{"id":"6b1e747b-4b0d-4110-a98a-7d00fee6cfba","arxiv_id":"2607.22166","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prior-guided planning framework distills expert trajectories into local spatiotemporal decision priors and injects them into A* and ACO planners, cutting explored nodes under partial observability.","lead":"ImiPath trains a spatiotemporal attention network on expert navigation trajectories and uses its predictions to bias A* and ant-colony planners inside a local 11x11 observation window, reducing node expansions while keeping path quality. The value for a generalist is a concrete test of whether learned local priors can replace global heuristics for robots that can only see nearby obstacles.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'heterogeneous planners' claim is untested: the prior is only plugged into the same PFACO that generated its training data, so the observed stochastic gains may be planner-specific rather than reusable decision priors.","rationale":"Reader's weakest_assumption (11x11 observation sufficiency) is a valid concern, but I find a more directly load-bearing gap in the evaluation logic: the paper's central claim is about reusable, planner-agnostic priors, yet the stochastic experiments are self-referential (train on PFACO, test with PFACO), and no cross-planner stochastic test is reported. This is not a matter of disagreement with consensus; it is a mismatch between the claim's scope and the evidence. The deterministic results do show a clear node-expansion reduction, so the paper supports a weaker claim: a learned prior trained by an expert can improve the same or closely related planner. The reader's rationale already notes the self-referential PFACO issue and the missing ACO-PAL baseline, so my concern is complementary rather than contradictory. A focused cross-planner experiment would settle it. I would keep the CONDITIONAL verdict: the paper is plausible and internally coherent, but its strongest advertised application (heterogeneous planners) is not yet supported.","tokens_in":15503,"tokens_out":8430,"duration_ms":90095,"concrete_test":"Using the same trained STAPNet checkpoint and the same held-out maps, integrate the prior into two stochastic planners not used for training (e.g., AS and EAS) via the same transition fusion as Eq. (19), and into a deterministic planner outside the A* family (e.g., Dijkstra or JPS) via Eq. (18). Measure APL, time, SR, and explored nodes. If the prior-guided AS/EAS do not beat their unguided counterparts as clearly as PFACO+prior beats PFACO, the heterogeneous-transfer claim should be narrowed to same-planner/same-family guidance. Also add ACO-PAL [18] as a baseline under identical FoV conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ImiPath's central novelty is that STAPNet learns planner-agnostic spatiotemporal decision priors that can be reused across heterogeneous planners (§3.5.2). The experiments do not establish this. STAPNet is trained on expert demonstrations from PFACO and A* Global (§4.1). The stochastic variant is evaluated only as ImiPathPFACO (Tables 1–2) — the same algorithm whose trajectories were used for training. The deterministic variant is combined with A* Local, which is the same A* family as the A* Global expert. No result combines the trained prior with a stochastic planner not used for training (AS, EAS, MMAS, IHMACO appear only as baselines), and the closest prior-aware local planner ACO-PAL [18] is not compared. Therefore the improvement of ImiPathPFACO over PFACO could be a specialized correction to PFACO's own transition rule (Eq. 19) rather than evidence for a reusable prior. The abstract's claim that the prior generalizes to heterogeneous planners and §3.5.2's claim of planner-agnostic guidance are extrapolations beyond the reported data. This is an experimental-scope gap, not an internal inconsistency; the deterministic node-expansion results stand, but they support only same-family transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ImiPath, an imitation-learning framework for path planning under partial observability. STAPNet is trained on expert demonstrations generated by PFACO and A* Global, using an 11×11 spatiotemporal observation that encodes obstacle layout, projected start/goal directions, and temporally decayed trajectory history. At inference, the predicted action distribution is fused into downstream planners: as a bias term in A*-style evaluation (Eq. 18) and as a multiplicative factor in ACO transition probabilities (Eq. 19). Experiments compare stochastic variants (ImiPath_PFACO) against ACO baselines and deterministic variants (ImiPath_A*_Local) against A* Local on maps of several sizes, with additional dynamic-scenario and magnetic-microrobot demonstrations. The main reported results are that the deterministic variant reduces explored nodes on large maps while preserving 100% success rate, and that the stochastic variant achieves competitive or better path quality than PFACO.","tokens_in":15821,"tokens_out":5696,"duration_ms":59050,"significance":"If the deterministic node-reduction results replicate, the paper makes a useful empirical contribution: a learned local prior can partially substitute for global heuristics in partially observable grid navigation. The use of Wilcoxon signed-rank tests, 100% success rates across map scales, and a physical robot validation are strengths. However, the central novelty claim — that ImiPath learns planner-agnostic priors reusable across heterogeneous planning paradigms — is not established by the reported experiments. The stochastic variant is only fused into the same PFACO algorithm that generated its training demonstrations, and the deterministic variant is fused into A* Local, which belongs to the same A* family as the A* Global expert. The closest prior-aware baseline, ACO-PAL [18], is not compared. These gaps are load-bearing for the generality claim.","major_comments":[{"comment":"The central claim of reusable priors for heterogeneous planners is not tested. STAPNet is trained on PFACO and A* Global demonstrations (§4.1); the stochastic evaluation fuses the prior only into PFACO (Tables 1–2), and the deterministic evaluation only into A* Local (Table 3), which is the same A* family as the A* Global expert. No stochastic planner not used in training (AS, EAS, MMAS, IHMACO) is tested as a fusion target, and the closest prior-aware local planner ACO-PAL [18] is absent. The observed gains could therefore be a same-algorithm correction rather than evidence of transfer. Please add cross-planner fusion experiments with at least one held-out stochastic planner and compare against ACO-PAL, or substantially narrow the claimed scope.","section":"§4.2.1, Tables 1–2 vs §3.5.2"},{"comment":"Search efficiency is measured primarily by explored-node counts, but wall-clock time is substantially higher for ImiPath_A*_Local than for A* Local on every map scale (e.g., 7.30e-1 s vs 1.839e-2 s on Large). For the stochastic variants, no node-expansion counts are reported, and the time comparisons are mixed (e.g., FoV scale 3.344e-1 vs 3.132e-1; Small 8.160e-1 vs 6.407e-1; Medium-Large 1.426 vs 1.410). Since the title and abstract claim 'efficient path planning', the paper should either report node counts for stochastic planners and time for all variants, or explicitly state that node reduction does not imply wall-clock speedup. Without this, the efficiency claim is not fully supported.","section":"Table 3 and §4.2.1"},{"comment":"The load-bearing premise is that the 11×11 local observation (Eqs. 8–12) — obstacles, projected start/goal directions, and trajectory history — suffices to recover globally informed expert preferences. The ablations in §4.2.3 vary dataset size, network architecture, and fusion strategy, but never remove or replace an observation channel (e.g., M_T, M_S, or M_G). A channel-ablation study and a sensitivity analysis for the temporal decay γ (Eq. 11) and the fusion weights ρ, μ (Eq. 18) are needed to show that the prior exploits spatiotemporal structure rather than acting as a simple goal-direction bias.","section":"§3.1.2 and §4.2.3"}],"minor_comments":[{"comment":"The word 'dicision' appears several times (e.g., abstract, §3, §3.5.2); it should be 'decision'.","section":"Abstract and throughout"},{"comment":"The sentence beginning 'On FoV scale, ImiPathA*Local achieves a 100% success rate...' is duplicated verbatim. Please remove the repetition.","section":"§4.2.1"},{"comment":"STAPNet standalone runtime is inconsistent: Table 1 reports 3.361e-3 s while Table 6 reports 3.365e-1 s for essentially the same entry (APL and SR match). Please clarify which is correct and ensure consistent reporting.","section":"Table 6 vs Table 1"},{"comment":"In the deterministic-planner review, 'Xu et al. [24]' and 'Huang et al. [24]' both cite the same reference [24], but the bibliography lists only Xu et al. (2024). The 'Huang et al.' citation appears to be a reference error.","section":"§2.1"},{"comment":"Training details are incomplete: learning rate, weight decay, batch size, epochs, gamma, alpha, beta, and the specific values of rho/mu for the main experiments are not collected in one place. The paper would benefit from a hyperparameter table. Additionally, no code or trained model is provided, which limits reproducibility.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid deterministic core — the node-reduction results with 100% SR and Wilcoxon tests are credible — but the 'heterogeneous planners' claim, which is the main novelty, goes beyond the evidence. I would request either new cross-planner experiments (including ACO-PAL) or a careful re-scoping of the claims. This is fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about learned guidance for local path planning. What's actually new is the packaging: a local spatiotemporal observation (obstacles, projected start/goal directions, decayed trajectory history) plus a cross-attention policy network whose output is injected into both A* and ACO transition rules. Each ingredient has been published before, but the specific combination for partial observability is a legitimate extension, not a paradigm shift.\n\nThe deterministic experiments are the strong part. On large maps, ImiPath_A*_Local expands about 90 nodes versus 242 for A* Local, with 100% success and comparable or better path length. That is a real, repeatable-looking result and it gives the paper independent grounding: the prior is trained on A* Global demonstrations and transferred to A* Local, so this is genuine cross-instance, within-family transfer. The microrobot demo is a nice bonus, though preliminary.\n\nThe soft spots are exactly where the reader and stress-test point. The claim that the prior is planner-agnostic and reusable across heterogeneous planners is not established. STAPNet is trained on PFACO and A* Global; the stochastic variant is only evaluated as ImiPath_PFACO, which is the same planner that produced its training data. That means the stochastic gains could simply be a specialized correction to PFACO's own transition rule. No result plugs the prior into AS, EAS, MMAS, or any other stochastic planner, and the closest prior-aware baseline ACO-PAL is absent. The deterministic variant uses A* Local, which is the same search family as the A* Global expert, so it only supports same-family transfer. The paper's abstract and §3.5.2 overstate what the data show.\n\nAlso, the stochastic experiments report no node-expansion counts, and wall-clock time for the full ImiPath variants is often worse than the lightweight baselines. No code or data is provided, and key hyperparameters (gamma, rho, mu, alpha, beta) are not fully specified. These are addressable, not fatal.\n\nBottom line: the deterministic result deserves a referee, and the framework is coherent and clearly presented. But the central generalization claim needs either more experiments or a more modest wording. I would send it to peer review with the expectation of major revision—add ACO-PAL and at least one non-PFACO stochastic host, report stochastic node expansions, and release code/data. Then the paper would matching its claims.","headline":"A useful local-prior planner with a solid deterministic result, but the 'heterogeneous planners' claim is not supported by the experiments as written.","tokens_in":16358,"tokens_out":1618,"would_cite":true,"duration_ms":19151,"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 local 11x11 view can supply the directional priors that global heuristics provide, making partially observable planning faster without sacrificing path quality.","keywords":["path planning","partial observability","imitation learning","spatiotemporal priors","ant colony optimization","A* search","grid navigation","decision priors"],"falsifier":"Train STAPNet on maps whose local obstacle statistics differ sharply from the test maps (e.g., open rooms in training, narrow corridors in testing), then compare ImiPath-guided A* against plain A* Local on the test corridors: if node expansions and success rate do not improve over A* Local, the prior has overfit to training layout statistics rather than learned transferable local decision knowledge.","tokens_in":15386,"feed_emoji":"🧭","tokens_out":7295,"duration_ms":73059,"temperature":0.7,"pith_summary":"The paper claims that an agent navigating with only an 11x11 local view can recover the directional preferences a global planner would use, by learning them from expert trajectories. ImiPath encodes the local obstacle layout, projected start and goal directions, and recency-weighted motion history into a compact observation, and trains STAPNet to predict expert-like action distributions. Those distributions are injected into classical planners — A* and ant-colony optimization — as directional guidance instead of being used as a standalone controller. If the claim holds, partially observable grid navigation can match near-global path quality while expanding far fewer nodes, across map sizes, dynamic scenes, and a physical magnetic microrobot.","feed_headline":"90 nodes instead of 243: local priors speed partial-view planning","feed_subtitle":"The same trained prior guides both grid-search and ant-colony planners without retraining.","key_machinery":"The key machinery is the local spatiotemporal observation — four aligned 11x11 matrices encoding obstacle occupancy, the agent's start and goal as projected boundary cues, and recent trajectory history with exponential temporal decay — together with STAPNet, a cross-attention policy network whose temporal-history branch queries the spatial branch and outputs a distribution over eight neighbor actions. The prior enters planners through two fusion rules: a logarithmic probability term subtracted in A*'s node evaluation, and a multiplicative probability factor in the ant-colony transition probability. Together they convert a purely local view into an expert-informed directional bias that reduce","core_discovery":"The paper's central claim is that a policy network trained on 11x11 local windows extracted from globally optimal expert trajectories can serve as a reusable spatiotemporal decision prior for partially observable planning. The learned policy outputs a distribution over eight movement directions, and ImiPath fuses that distribution into the search: for A*, as a log-probability bonus in the node evaluation function; for ant-colony optimization, as a multiplicative factor in the state-transition rule. Under an 11x11 field of view, this reduces node expansions substantially on maps larger than the view (e.g., about 90 versus 243 explored nodes on 30x30 maps) while maintaining 100% success and eq","pith_inferences":["Beyond the paper: the same observation encoder could bias sampling in continuous planners such as RRT-style algorithms by turning the predicted action distribution into a non-uniform sampling density, since the prior is defined on directions rather than grid cells.","Beyond the paper: if the trajectory-history channel is what suppresses re-exploration, the idea could extend to multi-robot coordination, where each agent's recent history biases its search away from regions teammates just covered.","Beyond the paper: sweeping the field-of-view size at test time would separate true locality from goal-direction leakage — if gains persist at 7x7 or 9x9 views, the prior is genuinely local; if they require the full 11x11, part of the signal is the projected goal cue.","Beyond the paper: replacing the grid observation with a local cost patch plus bearing-to-goal would test whether the mechanism transfers to non-grid costmaps; the paper does not report such an experiment."],"forward_implications":["ImiPath-guided A* keeps 100% success on maps from 15x15 to 30x30 under an 11x11 field of view, with the largest map showing about 90 explored nodes versus 243 for local A* and a shorter average path.","The same trained prior, fused into ant-colony optimization as a transition bias, reaches 100% success on every tested map scale with path quality that matches or beats the best stochastic baselines.","Prior inference takes milliseconds, making online closed-loop replanning feasible in dynamic environments; the dynamic-scenario tests show 100% success with at most 15 explored nodes.","Because the prior comes from local observations rather than planner-specific data structures, the same network can guide both deterministic and stochastic planners without retraining or parameter reselection.","The framework transfers from simulation to a physical magnetic microrobot, completing navigation from sequential local observations in a constrained channel environment."],"fun_headline_variants":["Learned priors cut path-planning node expansions by over 60%","ImiPath: one prior guides grid and ant colony planners under partial views","Spatiotemporal priors slash node expansions in partially observable maps","Reusable decision priors trim planning nodes from 243 to 90","From 243 to 90 nodes: local priors for efficient partial-view planning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that an 11x11 window showing obstacles, projected start and goal directions, and recency-weighted trajectory history is rich enough for a policy trained on globally informed expert paths to produce direction preferences that transfer to unseen maps and dynamic scenarios; if that local view is too impoverished, the learned prior collapses into a simple goal-direction bias and the efficiency gains disappear.","fun_headline_variants_meta":{"raw":{"variants":["Learned priors cut path-planning node expansions by over 60%","ImiPath: one prior guides grid and ant colony planners under partial views","Spatiotemporal priors slash node expansions in partially observable maps","Reusable decision priors trim planning nodes from 243 to 90","From 243 to 90 nodes: local priors for efficient partial-view planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1824,"prompt_tokens":749,"completion_tokens":1075,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":976}},"tokens_in":493,"tokens_out":1075,"duration_ms":10215,"temperature":1.0,"reasoning_tokens":976,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:34:16.514239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train STAPNet on maps whose local obstacle statistics differ sharply from the test maps (e.g., open rooms in training, narrow corridors in testing), then compare ImiPath-guided A* against plain A* Local on the test corridors: if node expansions and success rate do not improve over A* Local, the prior has overfit to training layout statistics rather than learned transferable local decision knowledge.","supporting_citations":[],"review_version":1}