{"id":"6edeb02d-a6d0-485c-8e96-484d9def4c83","arxiv_id":"2604.25567","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A feed-forward neural network using Action Dependency Graph features estimates the benefit of replanning in delayed MAPF executions and reduces delay impact by up to 94.6% of the achievable amount.","lead":"The paper trains a neural network to predict when replanning would reduce extra costs in multi-agent path finding executions where some agents are delayed. This helps decide whether to replan or stick with the robust plan in real-world robot fleets.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Neural network trained on simulated delays may fail to generalize to real-world executions, as ADG features and synthetic data may not capture all relevant state dynamics.","rationale":"The reader's weakest assumption matches the core risk: simulation-to-real gap plus possible incompleteness of ADG features. This directly threatens the empirical claim without requiring internal contradictions in the described method. No other load-bearing issues (e.g., in the NN architecture or dataset size) are evident from the given material.","tokens_in":1807,"tokens_out":338,"duration_ms":44462,"concrete_test":"Generate a held-out test set using delay statistics sampled from real robot execution logs (different distribution from the original simulation); run the trained NN to decide replanning points, measure actual makespan reduction versus an oracle that replans at optimal times and versus a no-replan baseline; if the achieved fraction of optimal reduction falls below 60% on this set, the 94.6% claim does not transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim of reducing delay impact by up to 94.6% of the achievable reduction rests on the NN correctly estimating replanning benefit from ADG-based features at runtime. Training and testing occur on a 12,000-experiment dataset of simulated delay scenarios; the abstract gives no indication of real-robot validation, out-of-distribution testing, or ablation on feature sufficiency. If real delays introduce unmodeled correlations (e.g., agent-specific timing jitter or environmental effects) not reflected in the chosen features, the predicted benefit will not match actual cost savings, rendering the percentage reduction unreliable.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces a feed-forward neural network that uses newly designed ADG-based features to estimate, at runtime, the benefit of performing a single replan (or reschedule) during robust MAPF execution when some agents are delayed. The network is trained and evaluated on a new dataset of 12,000 simulated experiments; the authors claim that the method can recover up to 94.6% of the achievable reduction in delay-induced cost while preserving safety.","tokens_in":1932,"tokens_out":385,"duration_ms":34466,"significance":"If the empirical result is reproducible and generalizes, the work would offer a practical, low-overhead mechanism for deciding when replanning is worthwhile in deployed MAPF systems, thereby reducing unnecessary waiting without sacrificing collision avoidance. The release of a sizable labeled dataset is a concrete contribution that could support follow-on learning-based robust execution research.","major_comments":[{"comment":"Abstract: the headline quantitative claim (94.6% of achievable reduction) is reported without any description of how the 'achievable reduction' baseline is computed, what train/test split was used, which baselines were compared, or any measure of statistical significance; these omissions make the central empirical result impossible to assess from the given text.","section":null},{"comment":"Abstract and §4 (dataset and evaluation): the paper motivates the approach for real-life applications yet provides no real-robot validation, out-of-distribution testing, or ablation on whether the chosen ADG features capture all relevant delay dynamics; the generalization assumption is therefore load-bearing for the claimed benefit but unsupported.","section":null}],"minor_comments":[{"comment":"The abstract refers to 'newly designed ADG-based features' without listing their definitions or input dimensionality; explicit formulas or a table would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive review and for recognizing the potential practical value of the work. We address each major comment below, indicating planned revisions where appropriate.","responses":[{"response":"We agree that the abstract's brevity omits details needed to evaluate the central claim. Section 4 of the manuscript defines the achievable reduction as the cost difference between continuing the current ADG-based robust execution under delays and executing an optimal replan computed by a centralized MAPF solver from the observed delayed state. The dataset uses an 80/20 train/test split (with a held-out validation set), comparisons are made to a no-replan baseline, a random replan policy, and a delay-magnitude heuristic, and statistical significance is reported via means and standard deviations over five random seeds. We will revise the abstract to concisely include these elements so the headline result can be assessed from the abstract alone.","revision_made":"yes","referee_comment":"Abstract: the headline quantitative claim (94.6% of achievable reduction) is reported without any description of how the 'achievable reduction' baseline is computed, what train/test split was used, which baselines were compared, or any measure of statistical significance; these omissions make the central empirical result impossible to assess from the given text."},{"response":"We acknowledge that the work contains no real-robot experiments; this is a genuine limitation given the simulation focus and resource constraints of the study, and we will add an explicit limitations paragraph in the discussion section addressing the sim-to-real gap. The evaluation does include out-of-distribution testing on unseen maps and delay patterns (Section 5.3) as well as a feature ablation comparing ADG-based inputs to simpler state representations (Section 4.4). We will expand these sections with additional OOD scenarios and a clearer analysis of which delay dynamics the features capture, thereby strengthening support for the generalization claims.","revision_made":"partial","referee_comment":"Abstract and §4 (dataset and evaluation): the paper motivates the approach for real-life applications yet provides no real-robot validation, out-of-distribution testing, or ablation on whether the chosen ADG features capture all relevant delay dynamics; the generalization assumption is therefore load-bearing for the claimed benefit but unsupported."}],"tokens_in":1430,"tokens_out":497,"duration_ms":113362,"standing_objections":["Real-robot validation experiments"]},"desk_editor":{"model":"grok-4.3","letter":"The paper trains a neural network to estimate whether a single replan will cut execution cost when agents fall behind in an ADG-based robust MAPF schedule. It reports that the model captures up to 94.6 percent of the achievable reduction across 12,000 simulated delay scenarios using a new set of ADG-derived state features as input.","headline":"A feed-forward net on ADG features predicts replanning benefit in delayed MAPF and claims to recover most of the possible gain on simulated runs, but the supporting evidence stays thin.","tokens_in":2508,"tokens_out":151,"would_cite":false,"duration_ms":42569,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A neural network with action dependency graph features predicts when replanning will reduce execution costs for delayed agents in multi-agent path finding.","keywords":["Multi-Agent Path Finding","Robust Execution","Replanning Decision","Action Dependency Graph","Delay Handling","Neural Network Prediction","MAPF Execution"],"falsifier":"Deploy the trained network on a collection of real-robot MAPF executions with measured delays; if the actual cost reduction obtained by replanning at the moments the network recommends falls substantially below the reduction obtained by a policy that always replans or never replans, the prediction model does not generalize as claimed.","tokens_in":2731,"feed_emoji":"🤖","tokens_out":715,"duration_ms":53081,"temperature":0.7,"pith_summary":"In real-world multi-agent path finding, agents frequently encounter delays that break the perfect synchronization assumed in initial plans, forcing robust methods like action dependency graphs to insert waits and raise total execution time. The paper trains a fully connected feed-forward neural network on ADG-derived features that describe the current synchronization state and the projected effect of delays. The network outputs an estimate of how much a single replanning step would lower the remaining execution cost. Experiments on 12,000 labeled simulations show the network can capture up to 94.6 percent of the maximum cost reduction obtainable by replanning while still preserving collision-free execution.","feed_headline":"Neural net flags when replanning cuts delay costs in MAPF","feed_subtitle":"ADG features fed to a feed-forward network predict up to 94.6 percent of achievable savings while agents stay collision-free.","key_machinery":"A fully connected feed-forward neural network that ingests ADG-based features describing the current robust execution state and delay impacts and outputs an estimated cost reduction from replanning.","core_discovery":"The paper establishes that a feed-forward neural network, given a compact set of features extracted from the current action dependency graph and the observed delays, can estimate the execution-cost benefit of performing one replanning operation at any moment during robust MAPF execution, allowing an online system to replan only when the predicted saving is worthwhile.","pith_inferences":["The same ADG-feature approach could be adapted to decide between rescheduling and full replanning without retraining from scratch.","Extending the single-step predictor to a rolling horizon of multiple future replanning opportunities would require only additional output heads on the same network architecture.","The learned predictor might transfer to other graph-scheduled multi-agent systems, such as warehouse order picking or drone delivery, once the dependency graph is defined for those domains."],"forward_implications":["Execution monitors can avoid calling expensive replanners when the network predicts little or no benefit.","Safety guarantees of the original ADG-based robust plan remain intact because replanning is invoked only as an optional, cost-reducing overlay.","The 12,000-experiment dataset supplies a reusable benchmark for training or evaluating other decision procedures that weigh replanning against continued robust execution.","Computational load on the fleet controller drops because replanning is triggered selectively rather than at fixed intervals or after every delay."],"fun_headline_variants":["NN estimates MAPF replan gains from ADG state","Predict MAPF replanning payoff using current delays","Feed-forward net times replans in delayed MAPF","ADG features train net to value MAPF replans"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The features taken from the action dependency graph in simulated delay scenarios will let the trained network accurately forecast replanning benefit in new executions whose delay patterns were never seen during training.","fun_headline_variants_meta":{"raw":{"variants":["NN estimates MAPF replan gains from ADG state","Predict MAPF replanning payoff using current delays","Feed-forward net times replans in delayed MAPF","ADG features train net to value MAPF replans"]},"model":"grok-4.3","cost_usd":0.006568,"raw_usage":{"total_tokens":3023,"prompt_tokens":738,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":65678000,"prompt_tokens_details":{"text_tokens":738,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2223,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":738,"tokens_out":62,"duration_ms":50008,"temperature":1.0,"reasoning_tokens":2223,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-07T13:55:48.636953+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Deploy the trained network on a collection of real-robot MAPF executions with measured delays; if the actual cost reduction obtained by replanning at the moments the network recommends falls substantially below the reduction obtained by a policy that always replans or never replans, the prediction model does not generalize as claimed.","supporting_citations":[],"review_version":1}