{"id":"5471cfe8-7d5b-407a-b9df-dc98ec7b8da0","arxiv_id":"2505.07855","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural occupancy predictor trained to imitate artificial potential field maps is reported to improve planning metrics over the APF baseline, but the comparison is confounded because the training target is the baseline itself.","lead":"This paper trains a convolutional and recurrent neural network to reproduce occupancy maps generated by artificial potential fields (APF), then evaluates the resulting planner against APF on 2,000 CommonRoad scenarios. It claims higher task completion, larger safety margins, and fivefold lower execution time, but the evaluation is weakened by a teacher-student circularity and missing experimental details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation is circular and lacks a train/test split, so the reported gains over APF are uninterpretable.","rationale":"The decisive issue is that the experimental section cannot establish the stated central claim, because the only reported comparison is between a network trained to imitate APF and APF itself, with no evidence that the test scenarios are disjoint from training. The paper's own Fig. 2 shows a validation loss, which implies a split exists, but Section IV never states how the 2,000 CommonRoad scenarios were partitioned, nor the size of the validation set. If the same scenarios were used for both training and evaluation, the TCR/TTC/jerk numbers in Table I are consistent with memorization: the network can recall the APF ideal map for each input and the downstream planner then produces trajectories that are at best APF-quality, not better. If a proper split was used, the paper still does not specify the trajectory-generation procedure for 'Ours' or the parameters of the APF baseline, so the comparison is not reproducible. The internal paradox reinforces this: at convergence of Eq. (9), the predicted occupancy map equals the APF ideal map, so under identical planning the network should match, not beat, APF. Therefore any reported improvement must come from a source other than the physics-informed supervision as described, such as a mismatched baseline or unconstrained network deviations. No code, data split, or planner details are provided, so independent verification is impossible. The core idea may be salvageable, but the current evidence does not support the abstract's claim of improved task completion, safety margins, and efficiency.","tokens_in":8429,"tokens_out":3571,"duration_ms":38736,"concrete_test":"Re-run the evaluation with an explicit held-out split: train on a random subset of, say, 1,500 CommonRoad scenarios and test on the remaining 500, reporting the same metrics (TCR, TTC, Jerk, Headway, Time) separately for seen and unseen scenarios. In addition, take the APF ideal occupancy map O* for each test scenario and run the exact same trajectory-generation procedure used for 'Ours' on O*; if the network's map does not beat O* under identical downstream planning, the comparison against APF is not evidence of improvement.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that the learned occupancy maps improve planning over APF. But the learning target in Eq. (9) is the APF-generated ideal occupancy map O*, and the baseline in Table I is APF/CommonRoad-Reactive-Planner. If training converges, the network reproduces O*, so applying the same planner to the predicted map should recover APF behavior, not beat it. The reported gains can only arise from (i) evaluating on scenarios used in training (no train/test split is reported anywhere; Section IV only says '2000 CommonRoad scenarios'), (ii) a mismatch between the APF map used to generate labels and the APF planner used as baseline, or (iii) unconstrained deviations from the APF target that happen to improve metrics, which would contradict the claim that APF guidance is the cause. Section IV.D also switches baseline names between 'APF' and 'CommonRoad-Reactive-Planner', and no trajectory-generation or planning details are given for 'Ours'. Without a held-out split and a matched planner, the improvements are uninterpretable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end occupancy prediction framework for autonomous vehicle motion planning. A convolutional-recurrent network takes a binary scene map as input and outputs an occupancy map of shape [B, 36, 9]. The training objective, Eq. (9), is an MSE loss that regresses the network output toward an 'ideal occupancy map' generated from an artificial potential field (APF) defined by Eqs. (6)-(8). The authors claim that embedding APF as a physics-informed constraint improves safety, smoothness, and efficiency, and they report experiments on 2,000 CommonRoad scenarios in which their method outperforms an APF/CommonRoad-Reactive-Planner baseline on task completion rate, TTC, jerk, headway, and execution time (Table I). The paper also provides qualitative comparisons of potential-field maps and two sample trajectories.","tokens_in":8683,"tokens_out":4598,"duration_ms":48923,"significance":"If the empirical claims were supported, the framework would be a useful, modular way to combine a learned occupancy predictor with a fast planner, and the reported five-fold speedup would be practically interesting. The paper has some strengths: the network architecture is clearly described in Eqs. (1)-(5), the loss is explicit, and the evaluation is conducted on a standard benchmark (CommonRoad). However, the central claim is not established by the reported evidence. The training target is the APF-generated map, and the baseline is APF, so the comparison is between APF and a learned approximation of APF. Without a held-out train/test split, a matched planner description, and uncertainty quantification, the reported improvements over APF are uninterpretable. The contribution as stated, namely that physics-informed guidance during training leads to safer planning, is therefore not supported by the manuscript in its current form.","major_comments":[{"comment":"The training target O* in Eq. (9) is the APF-generated ideal occupancy map, and the only baseline in Table I is APF (or the CommonRoad-Reactive-Planner; see the next comment). A network trained to minimize this MSE can at best reproduce APF's occupancy field, so the reported gains in TTC, jerk, and headway are not explained by the stated mechanism. If the gains instead arise from imperfect regression, they contradict the claim that APF guidance is the cause. In addition, the paper does not report any train/test split for the 2,000 CommonRoad scenarios, so the improvements could also be an artifact of evaluating on scenarios used in training.","section":"§III.B, Eq. (9), and Table I"},{"comment":"The baseline is called 'CommonRoad-Reactive-Planner' in the text but 'APF' in Table I, and the paper does not state how either planner consumes the predicted occupancy map to generate trajectories. Without a precise specification of the planning algorithm, the inputs, and the matching between the two methods, the metric differences cannot be attributed to the proposed occupancy prediction, and the experiments cannot be reproduced.","section":"§IV.D and Table I"},{"comment":"The reported metrics are single-point values with no standard deviations, no multiple seeds, and no statistical significance tests. The differences are moderate (e.g., TTC 2.798 s vs 2.979 s, task completion rate 0.902 vs 0.946), and with no uncertainty quantification the claimed improvements are not statistically grounded.","section":"§IV.D, Table I"},{"comment":"The paper describes an occupancy-prediction network but never specifies the trajectory-generation or planning module that converts a predicted occupancy map into the trajectories used to compute TTC, jerk, headway, and completion rate. Since these quantities are properties of trajectories, the absence of a planning specification makes the experimental results unverifiable and prevents the reader from assessing whether the comparison is fair.","section":"§III and §IV"}],"minor_comments":[{"comment":"The text says the loss is 'the sum of all elements in the difference' between the two maps, but Eq. (9) defines an average MSE over H*W cells; these two descriptions should be aligned.","section":"§IV.A and Eq. (9)"},{"comment":"Equation (5), Z_t = GRU(LSTM(F3, h_{t-1}), h_t), is syntactically unclear because h_t appears as an input before it is computed; the state-update equations for the LSTM and GRU should be written explicitly.","section":"§III.A, Eq. (5)"},{"comment":"The x-axis of Fig. 2 is called 'epochs' in the caption but 'total number of training iterations' in the text; please use consistent terminology and specify the number of epochs and the dataset sizes.","section":"§IV.A and Fig. 2"},{"comment":"The text says the model's trajectory is compared with 'the ground truth,' but Fig. 4 shows only predicted trajectories of APF and the proposed method; the ground-truth trajectory should be defined and plotted.","section":"§IV.C"},{"comment":"The APF potential is a geometric heuristic, not a physical law of vehicle motion; calling it 'physics-informed' is misleading unless the potential is connected to vehicle dynamics or a physical safety constraint.","section":"§III.B"},{"comment":"There are several typographical issues, including 'Head-Way' in Table I, 'A Vs' and 'A V' spacing in the abstract and introduction, and 'face limitations' in §IV.C; these should be corrected in a revision.","section":"Table I and throughout"}],"recommendation":"reject","confidential_remarks":"I agree with the reader's assessment: the evaluation is circular and the baseline/planner details are missing. The paper would need a substantially new experimental setup, including a proper train/test split, a matched non-APF baseline or an independent downstream planner, and statistical uncertainty quantification, before the central claim could be assessed. In its current form, the empirical results do not support the stated conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the central comparison doesn't support the conclusion. The network is trained to regress APF-generated occupancy maps (Eq. 9), and the baseline in Table I is APF. If training converges, the network reproduces the teacher, so you'd expect similar planning metrics, not better. The reported gains in TTC, jerk, and headway only make sense if the network is evaluated on scenarios it memorized, and the paper never reports a train/test split. The baseline label also switches between \"APF\" and \"CommonRoad-Reactive-Planner\" in Section IV.D, so it's genuinely unclear what was compared.\n\nTo be fair, the architecture is simple and reasonable: conv layers with a residual connection, plus LSTM/GRU for temporal context. Using APF as a training target is a natural idea, and the CommonRoad benchmark is a sensible testbed. The execution time improvement (0.0019 s vs 0.01 s) is plausible because a small network forward pass can be faster than recomputing potential fields. The visualizations do show smoother, more adaptive fields, though that's qualitative.\n\nThe real problem is the logic of the comparison. The paper explicitly says the model \"accurately mimic[s]\" APF maps, then claims it outperforms APF on safety and comfort. If the network deviates from the target in ways that improve metrics, then the APF guidance isn't the cause of the improvement; if it stays close to the target, the metrics should match APF. Either way, the claim that physics-informed guidance improves planning is unsupported. There are no error bars, and the \"2000 CommonRoad scenarios\" are described without any split details. Figure 2 shows training and testing loss, so a split exists, but the paper never says how it was made.\n\nThis is a paper a serious referee could help fix. The idea is worth exploring, but the evaluation must be redone with a held-out split, a matched planner, and a direct comparison against the APF teacher. I wouldn't accept it in its current form, but I'd send it to review with the expectation of major revision.","headline":"A network trained to regress APF occupancy maps can't claim to beat APF without a held-out split and a matched planner.","tokens_in":9178,"tokens_out":2905,"would_cite":false,"duration_ms":29921,"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 claims that an end-to-end occupancy network, trained to match artificial potential field maps, improves autonomous vehicle task completion, safety margins, planning smoothness, and speed compared with the rule-based APF planner.","keywords":["occupancy prediction","artificial potential fields","physics-informed learning","autonomous driving","motion planning","end-to-end","LSTM","safety metrics"],"falsifier":"Re-run the evaluation on a held-out set of CommonRoad scenario IDs that are provably absent from training, and check whether the 0.946 versus 0.902 completion advantage and the TTC and headway gains persist; if the advantage disappears or reverses on unseen scenarios, the claim of generalization from physics-informed learning is falsified. A second, faster check is to inspect the training curves in Fig. 2 for a large gap between training and validation loss on the same scenario distribution, which would indicate memorization.","tokens_in":8238,"feed_emoji":"🚗","tokens_out":6178,"duration_ms":58312,"temperature":0.7,"pith_summary":"This paper claims that a small end-to-end network can learn to turn a binary map of a driving scene into an occupancy map that encodes safety, and that this predicted map supports better motion planning than the rule-based artificial potential field (APF) method that was used to supervise it. The authors train a convolutional plus recurrent network to reproduce APF-generated ideal occupancy maps, so the network learns distance-based repulsion and attraction as part of its learned representation rather than as a post hoc correction. On 2,000 CommonRoad scenarios they report a higher task completion rate (0.946 vs 0.902), longer time-to-collision, larger headway, lower jerk, and an average execution time of 0.0019 seconds, about five times faster than the baseline. The paper's central proposal is that physical priors can be baked into the learning objective without sacrificing model flexibility or real-time operation.","feed_headline":"Trained on physics rules, occupancy network plans 5x faster and safer","feed_subtitle":"On 2,000 traffic scenarios it beats the APF baseline on completion, time-to-collision, jerk, and headway.","key_machinery":"The load-bearing object is the APF-generated ideal occupancy map used as the regression target. Artificial potential fields define total potential $U(x)=U_{\\text{att}}(x)+U_{\\text{rep}}(x)$ with quadratic attraction toward the goal and inverse-distance repulsion within a threshold $d_0$ (Eqs. 6–8), and the normalized squared difference between the network's predicted map $\\hat{O}_{i,j}$ and the ideal map $O^*_{i,j}$ is the training loss (Eq. 9). This single mechanism injects physical rules directly into learning: the network never sees the APF rule at inference time, yet its occupancy maps are shaped by distance-to-obstacle safety margins. The framework is presented as modular, so other physical priors could replace APF without changing the network architecture.","core_discovery":"The central claim is that an occupancy predictor trained by regression against APF-generated ideal maps inherits the physical plausibility of the potential field while keeping the adaptivity of a learned model. The network, built from residual convolutional layers followed by LSTM and GRU units, outputs a 36×9 occupancy map; training minimizes the mean squared error between its output and an ideal map computed from attractive and repulsive potentials (Eqs. 6–9). Because the ideal map assigns high cost near obstacles and low cost along safe corridors, the network learns spatial risk patterns rather than raw obstacle positions. The authors argue this makes predicted occupancy physically plausible and therefore usable for planning, and they support the claim with scenario visualizations and aggregate metrics: task completion 0.946 vs 0.902, TTC 2.979 s vs 2.798 s, jerk 1.361 vs 2.079 m/s³, headway 21.124 m vs 19.054 m, and execution time 0.0019 s vs 0.01 s.","pith_inferences":["Editorial inference: Because the supervision signal is APF-generated, the network's safety behavior cannot exceed what APF's Euclidean-distance logic encodes; if a scenario needs context-dependent rules such as lane priority, blind spots, or agent intent, the current training targets would not teach them, and a testable extension is to train the same architecture on hand-annotated or learned risk ","Editorial inference: The fivefold speed-up likely comes from replacing an iterative planner with one network forward pass, so a similarly fast planner could be obtained by distilling the same APF targets into any small network; the distinctive contribution is the distillation scheme and the learned smoothness, not the planner itself.","Editorial inference: A direct ablation, training the identical architecture against raw binary occupancy targets instead of APF targets, would isolate how much of the reported gains come from physics-informed supervision, but the paper does not report this comparison.","Editorial inference: The 36×9 output grid is coarse relative to real sensor ranges, and scaling the approach to larger maps with more agents may require structured sparsity or attention, which the current convolutional and recurrent design does not address."],"forward_implications":["A single lightweight forward pass can replace the classical reactive planner, since average planning time drops to 0.0019 seconds, leaving ample margin for replanning in dynamic scenes.","Predicted occupancy maps carry learned safety margins, reflected in longer time-to-collision and headway, without explicit collision checking at inference.","Network-produced potential fields are smoother and more adaptive than hand-designed APF fields, yielding lower jerk and therefore a more comfortable ride.","The physical-rule module is exchangeable, so substituting other rules for APF could adapt the same architecture to different safety constraints or task requirements.","The higher task completion rate across 2,000 CommonRoad scenarios suggests the learned representation generalizes across the diverse driving situations tested, supporting the case for deployment."],"supporting_citations":[{"why":"Supplies the deep-learning dynamic occupancy grid prediction approach that this work extends by adding physics-informed supervision.","marker":"[11]"},{"why":"Supplies the CNN plus LSTM temporal-occupancy architecture that the proposed network adapts.","marker":"[12]"},{"why":"Motivates the core idea that physics constraints can improve neural motion prediction models.","marker":"[13]"},{"why":"Represents the physics-informed safety controller approach that applies physics outside the network, which this paper contrasts with its end-to-end integration.","marker":"[14]"},{"why":"A recent end-to-end occupancy and flow prediction model positioned as lacking physical guidance and entity-level modeling.","marker":"[18]"},{"why":"Provides the time-to-collision safety metric used in the holistic evaluation.","marker":"[22]"},{"why":"Provides the task completion rate metric used to compare success across scenarios.","marker":"[23]"}],"fun_headline_variants":["Physics-aware occupancy net boosts AV safety and speed","APF-guided learning makes occupancy planning safer, faster","Physics-informed AI plans AV routes with fewer risks","Learned occupancy plus APF yields safer navigation","Physics-guided occupancy nets plan AVs 5x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the 2,000 CommonRoad evaluation scenarios were not part of the network's training data, but the paper reports no train/test split; if the network saw those scenarios, its higher task-completion rate could reflect memorization rather than learned physics.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware occupancy net boosts AV safety and speed","APF-guided learning makes occupancy planning safer, faster","Physics-informed AI plans AV routes with fewer risks","Learned occupancy plus APF yields safer navigation","Physics-guided occupancy nets plan AVs 5x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3109,"prompt_tokens":895,"completion_tokens":2214,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2140}},"tokens_in":511,"tokens_out":2214,"duration_ms":16411,"temperature":1.0,"reasoning_tokens":2140,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:15:22.218229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation on a held-out set of CommonRoad scenario IDs that are provably absent from training, and check whether the 0.946 versus 0.902 completion advantage and the TTC and headway gains persist; if the advantage disappears or reverses on unseen scenarios, the claim of generalization from physics-informed learning is falsified. A second, faster check is to inspect the training curves in Fig. 2 for a large gap between training and validation loss on the same scenario distribution, which would indicate memorization.","supporting_citations":[{"cited_title":"Dynamic occupancy grid prediction for urban autonomous driving: A deep learning approach with fully automatic labeling,","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-learning dynamic occupancy grid prediction approach that this work extends by adding physics-informed supervision."},{"cited_title":"Traffic scene prediction via deep learning: Introduction of multi-channel occupancy grid map as a scene representation,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN plus LSTM temporal-occupancy architecture that the proposed network adapts."},{"cited_title":"Mixnet: Physics constrained deep neural motion prediction for autonomous racing,","cited_arxiv_id":null,"evidence_quote":"Motivates the core idea that physics constraints can improve neural motion prediction models."},{"cited_title":"Ofmpnet: Deep end-to-end model for oc- cupancy and flow prediction in urban environment,","cited_arxiv_id":null,"evidence_quote":"A recent end-to-end occupancy and flow prediction model positioned as lacking physical guidance and entity-level modeling."},{"cited_title":"Road traffic safety assessment in self-driving vehicles based on time-to- collision with motion orientation,","cited_arxiv_id":null,"evidence_quote":"Provides the time-to-collision safety metric used in the holistic evaluation."},{"cited_title":"A literature review of performance metrics of automated driving systems for on-road vehicles,","cited_arxiv_id":null,"evidence_quote":"Provides the task completion rate metric used to compare success across scenarios."}],"review_version":1}