{"id":"e454aef7-01f3-48c7-9732-c88a8fb69245","arxiv_id":"2505.00586","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ParkDiffusion predicts multi-modal trajectories for heterogeneous agents in parking scenarios, reporting state-of-the-art accuracy on the DLP and inD datasets.","lead":"ParkDiffusion is a generative model that forecasts six possible future paths for every car and pedestrian in a parking lot, conditioning on the lot's map and each agent's type. It reports sharply lower prediction errors than existing methods on two datasets, with caveats about baseline fairness and reproducibility.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperforms all baselines by a large margin' claim rests on Table I, where ParkPredict+ is quoted from its own paper and every retrained baseline gets only 20 epochs; if the baselines are undertrained or the ParkPredict+ evaluation protocol differs, the reported margins are inflated.","rationale":"I read the central claim as the empirical benchmark claim in Table I, supported by the novelty claim of being the first heterogeneous VRU-aware parking trajectory predictor. The method is architecturally plausible and the ablations are informative, but the benchmark is the weakest link. The paper explicitly states in Section IV-B that ParkPredict+ is reported from published results and that all other baselines receive 20 epochs. Since minADE and minFDE are protocol-sensitive, a cross-paper comparison with an unmatched protocol and possibly undertrained baselines cannot support the phrase 'by a large margin' without additional evidence. This is a correctness risk, not a style or novelty complaint. It is addressable: retrain baselines fairly, report variance, and release code or hyperparameters. If the margins persist under a matched protocol, the headline claim is supported. I agree with the reader's weakest-assumption identification, so the CONDITIONAL verdict should remain; the decisive additional check is an exactly matched ParkPredict+ comparison.","tokens_in":12481,"tokens_out":3382,"duration_ms":37913,"concrete_test":"Re-run the DLP comparison with ParkPredict+ retrained under the authors' own preprocessing (using released code if available), train all baselines to convergence (following each original paper's schedule plus early stopping on the validation set), and report results over at least 3 random seeds as mean plus/minus standard deviation. Then recompute Table I and check whether ParkDiffusion's minADE and minFDE margins remain larger than the seed-to-seed spread. In addition, verify that the quoted ParkPredict+ numbers were computed for the same 4 s horizon, 0.4 s sampling, 20 m radius, and metric definition; if they were not, requote the baseline under an exactly matched protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim depends on the fairness of the baseline comparison in Section IV-B. Two features undermine it. First, ParkPredict+ is not integrated into the authors' preprocessing pipeline; its numbers are simply taken from the original paper because it is 'challenging to reproduce' it in their setting. This makes Table I a cross-pipeline comparison rather than a controlled one. The DLP data split is stated to match ParkPredict+ (51,750/5,750), but no evidence is given that the observation history length, prediction horizon, sampling interval, ego-centric radius, or metric implementation are identical. minADE and minFDE are sensitive to all of these choices, so the ParkPredict+ row may not be comparable. Second, all other baselines are trained for 20 epochs 'to ensure convergence,' with no learning-rate schedule, convergence curves, or early-stopping criterion reported. QCNet and SIMPL are large attention-based models that may require more epochs or different tuning in a new pipeline. If those baselines are undertrained, the large margins in Table I partly reflect asymmetric effort. The absence of repeated-seed variance or significance testing further means the point estimates cannot be distinguished from noise. The core architecture remains plausible, but the load-bearing evidence for the headline outperformance is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ParkDiffusion, a diffusion-model-based trajectory prediction method for heterogeneous agents (vehicles and pedestrians) in automated parking scenarios. The method combines a dual map encoder for soft and hard map constraints, an adaptive agent type embedding, a Leapfrog Diffusion initializer, and physics-based kinematic refinement. The authors evaluate on the Dragon Lake Parking (DLP) dataset and the Intersections Drone (inD) dataset, reporting strong improvements over several baselines and ablations that show the contribution of each component. The central claim is state-of-the-art performance on parking-scene trajectory prediction, with particular emphasis on pedestrian prediction.","tokens_in":12890,"tokens_out":4861,"duration_ms":52558,"significance":"If the reported results hold, ParkDiffusion would be a useful step for automated parking trajectory prediction, an underexplored area relative to urban driving. The architecture is plausible, and the paper has clear strengths: explicit consideration of VRUs, a dual map encoder that separates soft and hard constraints, an agent type embedding that conditions on agent class, a kinematic refinement layer that improves feasibility, and a set of ablations including context masking and agent-density analysis. The use of two datasets and the inclusion of scenario-specific analyses are also positive. However, the empirical evaluation as presented is not yet at the level needed to support the headline claims. The baseline comparison is not fully controlled, the training budget assertion is unsupported, and the absence of statistical significance testing leaves the reported margins difficult to evaluate. The kinematic feasibility claim is also stronger than the implemented point-mass model justifies. With these issues addressed, the work could become a solid contribution.","major_comments":[{"comment":"The headline claim that ParkDiffusion 'outperforms all baselines across the board' rests on an uncontrolled comparison. ParkPredict+ is not integrated into the authors' preprocessing pipeline; its numbers are quoted from the original paper because it is 'challenging to reproduce' it in the authors' setting. No evidence is provided that the observation history length, prediction horizon, sampling interval, ego-centric radius, number of modes K, or metric implementation match the authors' setup. Table I also reports no pedestrian or 'all' metrics for ParkPredict+, so the 'across the board' claim is not fully supported even by the table as printed.","section":"IV-B / Table I"},{"comment":"The statement that all baselines are trained for 20 epochs 'to ensure convergence' is not supported by any convergence curves, learning-rate schedules, early-stopping criteria, or repeated-seed variance. With single-run point estimates and no significance testing, the large margins in Table I could partly reflect asymmetric hyperparameter effort or incomplete training of the baselines. The paper should retrain all baselines under matched training budgets and report mean and standard deviation over multiple seeds, or otherwise justify that the reported single runs are representative.","section":"IV-B"},{"comment":"The vehicle kinematic model is a point-mass double integrator with only an acceleration norm bound (Eq. 11). It does not constrain heading, steering angle, or turning radius, so the claim that the model produces 'physically feasible' trajectories (Contribution 3) is overstated. The authors should either adopt a kinematic bicycle model or explicitly discuss why the point-mass approximation is sufficient for the parking maneuvers considered, including reversing and tight turns.","section":"III-D / Eq. (11)"},{"comment":"The very low absolute errors on DLP (e.g., pedestrian minADE of 0.15 m and minFDE of 0.32 m over a 4 s horizon) suggest that a large fraction of the samples may be near-stationary. Without reporting the proportion of static agents or the speed distribution in the validation set, the practical significance of the margin over baselines is unclear. The authors should quantify the share of stationary trajectories and, if substantial, report metrics separately for moving and stationary agents.","section":"Table I"}],"minor_comments":[{"comment":"The sentence 'Overall, our method achieves the best minADE and minFDE scores compared to all baselines' after Table II is ambiguous because SIMPL has a lower vehicle minADE. The claim should explicitly state that it refers to the combined 'All' metrics, where it is supported by the table.","section":"IV-C"},{"comment":"The weighted L2 norm in Eq. (18) is not defined: the weights W and how they are chosen are not specified. The cross-entropy target formation for the probability branch is also described only in words; please provide the exact loss formulation.","section":"III-F"},{"comment":"The notation \\tilde{Y} is used both for the candidate trajectory after kinematic refinement and for the noisy input to the denoiser. Please clarify the relationship and define the diffusion schedule parameters αβ, ᾱβ, and σβ explicitly, including any dependence on the timestep β.","section":"III-E / Eq. (16)"},{"comment":"For the inD dataset, the paper states that Bendplatz and Frankenburg are merged due to their high number of VRUs, but it does not state whether this matches the official data splits or how the train/validation split was made. Please specify the exact split procedure and the number of agents of each class in each split.","section":"IV-B"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the evaluation protocol: the direct comparison with ParkPredict+ using published numbers is not a controlled experiment, and the assertion that 20 epochs suffices for convergence is unsubstantiated. If the authors cannot reimplement ParkPredict+ under their own preprocessing, they should clearly separate claims about their own retrained baselines from claims about the published ParkPredict+ numbers. The novelty claim of being 'the first' to consider VRUs in parking scenarios should also be checked against recent parking-related trajectory prediction work; this is a scope issue rather than a blocker."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: solid architecture, a real niche, and experiments that are good enough to publish after a fair baseline rerun, but not good enough to justify the 'large margin' headline as written.\n\nWhat's new: the combination of a Leapfrog-style diffusion predictor with dual map encoding, agent-type embedding, and kinematic refinement for heterogeneous cars and pedestrians in parking scenarios. That combination is genuinely new, and the claim of being the first to explicitly handle VRUs in parking prediction holds up — ParkPredict+ is vehicle-only, and generic urban predictors are not set up for parking-lot free-for-all semantics. The ablations are a plus: each added module buys a clear, monotone improvement, and the mask-perturbation and agent-density robustness tests give a decent sense of failure modes. The inD transfer shows the method is not memorizing one lot.\n\nSoft spots, in order:\n1. The headline comparison in Table I is not controlled. ParkPredict+ numbers are lifted from its own paper; no evidence that history length, horizon, sampling rate, ego radius, or metric implementation match. That is a real flaw, and it is acknowledged in the text, but the caption still says 'outperforms ... by a large margin' over that row.\n2. Baselines get 20 epochs with no convergence curves or hyperparameter tuning. QCNet and SIMPL are big models; undertraining would inflate the gap. This is addressable by rerunning with proper schedules and reporting seeds.\n3. No error bars, no significance tests. Point estimates alone cannot support the 'large margin' language.\n4. Minor: the inD evaluation merges only two of four scenes; the selection is unexplained, and the DLP train/val split follows ParkPredict+ but is not independently verified.\n\nNone of these kill the architecture. The method is plausible, the losses are standard, and the ablation story is consistent. But the empirical claims are currently overstrong. I would send this to a competent referee, and I would conditionally accept it after a fair baseline comparison, variance reporting, and code release.","headline":"A credible diffusion-based parking predictor with honest ablations, but the headline baseline comparison is not yet apples-to-apples.","tokens_in":13271,"tokens_out":2748,"would_cite":false,"duration_ms":29230,"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":"The paper's core claim is that parking-lot trajectory prediction improves substantially when a diffusion model conditions on soft and hard map constraints, adapts to agent type, and outputs kinematically feasible control signals.","keywords":["trajectory prediction","diffusion models","automated parking","heterogeneous multi-agent","multi-modal prediction","kinematic constraints","pedestrian prediction","vulnerable road users"],"falsifier":"Retrain QCNet and SIMPL under the paper's exact preprocessing and train them for substantially more than 20 epochs with several random seeds, while rerunning ParkDiffusion with the same seeds; if a baseline's minADE or minFDE reaches or falls below ParkDiffusion's reported 0.16 m / 0.30 m at convergence, the large-margin claim is refuted. Alternatively, if ParkDiffusion's per-seed error bars overlap the baseline numbers, the ranking is not statistically stable.","tokens_in":12297,"feed_emoji":"🚗","tokens_out":10404,"duration_ms":94246,"temperature":0.7,"pith_summary":"The paper tries to establish that trajectory prediction in automated parking is best done as a heterogeneous, multi-modal forecasting problem: one model should predict both cars and pedestrians, offer several plausible futures, and force every hypothesis through a physical motion model. ParkDiffusion does this by diffusing control signals instead of raw positions, then integrating those signals with kinematics tuned separately for vehicles and pedestrians, while a pretrained denoiser cleans the resulting trajectories. On the Dragon Lake Parking (DLP) dataset the paper reports an all-agent minADE of 0.16 m and minFDE of 0.30 m, about half the error of the strongest adapted automotive baselines, and on the inD intersection dataset it reports the best overall accuracy, driven mainly by pedestrian prediction. If these numbers hold, parking systems gain a predictor that covers the most safety-critical interaction in lots: a vehicle and a pedestrian sharing the same unmarked space, rather than treating parking as a vehicle-only problem.","feed_headline":"ParkDiffusion beats all baselines on parking-lot trajectory prediction","feed_subtitle":"Forecasting cars and pedestrians jointly, it reports 0.16 m minADE on the DLP parking dataset.","key_machinery":"The load-bearing mechanism is the kinematic refinement loop placed between the diffusion initializer and the denoiser. Instead of generating positions directly, the model lets the Leapfrog Diffusion (LED) initializer produce control signals; vehicles are integrated as point masses with acceleration bounded by the road-friction limit $\\|u_a(t)\\| \\le \\mu g$ with $\\mu = 0.7$ and $g = 9.81\\ \\mathrm{m/s^2}$, while pedestrians are integrated through a first-order neural ODE. A two-step cross-attention map encoder first conditions agent features on soft lane polylines and then on hard obstacle polylines, and an agent-type embedding modulates those features by scaling and shifting them. This gives the candidate trajectories a physical plausibility that a single pretrained denoiser then refines for all agent types.","core_discovery":"The core claim is that a diffusion-based trajectory predictor can be made parking-specific by splitting the scene into three conditionings: soft map semantics (lane markings that may be crossed), hard geometric obstacles (parked cars that must be avoided), and the agent's type (vehicle or pedestrian). The paper's distinct discovery is the way these conditions enter the generative process: a two-step cross-attention fuses agent features with soft then hard map polylines; a modulation layer scales and shifts the fused features by a learned agent-type embedding; and the decoder outputs control signals that are integrated under vehicle point-mass dynamics or a pedestrian neural ODE before the denoiser refines the path. The authors report that this combination produces the first heterogeneous, VRU-inclusive parking trajectory predictor and beats all compared baselines on the DLP dataset, with the largest gains on pedestrians.","pith_inferences":["The paper leaves open whether the soft-then-hard cross-attention order is essential; swapping the order would test whether the ordering itself carries the map-conditioning benefit.","The same soft/hard context split and type modulation could transfer to other shared-space domains, such as loading docks, garage ramps, or pedestrianized city centers, where the parking-specific kinematic models would need to be adapted.","A direct way to isolate the value of the learned pedestrian dynamics is to run the pedestrian branch with the same point-mass integrator used for vehicles; if pedestrian error stays low, the neural ODE is not the reason."],"forward_implications":["Parking-specific prediction no longer has to ignore pedestrians; the same model can forecast both classes, and the paper's largest margins are in pedestrian accuracy.","Diffusion models can run within an ADAS pipeline if an initializer skips most denoising steps and the denoiser only refines a kinematically plausible proposal.","Treating parked cars as hard constraints and lane markings as soft constraints, then fusing them in that order, is a reusable recipe for map conditioning in parking-like spaces.","Because the model outputs control signals rather than positions, the same pretrained denoiser can be paired with different kinematic models for different agent types."],"supporting_citations":[{"why":"Provides the Leapfrog Diffusion (LED) initializer and denoiser structure that ParkDiffusion builds on.","marker":"[11]"},{"why":"Supplies the DLP parking dataset and the ParkPredict+ baseline whose published numbers ParkDiffusion is compared against.","marker":"[14]"},{"why":"Supplies the inD intersection dataset used as a generalization benchmark.","marker":"[15]"},{"why":"QCNet baseline adapted for the parking comparison.","marker":"[7]"},{"why":"SIMPL baseline, the strongest competitor on vehicle metrics in the inD evaluation.","marker":"[8]"},{"why":"SceneTransformer baseline adapted for the parking comparison.","marker":"[16]"},{"why":"MultiPath++ baseline adapted for the parking comparison.","marker":"[17]"}],"fun_headline_variants":["ParkDiffusion predicts cars and pedestrians in parking lots","Diffusion model for heterogeneous parking trajectory prediction","ParkDiffusion beats baselines on parking lot forecasting","Heterogeneous parking prediction via diffusion with kinematic controls","Diffusion-based parking predictor handles cars and pedestrians jointly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported margin over prior work assumes that every baseline was trained to convergence on the same 20-epoch schedule with equivalent preprocessing, and that the published ParkPredict+ numbers were produced under conditions comparable to the authors' pipeline; if those assumptions fail, the size of the reported advantage shrinks.","fun_headline_variants_meta":{"raw":{"variants":["ParkDiffusion predicts cars and pedestrians in parking lots","Diffusion model for heterogeneous parking trajectory prediction","ParkDiffusion beats baselines on parking lot forecasting","Heterogeneous parking prediction via diffusion with kinematic controls","Diffusion-based parking predictor handles cars and pedestrians jointly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1356,"prompt_tokens":929,"completion_tokens":427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":545,"tokens_out":427,"duration_ms":5040,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:37:56.055831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain QCNet and SIMPL under the paper's exact preprocessing and train them for substantially more than 20 epochs with several random seeds, while rerunning ParkDiffusion with the same seeds; if a baseline's minADE or minFDE reaches or falls below ParkDiffusion's reported 0.16 m / 0.30 m at convergence, the large-margin claim is refuted. Alternatively, if ParkDiffusion's per-seed error bars overlap the baseline numbers, the ranking is not statistically stable.","supporting_citations":[{"cited_title":"Leapfrog diffusion model for stochastic trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Provides the Leapfrog Diffusion (LED) initializer and denoiser structure that ParkDiffusion builds on."},{"cited_title":"ParkPredict+: Multimodal intent and motion prediction for vehicles in parking lots with CNN and transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the DLP parking dataset and the ParkPredict+ baseline whose published numbers ParkDiffusion is compared against."},{"cited_title":"The inD dataset: A drone dataset of naturalistic road user trajectories at german intersections,","cited_arxiv_id":null,"evidence_quote":"Supplies the inD intersection dataset used as a generalization benchmark."},{"cited_title":"Query-centric trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"QCNet baseline adapted for the parking comparison."},{"cited_title":"SIMPL: A simple and efficient multi-agent motion prediction baseline for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"SIMPL baseline, the strongest competitor on vehicle metrics in the inD evaluation."},{"cited_title":"Scene Transformer: A unified architecture for predicting future trajectories of multiple agents,","cited_arxiv_id":null,"evidence_quote":"SceneTransformer baseline adapted for the parking comparison."},{"cited_title":"Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,","cited_arxiv_id":null,"evidence_quote":"MultiPath++ baseline adapted for the parking comparison."}],"review_version":1}