{"id":"212e6676-7d6c-47cc-a74e-cc44334ce82d","arxiv_id":"2411.14427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Training transformers to imitate a handcrafted 'expert' heuristic that marks grids on the optimal path lets a constrained A* planner explore fewer nodes on small synthetic maps, while slightly sacrificing path optimality.","lead":"Two transformer models learn to generate heuristics for a risk-constrained A* planner for drone flights, cutting node exploration by up to 53% on small grid maps. A generalist reader might care because it shows a standard deep-learning trick can be glued onto an exact search algorithm, at the cost of some optimality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported search times likely exclude transformer inference; the claimed end-to-end acceleration over Manhattan may not hold.","rationale":"The reader's weakest assumption focuses on the under-specified training labels (the unnamed solver and unreported off-path penalty) and generalization from synthetic maps. That is a real concern. However, I judge the more immediately falsifiable weakness to be the ambiguity about whether the reported search time includes the transformer inference needed to produce the Riskmap2.0 heuristic. The paper's stated purpose is to 'accelerate' planning, and the table of search times is the quantitative support. If those times omit heuristic generation, then the central practical claim is not established, even on the training distribution. The Riskmap-state section shows the authors are aware that generation time can dominate on small problems, but they do not apply this analysis to Riskmap2.0. A simple timing experiment settles the matter. This is a concrete, technical issue that does not require access to the original solver or penalty. It could be addressed by a small code addition in a revised version. Therefore I keep the reader's CONDITIONAL verdict: the paper should be accepted only after the end-to-end timing is reported and shown to support the acceleration claim.","tokens_in":8495,"tokens_out":10630,"duration_ms":102430,"concrete_test":"Implement a timing harness that, for each of the 1000 test tasks per map size, measures total wall-clock time from receiving the risk map and start/goal to returning the path, using the same code path and hardware. Run this for (a) Riskmap2.0 heuristic (including its forward pass) and (b) Manhattan heuristic (no generation). If (a) is not faster than (b) in geometric mean or median across the 1000 tasks, the headline acceleration is unsupported and the paper should be revised to report end-to-end timings and hardware details.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the learned heuristic accelerates ASD A*. Table I reports 'Search time (ms)' for Riskmap2.0 and Manhattan, but the paper never states whether this is A*-only time or end-to-end time. Riskmap2.0 must run a transformer encoder over the full flattened map (e.g., a 1024-token sequence for 32x32) before A* begins; that forward pass has real computational cost, especially on CPU. In the Riskmap-state experiments the authors explicitly warn that 'the heuristic generating time may be longer than the search time, because the Riskmap-state must generate a heuristic for every node A* found.' The same caveat is never analyzed for Riskmap2.0, where the heuristic is generated once per task. On the 32x32 maps the reported saving is only 2.64 ms (27.19 ms vs 24.55 ms); a single transformer inference on a 1024-token sequence could easily exceed that on commodity hardware. If the reported 'search time' excludes the forward pass, then the paper's acceleration claim is not an end-to-end claim, and the method could be slower than the Manhattan baseline in practice. This is a load-bearing ambiguity because the motivating purpose of the method is acceleration, and it cannot be resolved from the text.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two transformer-based heuristic generators for the ASD A* algorithm, which solves a risk-constrained shortest path (CSP) problem for Advanced Air Mobility planning. The first generator, Riskmap2.0, takes a risk map, start, and destination and outputs a heuristic for every grid; the second, Riskmap-state, takes the current node, destination, and risk map and outputs a heuristic for that node. The heuristics are trained in a supervised manner on labels produced by an unspecified traditional solver. Experiments on 16x16, 24x24, and 32x32 random maps and on a 16x16 city map report that Riskmap2.0 reduces node exploration and search time relative to a Manhattan baseline while maintaining SPL above 97.55%, and that Riskmap-state produces heuristics close to expert labels with smaller node counts. The central claim is that the learned heuristics accelerate ASD A* while preserving near-optimal path length.","tokens_in":8709,"tokens_out":3696,"duration_ms":38983,"significance":"If the reported results hold, the paper would demonstrate a practical way to inject learned, risk-aware guidance into a constrained A* planner, with a useful cross-distribution test on a city wind-flow map. Strengths include the large dataset sizes (over 128,000 entries per map size), the two distinct heuristic-generation strategies, and the explicit use of a classification-style output for Riskmap2.0 that keeps heuristics within a bounded value set. However, the central acceleration claim is currently not fully supported because the paper does not clarify whether the reported search times include transformer inference, does not disclose the expert-label generation procedure (solver and penalty value), and provides no variance or statistical tests for any of the reported averages. The comparison against Manhattan alone also leaves open how much of the speedup comes from the oracle-like construction of the training labels rather than from learned generalization. These issues are load-bearing for the paper's motivating claim of computational acceleration, but they are addressable with additional experiments and reporting.","major_comments":[{"comment":"The paper reports 'Search time (ms)' for Riskmap2.0 and Manhattan, but it never states whether this is A*-only node-expansion time or end-to-end time including heuristic generation. Riskmap2.0 requires a transformer forward pass over the entire flattened map before A* begins; for a 32x32 map this is a 1024-token sequence, and the reported saving over Manhattan is only 2.64 ms (27.19 vs 24.55 ms). A single transformer inference can plausibly exceed that saving on commodity hardware, so if the reported times exclude inference, the abstract's 'accelerate' claim is not supported end-to-end. The authors must specify the timing protocol and report wall-clock time including all heuristic-generation costs for both the learned and baseline heuristics.","section":"V.A, Table I"},{"comment":"The expert heuristic for Riskmap2.0 is defined using an unnamed 'traditional solver' and an undisclosed penalty added to the Manhattan distance for grids not on the shortest path. This makes the dataset irreproducible and leaves a free parameter that directly controls how aggressively ASD A* is steered toward the oracle path. The manuscript also asserts that this heuristic 'maintains consistency' without proof; if the penalty is positive, the heuristic can violate consistency because an off-path node can have h(n) = Manhattan(n) + P while a neighboring on-path node has h(n') = Manhattan(n) - 1, giving h(n) > 1 + h(n'). The authors must report the solver, the penalty value, and either prove or check consistency and admissibility, since these properties affect both optimality (SPL) and node counts.","section":"III.B.1"},{"comment":"All reported improvements are averages over 1000 tasks, but no standard deviations, confidence intervals, or statistical tests are given. Several differences in Table I are small in absolute terms, such as the 32x32 node count (826.14 vs 754.19) and search time (27.19 vs 24.55 ms), and without variance information it is impossible to tell whether these differences are robust across seeds, tie-breaking rules, or task draws. The paper should include error bars and, where relevant, paired tests over the same task set.","section":"V.A, V.C"},{"comment":"For Riskmap-state the authors acknowledge that 'the heuristic generating time may be longer than the search time, because the Riskmap-state must generate a heuristic for every node A* found.' The same caveat applies to Riskmap2.0, where the heuristic is generated once per task, but no attempt is made to measure this overhead. Since the stated purpose of the method is acceleration, the absence of any end-to-end timing comparison for either model leaves the core practical claim unverified. A concrete test would be to report total planning time including all neural network forward passes and compare it with the Manhattan heuristic on the same hardware.","section":"III.B.2, V.B"}],"minor_comments":[{"comment":"The label 'Manhanton' is a typo and should read 'Manhattan'.","section":"Table I"},{"comment":"The phrase 'as less nodes as possible' should be 'as few nodes as possible' for grammatical correctness.","section":"Introduction"},{"comment":"The sentence describing the 16x16 Riskmap-state result is duplicated: 'For the 16*16 dataset, the number of nodes explored is 19.23% less than the Manhattan. for the 16*16 dataset, the number of nodes explored is 10.51% less than the Manhattan.' The second sentence likely refers to a different map size or condition and should be corrected.","section":"V.B"},{"comment":"The text says 'Fig 2 shows an example of the ADS A*' but the algorithm is consistently called ASD A* elsewhere; please fix the acronym.","section":"II, Figure 2"},{"comment":"The phrase 'while maintaining consistency' is a substantive algorithmic claim, not a formatting issue; if the authors keep it, they should provide a proof or a citation. If they remove it, the sentence remains complete without it.","section":"III.B.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a direct continuation of the authors' previous conference paper, and the novelty over that prior work appears incremental. More importantly, the reproducibility of the empirical claims depends on an undisclosed solver and penalty value, and the timing protocol is not defined. These are fixable but central issues. The absence of any code or data release also makes independent verification difficult. I would encourage the editor to require the authors to provide the missing experimental details and an end-to-end timing comparison before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a straight continuation of the authors' AIAA 2024 work ([26]): same ASD A* algorithm, same transformer-based heuristic idea, now with a second network (Riskmap-state), a bigger dataset, and a single 16x16 city-map evaluation. That is the extent of the novelty. What is genuinely new is the Riskmap-state regression network, which outputs a heuristic for the current node rather than the whole map, and the city-map test showing 53% fewer nodes explored than Manhattan. Those results are worth a look if you work on learned heuristics for constrained A*.\n\nThe paper does several things well. It reports consistent node-exploration reductions over Manhattan across 16x16, 24x24, and 32x32 grids, and it is candid about the Riskmap-state limitation: generating a heuristic per visited node can take longer than the search itself. That honesty is a credit.\n\nThe soft spots are real and one is load-bearing. The stress-test concern about timing is correct: Table I reports \"Search time (ms)\" but never states whether it includes the transformer forward pass. Riskmap2.0 must encode the full flattened map before A* starts; on 32x32 that is a 1024-token transformer inference. The reported saving over Manhattan is only 2.64 ms, and that could easily be swallowed by a single forward pass on commodity hardware. The paper applies the \"heuristic generating time may be longer\" caveat to Riskmap-state but never to Riskmap2.0, where it matters just as much. This must be fixed for the acceleration claim to stand.\n\nBeyond that, the expert heuristic generation is under-specified: no name for the \"traditional solver,\" no penalty value for off-path grids, no error bars or statistical tests, and no code release. The comparison is only against Manhattan, not against the authors' own prior model, and the city-map test is limited to 16x16. The circularity concern is partly inherent to learned heuristics trained on solver output, but the expert heuristic is deliberately constructed so that ASD A* explores only shortest-path nodes; that makes the gains over Manhattan a measure of how well the transformer approximates the oracle, which is fine if framed as an engineering result. The conclusion overreaches by claiming to \"100% solve the problem\" and asserting less training time without evidence.\n\nWho gets value: anyone working on learning-augmented search or AAM planning who wants a concrete data point and a baseline for their own comparisons. It deserves peer review because the empirical claim is potentially useful and the issues are fixable with better measurement and reporting. As it stands, I would not cite it yet, but I would engage with a revised version that addresses the timing ambiguity and releases the oracle details.","headline":"An incremental but honest extension of the authors' own learned-heuristic work; the main acceleration claim hinges on an unresolved timing ambiguity and an under-specified oracle.","tokens_in":9237,"tokens_out":1842,"would_cite":false,"duration_ms":18980,"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 transformer trained on solver-generated heuristics can speed safety-constrained A* search for drone planning by up to 39.5% while keeping paths near-optimal.","keywords":["Constrained Shortest Path","A* algorithm","Advanced Air Mobility Planning","Supervised Learning","Transformer","heuristic search","risk-aware path planning","unmanned aerial vehicles"],"falsifier":"Run ASD A* with the Riskmap2.0 heuristic on maps whose risk distribution differs from the equal-thirds training distribution (for example, mostly safe cells with a few high-risk corridors) and compare node expansions against the Manhattan heuristic; if the learned heuristic does not reduce expansions or SPL drops materially below the Manhattan baseline on any such map family, the generalization claim fails.","tokens_in":8254,"feed_emoji":"🚁","tokens_out":8682,"duration_ms":77816,"temperature":0.7,"pith_summary":"This paper tries to establish that a transformer can learn the search heuristic that makes A*-style planners fast on risk-constrained drone routing. The problem is hard in the worst case: constrained shortest path (CSP) is NP-hard, and exact solvers are slow. The authors feed their ASD A* algorithm—A* augmented with a safety dimension—heuristics generated by two transformer networks, one that scores every grid cell given start and goal, and one that scores only the current search node. On random maps up to 32×32, the first network's heuristic beats the Manhattan baseline by exploring 8.71–39.5% fewer nodes and searching 9.71–24% faster while keeping success-weighted path length (SPL) above 97.55%. If true, this means learning can supply the guidance that makes constrained planning practical for urban air mobility without surrendering the planner's validity guarantee.","feed_headline":"Transformer-learned heuristic cuts drone A* searches by 39.5%","feed_subtitle":"Learned guidance for safety-constrained drone search cuts explored nodes by up to 39.5% without losing path quality.","key_machinery":"The load-bearing mechanism is the Transformer encoder used as a heuristic generator. For Riskmap2.0, the flattened risk map and tokenized start/destination are embedded and combined, then processed by the encoder; the output layer classifies each grid's heuristic into a discrete value set and takes an argmax, which keeps the generated heuristics within a bounded range. For Riskmap-state, the risk map is padded to a fixed size, combined with sinusoidal positional embeddings, concatenated with a task embedding for the current node and destination, and processed by the same encoder, with an MSE loss on the single scalar output. The training labels are constructed so that grids on the solver's shortest path receive lower heuristic values (Manhattan distance), while off-path grids receive a penalty; an accurate learned version of that pattern is what lets ASD A* concentrate its expansion on the shortest path.","core_discovery":"The central claim is that a transformer trained on solver-generated heuristic labels can produce, at inference time, a heuristic that guides ASD A* to near-shortest feasible paths while expanding far fewer nodes than the Manhattan-distance baseline. The Riskmap2.0 network takes a risk map, start, and destination and outputs a heuristic value for every grid, trained with a classification loss over a bounded set of heuristic values; the Riskmap-state network takes the risk map, the current search node, and the destination and regresses a single heuristic. The reported results put Riskmap2.0's node-exploration reduction at 8.71–39.5% across map sizes with SPL at 97.55–99.58%, and on a 16×16 city wind-flow map the reduction reaches 53.53% with SPL 99.73%. The paper presents these numbers as evidence that learned heuristics can accelerate safety-constrained air mobility planning while preserving near-optimal path quality.","pith_inferences":["The reported advantage shrinks as map size grows (39.5% at 16×16, 16.75% at 24×24, 8.71% at 32×32); the authors attribute this to reduced training data per map, which suggests the larger-map gap could close with more labels or longer training rather than a fundamental limit of the heuristic.","The method's ceiling is set by the undisclosed solver that generates the labels; if that solver is approximate, the learned heuristic inherits its errors, and an exact constrained-shortest-path solver would provide a stronger oracle and an upper bound on achievable speedup.","The only out-of-distribution test is a 16×16 city map, so claims of adaptability to different map sizes and risk structures would be much stronger if repeated at 64×64 or on real wind-flow maps without downsampling."],"forward_implications":["On random 16×16, 24×24, and 32×32 risk maps, ASD A* with the Riskmap2.0 heuristic explores 8.71–39.5% fewer nodes and searches 9.71–24% faster than with Manhattan distance, while SPL stays between 97.55% and 99.58%.","On a structured 16×16 city wind-flow map, the same heuristic explores 53.53% fewer nodes and searches 48.65% faster, with SPL 99.73%, showing that the learned heuristic can transfer to at least one non-random risk distribution.","The Riskmap-state network, trained jointly on 16×16 and 64×64 maps, produces per-node heuristics with MSE 1.742 on 16×16 and 6.234 on 64×64, reducing node exploration by 19.23% and 10.51%, respectively.","Because the transformer only proposes heuristics and ASD A* remains the planner, every returned path is still checked against the safety constraint—unlike end-to-end transformer planners that can fail to produce valid plans."],"supporting_citations":[{"why":"Supplies the Transformer encoder backbone used by both heuristic networks.","marker":"[19]"},{"why":"Defines the ASD A* algorithm and the original learned-heuristic approach this paper extends.","marker":"[26]"},{"why":"Provides the SPL metric used to report path optimality.","marker":"[27]"},{"why":"Generates the city wind-flow risk maps used for the realistic transfer test.","marker":"[28]"},{"why":"Serves as the comparison transformer planner that the paper argues cannot guarantee valid solutions.","marker":"[24]"}],"fun_headline_variants":["AI heuristic speeds safety-constrained drone pathfinding by up to 39.5%","Learned heuristic accelerates AAM path planning, preserves quality","Transformer heuristic shrinks search space for safe drone routes","Risk-aware planning gets faster with transformer-based heuristics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speedup claim depends on the unreported traditional solver's labels and the hand-chosen Manhattan-distance penalty being a reliable oracle for the planner, and on the transformer generalizing from synthetic maps with equal thirds of safe, low-risk, and high-risk grids to the deployment environment.","fun_headline_variants_meta":{"raw":{"variants":["AI heuristic speeds safety-constrained drone pathfinding by up to 39.5%","Learned heuristic accelerates AAM path planning, preserves quality","Transformer heuristic shrinks search space for safe drone routes","Risk-aware planning gets faster with transformer-based heuristics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3248,"prompt_tokens":928,"completion_tokens":2320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2249}},"tokens_in":544,"tokens_out":2320,"duration_ms":15109,"temperature":1.0,"reasoning_tokens":2249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:11:12.660301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ASD A* with the Riskmap2.0 heuristic on maps whose risk distribution differs from the equal-thirds training distribution (for example, mostly safe cells with a few high-risk corridors) and compare node expansions against the Manhattan heuristic; if the learned heuristic does not reduce expansions or SPL drops materially below the Manhattan baseline on any such map family, the generalization claim fails.","supporting_citations":[{"cited_title":"Learning-accelerated a* search for risk-aware path planning","cited_arxiv_id":null,"evidence_quote":"Defines the ASD A* algorithm and the original learned-heuristic approach this paper extends."},{"cited_title":"Safe path planning of uav based on reinforcement learning in probabilistic environments","cited_arxiv_id":null,"evidence_quote":"Generates the city wind-flow risk maps used for the realistic transfer test."}],"review_version":1}