration is not documented: no number of Gaussians K, no input/output dimensional setup, no cross-validation, and no statement that the training windows match the GAN's ten-point-input/ten-point-output protocol. Without per-trajectory paired error statistics and a reproducible baseline configuration, the headline advantage could be sampling noise or an artifact of a weakly tuned baseline. This is load-bearing because if the real-data advantage is not statistically robust, the paper's contribution reduces to providing a dataset plus a framework without demonstrated predictive benefit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an LSTM-based generative adversarial network for predicting UAS landing trajectories. The generator observes the previous 10 position samples and outputs 10 future positions, while a discriminator is trained to score trajectory quality. The method is compared with Gaussian Mixture Regression (GMR) on two simulated datasets and a real drone-landing dataset collected by the authors. The paper claims that the GAN outperforms GMR in accuracy across datasets and also claims that the trained discriminator can be used to evaluate other predictions.","tokens_in":7895,"tokens_out":5330,"duration_ms":48110,"significance":"If the real-data advantage holds under rigorous statistical testing and the dataset is released, the paper makes a useful empirical contribution: a GAN-based trajectory predictor and a real pilot-controlled UAS landing dataset. The idea of reusing the trained discriminator as a score function is also interesting, though it is not validated beyond the narrow setting of distinguishing true versus GAN-generated trajectories. The paper's main weakness is that the evidence is currently insufficient: the abstract overstates the results, the simulated-dataset rows of Table 1 contradict the 'various datasets' claim, the statistical basis for the real-dataset advantage is missing, and the GMR baseline configuration is undocumented. No code or dataset link is provided, which limits reproducibility.","major_comments":[{"comment":"The abstract states that the proposed model 'can output more accurate predictions than the baseline method (GMR) in various datasets.' This is contradicted by Table 1: for both the vertical and linear simulated datasets, the GAN has higher mean ADE than GMR for points 5 through 10 (e.g., point 10 vertical: GAN 14.88±8.41 vs GMR 9.0±3.69; point 10 linear: GAN 17.01±8.97 vs GMR 10.93±5.72). The real dataset is the only one where the GAN is consistently better. The claim must be narrowed to the real dataset, or the discrepancy must be analyzed and explained.","section":"Abstract and Section IV.A, Table 1"},{"comment":"The table reports values as mean ± spread but never states whether the spread is a standard deviation or a standard error, and no significance tests are provided. For the real dataset, the intervals overlap at later horizons (point 9: GAN 0.14±0.07 vs GMR 0.21±0.16; point 10: GAN 0.18±0.08 vs GMR 0.24±0.21). The paper should report per-trajectory paired errors, state the exact definition of the spread, and apply a paired statistical test (e.g., Wilcoxon signed-rank test) or bootstrap confidence intervals to demonstrate that the real-data advantage is not sampling noise.","section":"Section IV.A, Table 1"},{"comment":"The GMR baseline is not documented sufficiently for the comparison to be reproducible. The number of Gaussian components K in Eq. (2) is not given, the input and output dimensional setup is not described (e.g., whether GMR also uses 10 past positions to predict 10 future positions), and there is no statement about hyperparameter selection or cross-validation. Without these details, the reader cannot rule out that the GAN's advantage comes from a poorly tuned baseline. The authors should specify the GMR configuration and ideally include a sensitivity analysis over K.","section":"Section IV.A, GMR baseline"},{"comment":"The modeling premise that 'environment factors and previous status are only variables required to derive a trajectory' and the resulting mapping T_{t1,t2} = G(T_{t0,t1}) assume that ten consecutive position samples contain all information needed to forecast the next ten positions. This excludes explicit flight dynamics, control inputs, and environmental context. The assumption is load-bearing because if the input representation is insufficient, the model cannot generalize beyond its training distribution. The paper should explicitly acknowledge this as a limitation, justify it empirically (e.g., by an ablation over input length or comparison with a physics-based predictor), and state the conditions under which the assumption holds.","section":"Section III.A, Eq. (10)"},{"comment":"The contribution that the trained discriminator 'can be used to evaluate other predictions' is not supported by the presented evidence. Table 2 reports scores for true and GAN-generated trajectories from the same datasets used in training. For the real dataset, the score intervals overlap substantially (true 2.32±2.76 vs fake 0.21±0.86), and no significance test is reported. The discriminator is trained to distinguish real trajectories from those generated by this specific GAN, not to score arbitrary trajectory quality. To support the claimed use as a general evaluator, additional experiments on out-of-distribution trajectories or a clear operational definition of 'good' trajectory are needed.","section":"Section IV.B, Table 2"}],"minor_comments":[{"comment":"The number of trajectories is inconsistent: the abstract says 'more than 2600 trajectories,' while Section IV reports 2430 training trajectories and 100 evaluation trajectories, which sum to 2530. Please clarify the actual total.","section":"Abstract and Section IV"},{"comment":"The notation is unclear: the pooling function P, the embedding function phi, and the activation gamma are not defined. Eq. (14) introduces et = phi(x_{t-1}; W_em) after et was already defined in Eq. (11). Please define all symbols and remove the redundancy.","section":"Section III.C, Eqs. (13)-(15)"},{"comment":"The word 'experiences' should be 'experiments' in the sentence 'According to the results of the experiences.'","section":"Abstract"},{"comment":"The units of ADE are not stated. Please specify the units (e.g., meters) and the time interval between consecutive trajectory points.","section":"Table 1"},{"comment":"The conclusion says 'The prediction error is lower than the baseline method' without adding the qualifier that this holds for the real dataset at all points and for the simulated datasets only for the first four points. Also, the caption of Figure 6 reads 'GMM prediction' but the baseline is GMR; please use consistent terminology.","section":"Section V and Figure 6"},{"comment":"The axes and the plotted quantity (per-point mean displacement, standard deviation, or something else) are not labeled. Please clarify the content of the figure.","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The paper is quite short and the main quantitative support for the headline claim rests entirely on Table 1. The abstract and contribution list materially overstate the results relative to what Table 1 shows. I would recommend the editors require the authors to add paired statistical tests, document the GMR baseline, release the dataset and code, and revise the claims to match the evidence. The discriminator-as-evaluator contribution is currently unsupported and should be either removed or substantially qualified. If the authors cannot provide the statistical evidence, the paper's contribution reduces to a dataset description, which would be better suited to a data-description venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2411.14403. The paper ports Social-GAN to UAS landing trajectory prediction: an LSTM encoder-decoder generator with a pooling module, trained adversarially, predicting the next 10 positions from the previous 10. The genuinely new piece is the dataset: over 2,600 real landing trajectories of a DJI Tello flown by human pilots and tracked by motion capture. That is a real resource for the UAS community. The paper is also commendably honest in Section IV: it openly states that on the two simulated datasets the GAN only beats the GMR baseline for the first four predicted points. The real-data column, where GAN beats GMR at all ten horizons, is the actual result of interest.\n\nThe soft spots are real but mostly fixable. The abstract overclaims: \"more accurate predictions than the baseline method (GMR) in various datasets\" is contradicted by Table 1 on both simulated datasets after point 4. The table reports only mean ± spread over 100 evaluation trajectories, with no indication whether the spread is a standard deviation or an error, and no paired significance test. At horizons 9 and 10 on the real data the intervals overlap substantially, so the headline advantage is not self-evidently robust. The GMR baseline configuration (number of Gaussian components, training/input protocol) is undocumented, which makes the comparison hard to reproduce. The discriminator-as-evaluator claim is also under-supported: on the real dataset, true trajectories scored 2.32±2.76 while fake scored 0.21±0.86, which is not a clean separation. No code or dataset release is mentioned, and the modeling premise that past positions alone suffice ignores dynamics and control inputs; for an empirical comparison that is a defensible simplification, but it deserves a sentence of acknowledgment.\n\nNone of this sinks the central finding. The real-data comparison is a reasonable empirical contribution, and the weaknesses are omissions and framing rather than an obvious error. The paper deserves a serious referee, but with a clear mandate: fix the abstract, add paired per-trajectory statistics or confidence intervals, document the baseline, and release the dataset and code. As it stands I would not cite it for the method, but I would keep an eye out for the dataset.\n\nRecommendation: send to peer review, conditionally, expecting major revision.","headline":"A direct Social-GAN port to UAS landing with a genuinely useful new dataset and a plausible real-data result, but the abstract overclaims and the statistics are too thin.","tokens_in":8511,"tokens_out":2307,"would_cite":false,"duration_ms":21975,"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":"A GAN that reads ten past positions predicts drone landing trajectories with lower error than the GMR baseline on real pilot-flown data.","keywords":["generative adversarial network","trajectory prediction","unmanned aircraft systems","landing management","LSTM","Gaussian mixture regression","average displacement error","vertiport"],"falsifier":"A decisive test would be to retrain both models on the real dataset with a held-out pilot: if the GAN's all-ten-point ADE advantage over GMR disappears when the test trajectories come from a pilot whose landings were excluded from training, the claim that the model captures general pilot-control regularities would be refuted.","tokens_in":7450,"feed_emoji":"🛬","tokens_out":7311,"duration_ms":59804,"temperature":0.7,"pith_summary":"This paper tries to show that a generative adversarial network can predict the landing trajectory of an unmanned aircraft directly from its recent positions, without an explicit flight-dynamics model or control inputs. The proposed LSTM-based generator reads the previous ten position samples and outputs the next ten; a second LSTM discriminator scores whether a trajectory looks real. Comparing against Gaussian Mixture Regression on two simulated datasets and a newly collected dataset of more than 2,600 real pilot-flown drone landings, the authors report lower average displacement error for the GAN on the real data at all ten predicted points, and at the first four points on the simulated data. The paper also claims the trained discriminator can distinguish true from generated trajectories, making it usable as a learned evaluator. The authors note accuracy degrades as the prediction horizon lengthens, so the method is presented as a short-horizon predictor.","feed_headline":"GAN predicts real drone landings more accurately than GMR","feed_subtitle":"On 2,600+ piloted Tello trajectories, the GAN wins at all ten forecast points; on simulated data it wins the first four.","key_machinery":"The load-bearing mechanism is the adversarial training loop between an LSTM-based generator and an LSTM-based discriminator. The generator follows the Social-GAN structure: an LSTM encoder converts each embedded position into hidden states, a pooling module combines the hidden states with environment factors, and an LSTM decoder produces the ten predicted positions. The discriminator maps a trajectory to a scalar score, and the generator's loss is driven by that score. The identity carrying the argument is $T_{t_1,t_2}=G(T_{t_0,t_1})$, which replaces the ordinary differential equation $x_{t+1}=f(x_t,a_t(E))$ with a learned map from past positions alone.","core_discovery":"The central claim is that a GAN whose generator is an LSTM encoder-pooling-decoder can learn the mapping $T_{t_1,t_2}=G(T_{t_0,t_1})$ from a ten-point past trajectory to the next ten positions, and that this learned mapping is more accurate than the GMR baseline on real pilot-controlled landings. On the real DJI Tello dataset, the GAN beats GMR at all ten predicted points; on the vertical and linear simulated datasets, it beats GMR only at the first four points, and GMR is better afterwards. The authors interpret the real-data advantage as the GAN capturing subtle control regularities that the Gaussian model misses, and they present the discriminator's higher scores for true trajectories as evidence that it can serve as a learned judge of prediction quality. The same tables show a limitation stated in the conclusion: error increases with prediction length, so the method is not claimed to work for long-term prediction.","pith_inferences":["Because the model ingests only raw positions, it is likely sensitive to training distribution; a natural next test is whether the GAN retains its real-data advantage when wind, payload, or pilot identity shifts between training and test flights.","The simulated-data result (advantage only for four points) suggests the GAN-versus-GMR comparison is entangled with data regularity; a simulated dataset with pilot-like stochastic control would separate the model-class effect from the data effect.","The discriminator score could be repurposed as an online confidence monitor that flags predictions whose score drops, without retraining.","A straightforward extension is autoregressive rollouts, feeding predicted points back as input to extend the horizon, which the paper's stated short-horizon limitation implies."],"forward_implications":["Near-vertiport conflict detection and avoidance can use a GAN-based predictor that produces a ten-step landing forecast faster than a dynamics-based model.","On real pilot-flown data, the GAN's advantage over GMR persists across all ten forecast points, suggesting learned models are better suited than Gaussian regression to irregular human control styles.","On the simulated datasets, the GAN only improves on GMR for the first four points; users needing longer-horizon simulated forecasts should keep GMR as the baseline.","The trained discriminator can be reused to score new candidate trajectories, offering a learned evaluation signal without requiring a physics model."],"supporting_citations":[{"why":"Supplies the GAN min-max training objective and the generator/discriminator formulation the method is built on.","marker":"[29]"},{"why":"Provides the LSTM recurrent architecture used for the encoder, decoder, and discriminator.","marker":"[28]"},{"why":"Supplies the Social-GAN architecture of LSTM encoder, pooling module, and decoder that the generator adapts.","marker":"[24]"},{"why":"Demonstrates a GAN applied to aircraft trajectory prediction with weather conditioning, the closest prior use this paper extends.","marker":"[26]"},{"why":"Introduces the task-parameterized Gaussian Mixture Regression framework underlying the baseline predictor.","marker":"[30]"},{"why":"Provides the specific GMR implementation used as the comparison baseline.","marker":"[31]"}],"fun_headline_variants":["GAN beats GMR on 2,600+ real drone landings","Real Tello landings: GAN more accurate than GMR","Drone landing GAN wins on real data, GMR better on sim","For piloted UAS landings, GAN tops GMR","GAN improves landing trajectory prediction over GMR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a ten-point history of positions alone contains enough information to predict the next ten positions, so the flight-dynamics model, control inputs, and environment can be omitted from the predictor.","fun_headline_variants_meta":{"raw":{"variants":["GAN beats GMR on 2,600+ real drone landings","Real Tello landings: GAN more accurate than GMR","Drone landing GAN wins on real data, GMR better on sim","For piloted UAS landings, GAN tops GMR","GAN improves landing trajectory prediction over GMR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2406,"prompt_tokens":949,"completion_tokens":1457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1368}},"tokens_in":565,"tokens_out":1457,"duration_ms":12506,"temperature":1.0,"reasoning_tokens":1368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:12:31.723112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to retrain both models on the real dataset with a held-out pilot: if the GAN's all-ten-point ADE advantage over GMR disappears when the test trajectories come from a pilot whose landings were excluded from training, the claim that the model captures general pilot-control regularities would be refuted.","supporting_citations":[{"cited_title":"Generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"Supplies the GAN min-max training objective and the generator/discriminator formulation the method is built on."},{"cited_title":"Long short-term memory,","cited_arxiv_id":null,"evidence_quote":"Provides the LSTM recurrent architecture used for the encoder, decoder, and discriminator."},{"cited_title":"Social gan: Socially acceptable trajectories with generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Social-GAN architecture of LSTM encoder, pooling module, and decoder that the generator adapts."},{"cited_title":"Conditional generative adversarial networks (CGAN) for aircraft trajectory prediction considering weather effects,","cited_arxiv_id":null,"evidence_quote":"Demonstrates a GAN applied to aircraft trajectory prediction with weather conditioning, the closest prior use this paper extends."},{"cited_title":"On learning, representing, and generalizing a task in a humanoid robot,","cited_arxiv_id":null,"evidence_quote":"Introduces the task-parameterized Gaussian Mixture Regression framework underlying the baseline predictor."},{"cited_title":"gmr: Gaussian Mixture Regression,","cited_arxiv_id":null,"evidence_quote":"Provides the specific GMR implementation used as the comparison baseline."}],"review_version":1}