{"id":"c5c57f5e-8cd1-4ae5-944b-c635e0b43110","arxiv_id":"2412.14442","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"EPN predicts a target vehicle's future trajectory using the ego vehicle's planned path and refined CVAE endpoint guesses, reporting 30 to 65 percent lower errors than baselines, though the planned-path setup is undocumented.","lead":"An autonomous-driving trajectory prediction network, EPN, adds the ego vehicle's future planned path and predicted destination points to standard historical inputs. The authors report large error reductions on two highway datasets, but the source of the planned-path input is not described, which casts doubt on the comparisons.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3)'s ego planned trajectory P is never sourced for NGSIM/HighD; if P uses ground-truth future ego positions, EPN leaks future information and the reported gains are unfair.","rationale":"The reader's verdict is REJECT, and my stress-test pass identifies the same load-bearing weakness. The paper's significance rests entirely on the empirical performance gap, and that gap depends on an input that is not defined for the datasets used. NGSIM and HighD are naturalistic trajectory recordings with no planner outputs; any 'planned trajectory' must be synthesized by the authors. If the synthesis uses the ego vehicle's ground-truth future, then EPN is not a fair predictor at inference time, because it receives information about the future that the baselines do not. This is not a minor implementation detail: Eq. (13) and Eq. (16) explicitly fuse P into the social feature vector, and the mutual-influence argument in the introduction depends on P being a genuine plan that exists before the future unfolds. The paper gives no procedure, no pseudo-code, and no ablation with P removed or replaced by a causal planner in the main comparison, only the promised open-sourcing after acceptance. I also considered other issues: the oracle/min-of-k selection in Table III is a known limitation of multimodal prediction evaluation and applies to the baselines as well, so it is not the decisive problem; the absence of error bars and of comparisons to endpoint-based methods like DenseTNT weakens the paper but would not by itself justify rejection if the P-generation issue were resolved. The writing is clear and the architecture is plausible, and the ablation study in Table IV is internally consistent, which suggests the authors did run experiments. But internal consistency does not establish that the headline comparison is fair. A single concrete check—inspect the P-generation code and rerun with a causal planner—would settle whether the central claim survives. Since that check has not been done or documented, the reader's REJECT verdict is appropriate, and the paper should be revised to specify and justify the P-generation procedure or the claims should be withdrawn.","tokens_in":12037,"tokens_out":4172,"duration_ms":37102,"concrete_test":"In the released preprocessing code, or by directly requesting the authors' data-preparation script, trace how P in Eq. (3) is populated for NGSIM and HighD. Concretely, compare the values assigned to p_{t+1},...,p_{t+Tf} against the recorded ground-truth trajectory of the ego vehicle at the same timestamps. If P coincides with, or is derived from, the ego's ground-truth future positions, rerun the full NGSIM and HighD comparisons with P replaced by a history-only planner (e.g., constant-velocity extrapolation or a simple kinematic planner using only states up to time t), keeping all baselines unchanged. If the RMSE/ADE/FDE gaps over S-TF and PiP shrink substantially or vanish, the reported improvement is leakage-driven; if the gaps persist, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: EPN reduces RMSE/ADE/FDE by roughly 30–65% against planning-informed and intention-based baselines on NGSIM and HighD. The most load-bearing assumption is that the input P in Eq. (3), the ego vehicle's future planned trajectory, is available at inference without using future information about the target or the scene. Section V-A describes only downsampling, segment creation, and a 7:1:2 split for NGSIM and HighD; it never states how P is generated. Neither dataset contains planner outputs, so the in-dataset source for P can only be the recorded future trajectory of whichever vehicle is treated as the ego vehicle. If that is what was used, then EPN is given the ego's future positions while S-LSTM, CS-LSTM, S-GAN, WSiP, PiP, and S-TF are not. Eq. (13) encodes P via an LSTM, and Eq. (16) passes it through convolutional pooling into encsocial, so the model can exploit the strong correlation between the ego's future motion and the target's future motion, especially in highway car-following scenarios. This would make the reported reductions—particularly the 64.6% RMSE reduction on HighD—an artifact of information leakage rather than evidence that the architecture or endpoint correction is better. The paper also promises code only after acceptance, so the construction of P cannot currently be checked. The endpoint correction idea may still have merit, but the headline comparison is uninterpretable without a precise, non-future-dependent specification of P.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EPN, a multimodal trajectory prediction network that augments a social-pooling LSTM architecture with the ego vehicle's future planned trajectory P (Eq. (3)) and with a CVAE-based endpoint prediction module that corrects sampled endpoints before decoding complete trajectories. Experiments on NGSIM and HighD report large reductions in RMSE, ADE, and FDE relative to S-LSTM, CS-LSTM, S-GAN, WSiP, PiP, and S-TF, and ablations on NGSIM attribute part of the gain to speed/acceleration inputs, to the endpoint prediction module, and to the endpoint correction mechanism. Code is promised only after acceptance.","tokens_in":12368,"tokens_out":6059,"duration_ms":54812,"significance":"If the input P is produced causally, the core idea of conditioning target prediction on an ego plan is relevant and the endpoint-correction design is a reasonable engineering contribution. The ablations are a strength: they isolate the contributions of speed, acceleration, endpoint prediction, and endpoint correction, and the architecture description is coherent. However, the paper's headline significance rests almost entirely on the reported performance gap, and that gap depends on a key input whose source is not specified for either dataset. The manuscript therefore does not currently provide an interpretable comparison of EPN against the baselines, and its reproducibility is limited by the absence of code and by the unresolved construction of P.","major_comments":[{"comment":"The paper never states how the ego planned trajectory P is generated for NGSIM and HighD. Neither dataset contains planner outputs, and Section V-A describes only downsampling, segmentation, and a 7:1:2 split. If P is taken from the recorded future positions of the vehicle treated as the ego vehicle, then EPN receives information about the future during inference while the comparison methods do not. Because Eq. (13) encodes P through an LSTM and Eq. (16) passes it through convolutional pooling into encsocial, the network could exploit the strong correlation between the ego's future motion and the target's future motion, which would make the reported gains an artifact of information leakage. The authors must specify the exact causal procedure that produces P using only information available at time t, and must rerun the comparisons under that procedure; without this, the central empirical claim is uninterpretable.","section":"Section V-A, Eq. (3)"},{"comment":"The baseline numbers are presented without implementation details or sources, so the comparison cannot be reproduced. In particular, S-TF is reported with RMSE values (0.99 at 1s on NGSIM, 0.75 at 1s on HighD) that are much worse than the other learned baselines, and the claimed reductions of 34.9% and 64.6% are computed against this method. The authors should provide the exact evaluation protocol, the hyperparameters, and either code or a precise citation of where each tabulated value comes from. Otherwise the headline reductions and the conclusions drawn from them are not verifiable.","section":"Section V-C, Tables I-III"},{"comment":"The multimodal evaluation selects for each method the trajectory closest to the true future trajectory before computing ADE and FDE. This oracle selection is common in the literature, but the number of candidate trajectories must be matched across methods. The paper states k=6 endpoints, yet the CVAE sampling procedure in Section IV-B can generate any number of endpoint samples via multiple draws from the latent space. The authors should state the total number of output trajectories used for EPN and for each baseline, so that the comparison is not biased by a larger candidate set.","section":"Section V-C, Table III"}],"minor_comments":[{"comment":"\"RSME\" is a typo and should be \"RMSE\" throughout the tables and text.","section":"Throughout, Tables I-II"},{"comment":"Equation (13) uses the variables x_i^t, y_i^t for the ego planned trajectory, but P in Eq. (3) is defined as p_t; the notation should be made consistent so that the ego plan coordinates are clearly distinguished from those of other vehicles.","section":"Section IV-A, Eq. (13)"},{"comment":"The symbols Atar and Aner are used for the region of interest, but later equations use Anei for the set of neighboring vehicles; these notations should be unified.","section":"Section III-A"},{"comment":"The paper refers to a 7:1:2 split by scenario but does not state whether the split is performed at the track/segment level or the trajectory level; this should be clarified for reproducibility.","section":"Section V-A"}],"recommendation":"reject","confidential_remarks":"The decisive issue is that the key input P cannot be sourced from either NGSIM or HighD without either a planner that the paper does not describe or direct use of future recorded positions. The authors promise code only after acceptance, so the reviewer cannot currently check this. In my view this is a load-bearing flaw in the experimental design, not a presentation issue: the headline reductions are not interpretable until the causal construction of P is specified and the experiments are rerun. If the authors can supply a planner that uses only history and still reproduce substantial gains, a resubmission would be worth considering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines two established ideas—PiP's ego-planning input and PECnet-style endpoint-conditioned generation—with a small endpoint-correction module. That combination is new, and the writing is clear. The ablations on NGSIM are useful and suggest each added component contributes something. Credit where due: the architecture is coherent, the loss is standard, and the authors cite the relevant prior work.\n\nThe soft spot is load-bearing. Eq. (3) defines P as the ego vehicle's future planned trajectory, but NGSIM and HighD contain no planner outputs. Section V-A describes preprocessing but never states how P is generated. If P is the ground-truth future ego trajectory, the model gets future information the baselines do not, and the 30–65% reductions in Tables I–III are an artifact of information leakage rather than evidence for the architecture. The paper doesn't address this. Also, the oracle selection of the closest predicted trajectory inflates absolute numbers, and there are no error bars, no code, and no comparison to the cited endpoint-based methods (DenseTNT, ADAPT, DESTINE).\n\nBut don't overstate. The paper does not fit parameters to the test set; the concern is input leakage, not circular fitting. The central conceptual claim—that ego planning helps target prediction—is plausible and consistent with PiP, so the idea isn't wrong. The endpoint correction module is a minor but legitimate extension.\n\nWho is this for? Trajectory-prediction specialists. As presented, the evidence does not support the headline numbers because of the P ambiguity. A serious referee should ask for a precise, non-future-dependent definition of P (e.g., a constant-velocity or lane-keeping planner) and a rerun of the comparisons. If that can be provided, the paper could be a solid incremental contribution. Without it, the results are not interpretable.\n\nI'd send it to peer review rather than desk-reject, because the flaw is identifiable and potentially fixable, and the architecture plus ablations merit scrutiny. But I would not cite the reported numbers as they stand.","headline":"Solid incremental architecture, but the paper never says where the ego planned trajectory comes from on datasets that have no planner, so the headline gains are uninterpretable.","tokens_in":12884,"tokens_out":1749,"would_cite":false,"duration_ms":15486,"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":"EPN predicts target trajectories by conditioning on the ego vehicle's planned future path and on corrected target endpoints, reporting error reductions of about 35% on NGSIM and 64% on HighD over the strongest baselines tested.","keywords":["trajectory prediction","ego vehicle planning","multimodal prediction","endpoint prediction","conditional variational autoencoder","NGSIM","HighD","autonomous driving"],"falsifier":"Take a test split, replace P with a planner output computed only from history (for example, constant-velocity extrapolation or a learned policy), retrain and reevaluate EPN against PiP on the same metrics; if the large error reductions disappear, the reported gains were carried by leaked future information rather than by the architecture.","tokens_in":11840,"feed_emoji":"🚗","tokens_out":4137,"duration_ms":33866,"temperature":0.7,"pith_summary":"The paper argues that trajectory prediction for a target vehicle improves when the predictor is given the ego vehicle's own planned future trajectory as an extra input, alongside a module that first predicts where the target will end up and then corrects that endpoint. It claims this combination, called EPN, produces average reductions of 34.9% in RMSE, 30.7% in ADE, and 30.4% in FDE on NGSIM, and 64.6%, 64.5%, and 64.3% on HighD, compared with the strongest baselines in the experiments. If that is right, feeding planned ego motion and endpoint refinement is a practical way to cut prediction error on highway datasets without relying on high-definition maps.","feed_headline":"Ego plan plus endpoint fix cuts trajectory error 64%","feed_subtitle":"The planning-informed network beats intention-based baselines by 34.9% RMSE on NGSIM and 64.6% on HighD.","key_machinery":"The load-bearing mechanism is the pairing of two inputs: the ego planned trajectory $P$ (Eq. 3), fed through its own LSTM and social-pooling branch, and a target endpoint prediction module that uses a conditional variational autoencoder (CVAE) to sample endpoints, a correction decoder that adds an offset $d_{\\text{offset}}$ to the sampled endpoint, and an LSTM decoder that conditions full trajectories on the corrected endpoint. The correction mechanism is what separates EPN from plain endpoint-conditioned prediction; the ablation attributes a 14.1%, 18.4%, and 21.5% reduction in RMSE, ADE, and FDE to adding the endpoint module and correction on top of the planning-informed baseline.","core_discovery":"The central claim is that multimodal target trajectory prediction errors drop substantially when the predictor conditions on the ego vehicle's planned future trajectory and on corrected target endpoints rather than on discrete driving intentions. The model encodes historical states of the ego, target, and neighboring vehicles through LSTM encoders, pools them with a convolutional social network, and also encodes the ego plan; a conditional variational autoencoder samples candidate endpoints, a correction head predicts an offset between the sampled and true endpoint, and an LSTM decoder turns each corrected endpoint into a full trajectory. The authors report that EPN beats the prior planning-informed model PiP and the intention-based baselines by large margins, with the largest gains on the drone-recorded HighD dataset.","pith_inferences":["A reader should treat the headline numbers as conditional on the ego planned trajectory P being produced honestly from history, since neither NGSIM nor HighD contains planner outputs and the paper does not describe how P was generated.","One direct test of the claim is to replace P with a constant-velocity or kinematic extrapolation from history alone and see whether the reported reductions persist; if they vanish, the gains were carried by future information rather than by the architecture.","The endpoint-correction idea could be attached to other endpoint-driven predictors beyond CVAE, since the correction decoder only needs an initial endpoint estimate and an environmental feature vector.","Deploying EPN in a closed-loop simulator, where the ego plan is genuinely known before execution, would clarify whether the planning input helps in real systems rather than only in offline evaluation."],"forward_implications":["If P is genuinely available at test time, planning-informed prediction can be applied on mapless highway datasets and outperform history-only and intention-based models.","Endpoint correction reduces error more than simply adding more trajectory modalities, according to the ablation results.","On HighD, where positional data is more accurate, the plan encoding and endpoint module cut errors by roughly two-thirds, suggesting better sensor data amplifies the benefit.","The model produces multimodal predictions by sampling the latent space multiple times, with $k=6$ endpoints matching the six intention classes used by the baselines."],"supporting_citations":[{"why":"PiP, the planning-informed baseline that also uses ego vehicle plans; EPN extends and directly compares against it.","marker":"[19]"},{"why":"CS-LSTM, the convolutional social pooling interaction encoder that EPN builds on and uses as a baseline.","marker":"[14]"},{"why":"PECnet, the endpoint-conditioned prediction work that supplies the truncation trick ($\\sigma_T=1.3$) for sampling endpoints.","marker":"[28]"},{"why":"HighD, the drone-recorded highway dataset used for training and evaluation.","marker":"[31]"},{"why":"NGSIM US Highway 101 dataset used for training and evaluation.","marker":"[29]"},{"why":"NGSIM I-80 dataset used for training and evaluation.","marker":"[30]"},{"why":"S-TF, the sparse Transformer intention-based baseline against which EPN reports its average RMSE reduction.","marker":"[23]"},{"why":"S-LSTM, the social LSTM baseline with fully connected social pooling.","marker":"[32]"}],"fun_headline_variants":["Ego plan plus corrected endpoints cut trajectory error 64%","Planning-informed network cuts prediction error by 64% on HighD","Target trajectory prediction improved with ego plan and endpoint fix","EPN ego planning and endpoint correction improve trajectory prediction","Ego's future plan and corrected endpoints reduce error 64%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments treat the ego vehicle's planned future trajectory P as known input on datasets that do not contain planner outputs, and the paper does not say how P was produced; if P is actually the ground-truth future ego trajectory, the comparison hands EPN future information that the baselines lack.","fun_headline_variants_meta":{"raw":{"variants":["Ego plan plus corrected endpoints cut trajectory error 64%","Planning-informed network cuts prediction error by 64% on HighD","Target trajectory prediction improved with ego plan and endpoint fix","EPN ego planning and endpoint correction improve trajectory prediction","Ego's future plan and corrected endpoints reduce error 64%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00125,"raw_usage":{"total_tokens":5121,"prompt_tokens":937,"completion_tokens":4184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":4100}},"tokens_in":553,"tokens_out":4184,"duration_ms":22649,"temperature":1.0,"reasoning_tokens":4100,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:14:12.550777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test split, replace P with a planner output computed only from history (for example, constant-velocity extrapolation or a learned policy), retrain and reevaluate EPN against PiP on the same metrics; if the large error reductions disappear, the reported gains were carried by leaked future information rather than by the architecture.","supporting_citations":[{"cited_title":"Pip: Planning-informed trajectory prediction for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"PiP, the planning-informed baseline that also uses ego vehicle plans; EPN extends and directly compares against it."},{"cited_title":"Convolutional social pooling for vehicle trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"CS-LSTM, the convolutional social pooling interaction encoder that EPN builds on and uses as a baseline."},{"cited_title":"It is not the journey but the destination: Endpoint conditioned trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"PECnet, the endpoint-conditioned prediction work that supplies the truncation trick ($\\sigma_T=1.3$) for sampling endpoints."},{"cited_title":"The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,","cited_arxiv_id":null,"evidence_quote":"HighD, the drone-recorded highway dataset used for training and evaluation."},{"cited_title":"Us highway 101 dataset. federal highway administration,","cited_arxiv_id":null,"evidence_quote":"NGSIM US Highway 101 dataset used for training and evaluation."},{"cited_title":"Us highway i-80 dataset. federal highway administration,","cited_arxiv_id":null,"evidence_quote":"NGSIM I-80 dataset used for training and evaluation."},{"cited_title":"Multimodal trajectory prediction based on sparse weight sharing,","cited_arxiv_id":null,"evidence_quote":"S-TF, the sparse Transformer intention-based baseline against which EPN reports its average RMSE reduction."},{"cited_title":"Social lstm: Human trajectory prediction in crowded spaces,","cited_arxiv_id":null,"evidence_quote":"S-LSTM, the social LSTM baseline with fully connected social pooling."}],"review_version":1}