acceptodds
Under review as a conference paper at ICLR 2027

Learning Highly Accurate Conservative Strategy Based on Graph Neural Network for the Feasibility Prediction of Multi-UAV Task Planning

Abstract

High-throughput multi-UAV task-planning systems must evaluate diverse task instances under limited computational budgets. However, calling a search-based planner for every candidate instance is costly, whereas rejecting a planner-solvable instance is operationally more consequential. Therefore, we study pre-planner screening for a fixed planning pipeline with three budget-dependent outcomes: Feasible, Infeasible, and Unresolved. Since retaining Unresolved as a distinct outcome prevents budget-limited outcomes from being conflated with infeasibility, we propose CONS-GNN, which represents each candidate instance as a relation-aware heterogeneous UAV–task–region graph and predicts the reference planner’s outcome using only pre-planner information. CONS-GNN rejects an instance only when its predicted probability of infeasibility exceeds a validation-selected threshold; all other instances are forwarded to the planner. We evaluate CONS-GNN on three independently seeded test sets, each containing 10,000 3D multi-UAV task instances. CONS-GNN achieves mean accuracy, macro-F1, and macro-averaged one-vs-rest PR-AUC scores of 0.9517, 0.9388, and 0.9831, respectively. At the selected operating points, it avoids an average of 41.36% of planner calls while rejecting 3 of 12,449 planner-feasible test instances, corresponding to an observed deployment FRR of 0.0241%. Ablation and scalability analyses identify the contributions of individual graph components and assess performance across UAV team sizes. Together, these results indicate that planner-conditioned screening can reduce planner use with few observed false rejections under the evaluated conditions.

Primary area:
reinforcement learning
Keywords:
Multiagent Learning, Graph-based Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1eArOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CONS-GNN for planner feasibility prediction in multi-UAV task planning. The formulation of conservative screening based on three planner outcomes is reasonable, and the experimental evaluation includes multiple GNN baselines, ablation studies, and scalability analyses across different UAV team sizes. However, the proposed method primarily combines existing techniques, including heterogeneous GNNs, auxiliary multi-task supervision, and threshold-based selective classification, resulting in limited algorithmic novelty.

Critical strengths
  • Paper is generally well written and clearly organized, with the methodology and experimental setup presented in an understandable manner.

  • It achieves promising experimental results and outperforms existing methods on several evaluation metrics, although some concerns remain regarding the interpretation and validity of these results.

Critical weaknesses
  1. The core components of CONS-GNN include three parts: a heterogeneous UAV–task–region graph, relation-aware edge-conditioned message passing with an auxiliary difficulty prediction objective, and a validation-selected threshold for conservative screening. However, these components are largely based on established techniques. Relation-aware message passing, heterogeneous graph representations, and auxiliary multi-task learning are well-studied approaches in graph learning, while validation-based rejection thresholds are standard in selective classification. Section 4 mainly describes how these components are integrated, without introducing a clearly distinguishable new learning mechanism.

    Although Table 2 shows that CONS-GNN outperforms GAT, MPNN, R-GCN, HGT, and Adapted GRN, the improvements over stronger baselines are relatively modest. For example, CONS-GNN achieves an accuracy of 0.9517 compared with 0.9462 for Adapted GRN, while its PR-AUC of 0.9831 is only slightly higher than the approximately 0.981 achieved by several existing GNN models. Its planner-call savings exceed those of Adapted GRN by 3.71 percentage points.

    The main contribution appears to be an effective task-specific heterogeneous graph representation and selective screening pipeline, rather than a substantial methodological advance in GNN learning. The authors should clarify which aspects of the proposed algorithm cannot be readily reproduced using existing relation-aware GNN architectures and multi-task classification frameworks.

  2. One of the paper's central claims is that explicitly distinguishing Unresolved from Infeasible substantially reduces false rejections compared with binary outcome modeling. Table 4 reports that three-way CONS-GNN achieves 41.36% planner-call savings with an observed false-rejection rate (FRR) of 0.0241%, whereas the binary-collapse variant achieves 52.96% savings with an FRR of 0.2414%. Although the three-way model has an approximately tenfold lower observed FRR, the two models operate at substantially different rejection levels: the binary variant rejects 11.60 percentage points more instances.

    This comparison does not establish that three-way modeling provides a better safety–efficiency trade-off. A more conservative operating threshold generally reduces false rejections at the cost of rejecting fewer instances. The reported difference may therefore partly reflect the selected operating points rather than an inherent advantage of the three-class formulation.

    Furthermore, both models achieve observed test FRRs below the validation constraint of 0.5%. As the authors acknowledge in Section 5.4, these results do not establish superiority at matched test risk.

    A more convincing evaluation would compare planner-call savings at the same FRR levels or provide complete FRR–savings curves for both formulations. Without such comparisons, the current experiments demonstrate different operating trade-offs but do not conclusively support the claimed advantage of three-way outcome modeling.

  3. This paper reports that CONS-GNN avoids 41.36% of planner calls while falsely rejecting only 3 of 12,449 planner-feasible instances across three test sets, corresponding to an observed FRR of 0.0241%. This is a promising result under the evaluated conditions.

    However, this low false-rejection rate is achieved through empirical threshold selection on the validation set. Specifically, the authors jointly select the model checkpoint and rejection threshold to maximize planner-call savings while satisfying a validation FRR constraint of 0.5%. Such a procedure does not guarantee that the same risk level will be maintained on unseen or shifted task distributions.

    And all primary labels are generated using a fixed EECBS–3D-SIPP planning pipeline with a restrictive default budget of 0.03 seconds. In Section 5.4, the authors increase this budget by factors of two and five and find that 0.80% and 1.57% of the evaluated instance labels change, respectively. However, this analysis only examines label stability and does not evaluate the screening gate's false-rejection behavior under the modified planning budgets.

    This limitation is important because CONS-GNN predicts the outcome of a particular planner-budget pair rather than intrinsic task feasibility. Increasing the planning budget may turn some previously Unresolved instances into Feasible ones, potentially changing the false-rejection risk associated with the previously selected threshold.

    So, the current experiments support the effectiveness of CONS-GNN under a fixed planner, computational budget, and synthetic test distribution, but do not establish that its conservative behavior remains reliable across different planning budgets.

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to 1eAr (1/2): Weakness 1 — Contribution and baselines

Part 1/2: Weakness 1. The numbering below follows your original review.

Thank you for the careful assessment of the contribution and evaluation. We have clarified the method's scope, added a closer tuned baseline, and extended the gate evaluation to address the operating-point and budget concerns.

Part 2 addresses Weaknesses 2–3 on matched-FRR comparisons and the frozen gate under larger planning budgets.

Weakness 1. Contribution and comparison with existing relation-aware GNNs

We agree that heterogeneous graphs, relation-aware message passing, auxiliary supervision, and validation-based rejection are established techniques. Our contribution is their task-specific formulation for allocating multi-UAV planner calls. Section 6 now states this scope explicitly.

The graph represents UAV–task assignment compatibility, temporal and spatial competition between tasks, task–region usage, and regional connectivity. Its target is the outcome of a fixed EECBS–3D-SIPP pipeline under a stated computation budget: Feasible, Infeasible, or Unresolved. This separates failure to reach a decision within budget from a definitive negative result of that procedure. At deployment, requests are rejected only when their predicted Infeasible probability exceeds a validation-selected threshold; the reference planner handles the rest.

These principles can be implemented within existing relation-aware GNN and multi-task frameworks. We therefore position the contribution as the task-specific representation, three-outcome formulation, and conservative screening procedure, rather than a new general-purpose GNN learning rule.

To test whether the chosen configuration brings a practical benefit, we added a closer relation-aware baseline. It retains the node types, edge attributes, type-wise pooling, three-way output, and auxiliary difficulty supervision, but uses within-relation mean aggregation. CONS-GNN instead computes message gates from source-node, destination-node, and edge attributes. We also searched 12 hyperparameter configurations for this baseline and Adapted GRN (Supplementary Section H.7).

Thresholds were selected on validation data and frozen before testing. Supplementary Table S15 reports:

MethodFalse rejectionsFRR (%)Call savings (%)
Tuned Adapted GRN30.024139.70
Tuned relation-aware GNN30.024140.34
CONS-GNN30.024141.36

With the same three observed test errors, CONS-GNN avoids 498 and 306 additional calls relative to the two tuned baselines. These are comparisons of the complete tuned configurations, so we do not attribute the entire difference to the message gate alone.

Section 5.2 now reports this comparison alongside the online cost: mean latency falls from 18.84 ms/request without screening to 15.31 ms/request with CONS-GNN, including graph construction and inference. This 3.53 ms, or 18.74%, reduction supports the practical value of the screening procedure beyond its modest gains in classification metrics.

AuthorsOpenReview ↗

Response to 1eAr (2/2): Weaknesses 2–3 — Gate risk and planning budgets

Part 2/2: Weaknesses 2–3. The numbering below follows your original review.

Weakness 2. Three-way versus binary screening at the same FRR

We agree that the approximately tenfold FRR difference at the original operating points is not sufficient to establish a better trade-off: the three-way gate also rejects fewer requests. We have added a comparison at the same observed false-rejection counts in Supplementary Section H.5 and Table S13.

False rejectionsFRR (%)Three-way p_I savings (%)Binary savings (%)
30.024141.3639.92
120.096446.5246.03
300.241050.3852.96
620.498054.4056.12

At 3 and 12 errors, the three-way gate saves 1.44 and 0.49 percentage points more calls, corresponding to 432 and 147 additional avoided calls among 30,000 requests. At 30 and 62 errors, binary collapse has higher savings. FRR here uses pooled counts with 12,449 feasible instances; 30/12,449 gives 0.2410%, whereas the original 0.2414% is a run mean.

We also compare p_I with 1 − p_F = p_I + p_U from the same three-way model. At the two low-FRR points, the latter saves 38.87% and 45.23%, below the 41.36% and 46.52% obtained with p_I alone. Figure S8 presents all three scores. Its lines connect the evaluated points and do not imply separately evaluated intermediate thresholds.

We have revised Contribution 3, Section 5.4, and the Conclusion to limit the advantage to the two reported low-FRR points. These results are from post hoc test-set threshold scans and serve as diagnostics; deployment evaluations still use validation-selected frozen thresholds. They support a benefit in the evaluated low-FRR region, with the preferred trade-off depending on the tolerated cost of false rejection.

Weakness 3. False rejections when the planning budget changes

We agree that label stability alone does not measure the original gate's risk under a larger budget. We have added a frozen-gate evaluation in Supplementary Section H.6 and Table S14, and summarized the false-rejection results in Section 5.4. On 3,000 instances, we keep the model weights and threshold selected under the 0.03 s budget fixed, then rerun the reference pipeline with budgets of 0.06 and 0.15 s.

Budget (s)Feasible instancesCONS-GNN false rejectionsCONS-GNN FRR (%)Binary false rejectionsBinary FRR (%)
0.031,24500.000030.2410
0.061,26210.079260.4754
0.151,27820.1565110.8607

CONS-GNN rejects one newly feasible instance at 0.06 s and two at 0.15 s. These instances become solvable under the larger budget but are still rejected by the original threshold, directly measuring the effect raised in the review.

Because the inputs, models, and thresholds are fixed, the rejection decisions do not change: call savings remain 41.37% for CONS-GNN and 52.97% for binary collapse. What changes is the reference outcome, including the feasible denominator and false-rejection count. Binary collapse also saves more calls, so the FRR difference alone is not a comparison at matched risk.

The validation FRR constraint is empirical; it does not provide a formal guarantee that the same FRR will hold under a shifted data distribution. Our conclusion is restricted to the tested budgets: at two and five times the original time budget, CONS-GNN's observed FRRs are 0.0792% and 0.1565%, with savings of 41.37%. As stated in Section 6, changes in planner, search limits, or task distribution require renewed threshold validation.

Thank you for the additional experiments and clarifications. I appreciate the authors' efforts to address my concerns.

The response for W1 has helped me better understand the practical contribution of this work, particularly the reduction in planner calls. This partially alleviates my original concern. However, from the perspective of further improving the paper, I would still encourage the authors to clarify this contribution more explicitly in the paper's narrative. In particular, it would be helpful to articulate how the proposed components jointly contribute to reducing planner calls and what the central contribution of this domain-specific formulation is.

For W2,the additional results are largely consistent with my original concern. From the perspective of planner-call savings, the three-way does not demonstrate a clear overall advantage over the binary alternative, as its advantage is mainly observed in the low-FRR region, while the binary formulation performs better at higher FRR levels. Nevertheless, I still consider this additional analysis valuable.

For W3, I consider my original concern largely addressed.

I am willing to raise my rating from 2 to 3 in light of the additional experiments and clarifications. However, I would like to emphasize that my overall assessment remains closer to Borderline rather than Weak Accept. As reflected in the responses to W1 and W2, the proposed components have not yet demonstrated clear and substantial performance advantages individually, and the work, in its current form, is better characterized as a domain-specific framework integrating existing techniques than as a significant methodological advance. While I acknowledge its practical value, I believe these limitations should still be taken into consideration in the overall assessment.

Thank you for carefully reviewing the additional experiments and updating your assessment. We agree with your suggestions on the contribution and interpretation of the results and have revised the paper accordingly.

The Introduction and Section 4 overview now clarify the contribution as a task-specific formulation and screening procedure for multi-UAV planner-call allocation, built on established learning techniques. The graph encodes assignment compatibility, task competition, and airspace constraints; relation-aware message passing combines this information, while difficulty supervision is used only during training. The three outcome labels distinguish feasible, definitively negative, and unresolved results of the fixed pipeline under its budget. At deployment, only requests whose infeasibility score exceeds the validation-selected threshold are rejected; all others are forwarded to the planner.

Section 5.2 more clearly separates the practical benefit of the complete pipeline from its modest gains over tuned baselines. Section 5.3 and Figure S5 also clarify that the ablation variants differ in operating points and observed FRRs, so the original savings gaps cannot be interpreted as independent component contributions.

Regarding W2, we agree that the results do not establish a clear overall advantage for three-way screening. Section 5.4 and the Conclusion now explicitly state that, in the post hoc test-set diagnostics, three-way screening saves more calls at the two evaluated low-FRR points, whereas binary collapse performs better at the two higher-FRR points.

We appreciate your suggestions for presenting the work's practical value and limitations more precisely.

Reviewer_t4iNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Paper presents a GNN-based classifier for the feasibility of pre-planner designs for 3D UAV teams. The paper formulates a screening approach that expands standard Feasible/Infeasible to include Unresolved, for fixed planner-budge pairs, to allow for planners. Paper introduces CONS-GNN to encode assignment and constraints. The model is tested on 2 grids with a range of obstacles and no-fly zones and 4 to 10 UAVs. Results care compared against multiple baselines on Accuracy, F1 measure, Latency and other metrics, showing an improved performance. The Empirical results support the presented novelty of the model.

Critical strengths
  1. Technical contribution:
    • The provided method improves on the performance of pre-planner screening.
    • Presence of Discussion of Limitations provide a through discussion on what can be done in the future work.
  2. Empirical strength:
    • Empirical analysis shows that the presented model outperforms multiple baselines. Ablation and Scalability tests show that the design choices all contribute to improved performance.
  3. Relevance and impact:
    • MAPF in 3D space is a hard problem when in the presence of multiple constraints. The paper discusses real world applications in inpections, emergency respons and other areas where a fast filtering of plans if important for efficient plan generation.
Critical weaknesses
  1. Training times are not reported, or discussed. While the problem presented shows scalability, it is unclear how out of distribution or harder problems perform and whether or not re-training the model is needed. One of the primary limitations of ML methods is the added cost of training compared to non-learning models. Does the training time justify the presence of a filtering when it might be easier to simply run the planner at fraction of the time without the screening process.
  2. In Machine Learning, the policy quality is sensitive to hyperparemeters used and it is unclear if the baselines that have close performance to the presented methods can be further improved with sufficient hyperparameter tuning.
  3. It is unclear how much validating the feasibility of a pre-planner is providing an improvement, or how an infeasible flag leads to change in the input configuration to allow for feasible plan generation.
  4. Algorithm 1 needs a more detailed explanation.
  5. Table 5 is missing an indicator for whether or not the metric is higher is better, or lower is better.
Questions for authors
  1. Could the authors clarify if the evaluation on larger UAV team scales is zero-shot or if the model is trained and tested on the same number of tasks? This clarification would improve the quality of reproducibility.
2 comments
AuthorsOpenReview ↗

Response to t4iN (1/2): Weakness 1 and Question 1 — Cost and transfer

Part 1/2: Weakness 1 and Question 1. The numbering below follows your original review.

Thank you for the suggestions on computational cost, baseline tuning, and reproducibility. We have added the corresponding measurements and clarified how the screening stage fits into the planning pipeline.

Part 2 addresses Weaknesses 2–5 on baseline tuning, the role of screening, Algorithm 1, and Table 5.

Weakness 1. Training cost, model reuse, and harder inputs

We agree that the case for screening depends on both its online saving and the number of requests over which its offline cost can be reused. Supplementary Section H.3 and Table S11 now report measured offline time for one seed:

StageTime (s)
Label generation2,826
Training/tuning19,250
Calibration423
Total22,499

The total is approximately 6.25 hours and excludes the additional 12-configuration strong-baseline search. It covers one seed, not all three experimental runs. Online, mean latency is 18.84 ms/request without screening and 15.31 ms/request with CONS-GNN, including graph construction and inference (Section 5.2 and Supplementary Table S10). Dividing 22,499 s by the net saving of 0.00353 s/request gives approximately 6.37 million requests to amortize the recorded offline time.

We have added this accounting to Section 6. It supports reuse over a sustained workload; for a small, one-off batch, directly running the planner may be preferable. The calculation concerns elapsed time under the measured workload, rather than energy or monetary cost.

For larger inputs, the model trained on 3–4 UAVs and 3–8 tasks is reused without retraining, fine-tuning, or threshold selection at the new scale. Tests with 6, 8, and 10 UAVs use 6–12, 8–16, and 10–20 tasks, respectively (Section 5.5; Supplementary Section H.2 and Table S8). No false rejections are observed in these tests, and savings reach 70.50% at 10 UAVs. On the hard subset excluding instances rejected by deterministic pre-checks, the full model saves 33.48% of calls.

We also keep the original model and threshold frozen while increasing the reference budget from 0.03 to 0.06 and 0.15 s (Supplementary Section H.6 and Table S14). CONS-GNN rejects 1 of 1,262 and 2 of 1,278 feasible instances, giving FRRs of 0.0792% and 0.1565%, while savings remain 41.37%. These tests give evidence for reuse under the evaluated changes. They do not provide a risk guarantee for arbitrary shifts; changes in planner, search limits, or task distribution require renewed threshold validation, as stated in Section 6.

Question 1. Zero-shot evaluation at larger team sizes

The larger-team evaluation is zero-shot: it reuses the trained model and threshold without retraining, fine-tuning, or threshold selection at the target size. We have clarified this in Section 5.5 and specified the task-count ranges in Supplementary Section H.2.

SettingUAVsTasks per instance
Training3–43–8
Larger-team test66–12
Larger-team test88–16
Larger-team test1010–20

Each tested team size has 1,200 instances. The 4-UAV test is an in-range control; the 6-, 8-, and 10-UAV tests increase both team size and task count. Table 5 gives savings, and Supplementary Table S8 supplies the false-rejection counts and their denominators:

UAVsCONS-GNN savings (%)False rejections / feasible instances
441.250 / 480
652.330 / 360
861.170 / 280
1070.500 / 210

With the model and threshold unchanged, no false rejections are observed at these four sizes, and savings reach 70.50% at 10 UAVs.

AuthorsOpenReview ↗

Response to t4iN (2/2): Weaknesses 2–5 — Baselines and clarifications

Part 2/2: Weaknesses 2–5. The numbering below follows your original review.

Weakness 2. Additional tuning of strong baselines

We have extended the comparison with a 12-configuration search for Adapted GRN and a relation-aware GNN: learning rates {0.0001, 0.0003, 0.001}, hidden widths {64, 128}, and dropout rates {0, 0.1}. Supplementary Section H.7 describes the search and baseline construction.

The relation-aware baseline retains the same node types, edge attributes, type-wise pooling, three-way output, and auxiliary difficulty supervision as CONS-GNN. It uses within-relation mean aggregation, providing a closer comparison under the same input representation and training objectives.

Thresholds are selected on the validation set and fixed before evaluation on the test set. At its fixed threshold, each method falsely rejects three feasible test instances; thresholds are not adjusted on test labels to obtain that count. The results are in Supplementary Table S15 and summarized in Section 5.2:

MethodFalse rejectionsFRR (%)Call savings (%)
Tuned Adapted GRN30.024139.70
Tuned relation-aware GNN30.024140.34
CONS-GNN30.024141.36

At the same observed error count, CONS-GNN saves 1.66 and 1.02 percentage points more calls than the tuned baselines, corresponding to 498 and 306 additional avoided calls among 30,000 requests.

Weakness 3. What screening improves and how rejected inputs are handled

The screener decides whether to invoke the reference pipeline; it does not modify the task configuration. We have clarified the handling of rejected requests in the Introduction and the roles of the screener, planner, and upstream manager in Supplementary Section H.1.

On the main benchmark, screening avoids 41.36% of planner calls while retaining 99.9759% of planner-feasible instances. Including graph construction and inference, mean latency falls from 18.84 to 15.31 ms/request. Section 5.2 now includes this time comparison, so the measured benefit is explicit.

A request is rejected only when p_I exceeds the selected threshold. Forwarded requests still undergo planning and plan validation; no plan is generated for a rejected request. An upstream manager could revise time windows, the task set, or available resources and submit a new request. Automatic input repair is not evaluated, and we do not count any benefit from such repair. Section 3 also clarifies that the labels describe the fixed reference procedure: a screening rejection is not a proof of intrinsic infeasibility.

Weakness 4. Explanation of Algorithm 1

We have added a guide in Section 4.3 and a “Reading Algorithm 1” paragraph in Supplementary Section H.1. The algorithm has three stages:

  1. Offline training. Inputs are the instance generator, reference pipeline with budget B, candidate threshold set, FRR limit, and auxiliary-loss weight. Disjoint training and validation sets are generated. Pre-planner inputs form the graphs, and offline reference calls supply the Feasible/Infeasible/Unresolved labels and difficulty targets. Only the training set updates the model, using classification and auxiliary difficulty losses.
  2. Checkpoint and threshold selection. Evaluate retained checkpoint–threshold pairs on the validation set. Select the pair with the highest savings subject to FRR ≤ 0.5%, breaking ties by lower FRR. If none qualifies, use the best validation checkpoint with threshold 1, passing all requests. Test data do not participate in selection.
  3. Online screening. Freeze the selected model and threshold, construct the new request's graph, and obtain p_F, p_I, and p_U. Reject only when p_I exceeds the threshold; otherwise call the reference pipeline.

An instance is still forwarded when Infeasible is its most likely class but p_I does not exceed the threshold. The auxiliary difficulty head is used only during training and does not enter the online rejection decision.

Weakness 5. Metric directions and units in Table 5

We have added upward arrows for call savings and throughput, and downward arrows for FRR and latency, in Table 5 on page 9. We have also expanded “Lat.” and “Thr.” to “Latency” and “Throughput,” with units of seconds/instance and instances/second. The caption explains the arrows and notes that call savings should be read together with FRR.

Reviewer_MVv1OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes CONS-GNN, which decides whether to call a fixed multi-UAV planner (EECBS with 3D SIPP, 0.03 s budget) on each instance. From a heterogeneous UAV–task–region graph built before planning, it predicts the planner's outcome: Feasible, Infeasible, or Unresolved. An instance is rejected only if its infeasibility probability p_I exceeds a threshold that maximizes validation rejections under a 0.5% false-rejection rate (FRR) cap. On three synthetic 10,000-instance test sets with 3–4 UAVs, CONS-GNN avoids 41.36% of planner calls and wrongly rejects 3 of 12,449 feasible instances. It outperforms eight baselines that use the same gate. The paper also reports ablations, a hard subset, a binary-label variant, scaling to 10 UAVs, and ROS 2/Gazebo tests.

Critical strengths
  1. Defining outcomes relative to a fixed planner and budget, and keeping timeouts separate from definitive infeasibility, is a sensible way to pose learned pre-screening.
  2. Feature provenance is disciplined. Only pre-planner information is used at deployment, and planner statistics serve only as offline labels and auxiliary targets.
  3. All baselines share the same gate and validation-only threshold selection, with thresholds frozen before testing.
  4. The headline FRR is reported as raw counts. The hard-subset analysis is a useful attempt to separate learned screening from deterministic filtering.
  5. The limitations section is candid about the empirical FRR constraint and the missing cost breakdown.
Critical weaknesses
  1. The problem and setting are not clearly defined A. A few citations are not enough for the reader to understand the major problem this work is trying to solve, and the introduction does not give a fair picture of what the work is about. Its opening paragraph moves from UAV applications to MAPF/SIPP to allocating planner time across repeated requests in four sentences, relying on citations rather than explanation. It never describes a concrete scenario: who generates the candidate instances, why many must be screened, and what happens to a rejected request. B. What exactly is planning here? Does it account for both task allocation and path planning? Section 3 mentions assignment logic followed by EECBS–3D-SIPP. It never explains how tasks are assigned, how many tasks an instance contains, or whether a UAV can serve several tasks. C. What is pre-planning? It is not clear what pre-planner screening is. The term appears in the abstract and introduction before any definition, leaving the reader to infer it from Figure 1. The boundary of "pre-planner" information also needs defining, since the root-conflict features overlap with the planner's own first step. D. What is a release time? The term is used in Sections 3 and 4.1 without definition. Please provide examples that show the physical relevance of this problem: ◦ what a task, release time, and deadline correspond to in practice; ◦ what physical scale a grid cell and a time step represent; ◦ whether 3–4 UAVs on a 10×10×3 grid with a 0.03 s budget reflects a realistic deployment. E. What exactly is the pipeline in Section 3?

  2. The three-way vs. binary comparison does not support Contribution 3 Both models were selected under the same validation FRR budget, and the binary model simply used more of it. It avoids 11.6 percentage points more calls (about 3,500 instances across the test sets) for roughly 27 more false rejections. That is about one false rejection per 130 additional avoided calls. Whether that trade is worse depends on the cost ratio. Section 5.4 concedes that dominance at matched risk is not established, yet the introduction presents the result as a finding.

  3. The ablation results are hard to interpret Removing any single component, including the training-only difficulty head, reduces savings by a narrow 5.23–6.00 points. Every ablated variant (35.36–36.13%) ends up below every GNN baseline (37.03–39.11%). On the hard subset, the drops are only 1.71–2.53 points, less than half the full-test drops. So more than half of each ablation's loss comes from instances that deterministic pre-checks already reject, where architecture should matter least. This pattern points to threshold selection rather than the components themselves. Ablation FRRs are not reported.

Questions for authors
  1. Please describe the reference pipeline step by step, including the pre-checks, assignment procedure, SIPP constraints, search limits, and where each label can arise. Does planning include task allocation, and can a UAV serve multiple tasks?
  2. What is the end-to-end latency with no screening? How is per-instance cost split among graph construction, inference, and planning?
  3. How were the standard deviations in Tables 2 and 3 computed?
  4. What is the Feasible/Infeasible/Unresolved distribution for each test set and each team size?
  5. What thresholds were selected for each method and ablation, and what grid was searched? What were the ablated variants' validation and test FRRs?
  6. At matched FRR, how do savings compare for three models: the three-way model gated on p_I, the three-way model gated on 1 − p_F, and the binary model? A savings–FRR curve would answer this directly.
4 comments
AuthorsOpenReview ↗

Response to MVv1 (1/4): Weakness 1 — Problem setting

Part 1/4: Weakness 1. The numbering below follows your original review.

Thank you for the detailed review. We have clarified the planning setting and reference pipeline, added cost and error-count breakdowns, and revised the interpretation of the gate and ablation comparisons. The responses below point to the corresponding changes in the revised paper and supplementary material.

The remaining parts cover the gate comparison (Part 2), ablations and thresholds (Part 3), and the pipeline, latency, and statistics (Part 4).

Weakness 1. Problem setting, requests, and the meaning of pre-planner screening

A. Where requests come from and what rejection means. Consider a multi-UAV inspection service: an upstream manager assembles a task list, available UAVs, time windows, and airspace restrictions into a candidate request. The reference pipeline assigns tasks and searches for valid multi-UAV paths. Even an unsuccessful or timed-out call consumes computation. Our screener decides whether to invoke this pipeline for the current request. A rejected request receives no plan; the upstream manager may adjust its inputs and resubmit it. Automatic repair and resubmission are outside the evaluated system. We have added this example to the Introduction. The main benchmark uses synthetic requests; the example explains the intended application.

B. Assignment and path planning. The reference pipeline includes both stages. Each main-benchmark instance has 3–4 UAVs and 3–8 tasks, and one UAV can serve several tasks in sequence. Tasks are processed by deadline, release time, and identifier. A fixed greedy insertion procedure considers positions in each UAV's task sequence, excludes insertions that violate cumulative time or resource lower bounds, and selects the smallest increase in static path length, with deterministic tie-breaking. EECBS–3D-SIPP then searches for collision-free paths while keeping the assignment and task order fixed. Transfers between tasks count toward time and resource use. Section 3 now states the setting, and Supplementary Section H.1 gives the procedure.

C. The screening boundary. “Pre-planner” means before the full reference call. Inputs include the request, map, and features computed from them; the screener does not read the current instance's high-level search trace or final outcome. Root-conflict features use independently computed lower-bound paths, so their computation overlaps with planner initialization. We have clarified this in Section 4.1. The online measurements include this feature cost: with graph construction and inference included, mean latency falls from 18.84 to 15.31 ms/request (Supplementary Table S10).

D. Time windows and physical scale. A release time is the earliest task start, and a deadline is the latest arrival at its goal. For example, an inspection task may start no earlier than 9:00 and must reach its destination by 9:10. Section 3 now gives this example and states that the benchmark grid cells and time steps are abstract units, without a common conversion to metres or flight seconds. The 0.03 s limit is a computation budget. The small grids support controlled comparisons; the separate ROS 2/Gazebo tests cover six scenarios and 240 instances, saving 32.5%–50.0% of calls with no observed false rejections (Supplementary Tables S3–S4). We also evaluate the frozen gate at 0.06 and 0.15 s in Supplementary Section H.6 and Table S14.

E. The complete pipeline. The order is necessary-condition checks, task assignment and ordering, EECBS high-level conflict search with low-level 3D SIPP, and complete-plan validation. A valid complete plan yields Feasible; a definitive negative result from the fixed procedure yields Infeasible; reaching a time or search limit without a definitive result yields Unresolved. Infeasible is a pipeline outcome, not a proof that every possible assignment or planner would fail. Section 3 and Supplementary Section H.1 now make these stages and label meanings explicit.

AuthorsOpenReview ↗

Response to MVv1 (2/4): Weakness 2 and Question 6 — Gate comparison

Part 2/4: Weakness 2 and Question 6. The numbering below follows your original review.

Weakness 2. Contribution 3 and the three-way versus binary trade-off

We agree that satisfying the same validation FRR limit does not establish an advantage at the same test risk. The original binary operating point saves more calls and makes more false rejections, so the approximately tenfold FRR difference alone does not establish which trade-off is preferable.

We have added comparisons at the same observed false-rejection counts in Supplementary Section H.5 and Table S13. At 3 and 12 false rejections, the three-way gate saves 41.36% and 46.52% of calls, compared with 39.92% and 46.03% for binary collapse. These gains of 1.44 and 0.49 percentage points correspond to 432 and 147 additional avoided calls among 30,000 requests. At 30 and 62 false rejections, binary collapse instead saves more calls: 52.96% versus 50.38%, and 56.12% versus 54.40%.

We have revised Contribution 3, Section 5.4, and the Conclusion to state the advantage at the two reported low-FRR points, rather than across all risk levels. The preferred operating point depends on the relative costs of false rejection and planner invocation. These comparisons use post hoc test-set threshold scans; the deployment results continue to use thresholds selected on validation data and frozen before testing. The full three-score comparison and curve are provided in Table S13 and Figure S8.

Question 6. Comparing all three gate scores at matched FRR

We have added the requested comparison to Supplementary Section H.5, Table S13, and Figure S8, and summarized it in Section 5.4.

False rejectionsFRR (%)Three-way p_I savings (%)Three-way 1 − p_F savings (%)Binary savings (%)
30.024141.3638.8739.92
120.096446.5245.2346.03
300.241050.3851.1452.96
620.498054.4055.0856.12

FRR divides false rejections by 12,449 feasible instances; savings divide rejected requests by 30,000. The 0.2410% value is the pooled count 30/12,449, whereas the original binary FRR of 0.2414% is a mean across runs.

The two three-way scores come from the same model: 1 − p_F = p_I + p_U. At 3 and 12 errors, gating on p_I saves 1.44 and 0.49 percentage points more calls than binary collapse, and 2.49 and 1.29 points more than gating on p_I + p_U. Binary collapse has higher savings at 30 and 62 errors.

Figure S8 connects the evaluated points; the connecting lines do not represent separately evaluated intermediate thresholds. This is a post hoc diagnostic using test labels to locate comparable operating points, reported separately from deployment with validation-selected thresholds.

AuthorsOpenReview ↗

Response to MVv1 (3/4): Weakness 3 and Question 5 — Ablations and thresholds

Part 3/4: Weakness 3 and Question 5. The numbering below follows your original review.

Weakness 3. Interpreting the component ablations

The operating-point difference is important here. Supplementary Table S9 now reports thresholds, validation and test FRRs, and false-rejection counts for the original ablations. Their test FRRs range from 0% to 0.0161%, below the full model's 0.0241%. Thus, the original savings gaps of 5.23–6.00 percentage points combine component and operating-point effects; they cannot be read as isolated component contributions.

Supplementary Section H.4 and Table S12 provide additional operating points, with thresholds selected on validation data and frozen before testing:

ConfigurationOriginal savings (%)Additional savings (%)Gap to full model (pp)
Full CONS-GNN41.3641.36—
No region nodes35.5140.281.08
No task–task edges35.6939.551.81
No task–region edges35.5940.390.97
No difficulty head36.1340.800.56
No edge features35.3639.212.15
No heterogeneity35.9839.701.66

The full model has the highest savings at these reported points. The difficulty-head gap decreases from 5.23 to 0.56 percentage points, supporting the reviewer's concern that threshold selection affects the original gaps. We describe these as additional validation-selected operating points, not as a comparison at matched test FRR; their savings do not establish superiority over every ablation at equal risk.

On the hard subset, using the corresponding additional thresholds without further selection, the full model saves 33.48% of calls and the ablations save 31.37%–33.13%. Its savings advantage therefore also appears after removing instances rejected by deterministic pre-checks. The standard and hard subsets have different denominators, so subtracting their percentage-point gaps does not identify how many lost rejections came from each group. Section 5.3 now reflects these distinctions and avoids assigning the full original gap to an individual component.

Question 5. Threshold grid and ablation FRRs

The main comparisons and original ablations use the threshold grid {0.50, 0.60, 0.70, 0.80, 0.90, 0.95, 0.98}. Checkpoint–threshold pairs are selected on validation data to maximize call savings subject to FRR ≤ 0.5%, breaking savings ties by lower validation FRR. They are then frozen before testing. Supplementary Sections B and H.2 now state the selection procedure and grid together with the error counts.

The main-method thresholds remain listed in Supplementary Table S1:

MethodThreshold
Rule Filter0.90
XGBoost0.90
MLP-Global0.80
GAT0.80
MPNN0.70
R-GCN0.80
HGT0.70
Adapted GRN0.95
CONS-GNN0.80

Supplementary Table S9 gives the original ablation operating points. Parentheses contain false-rejection counts.

ConfigurationThresholdValidation FRR (%)Test FRR (%)
Full model0.800.0964 (6)0.0241 (3)
No region nodes0.950.0803 (5)0.0080 (1)
No task–task edges0.950.0643 (4)0.0080 (1)
No task–region edges0.900.0964 (6)0.0080 (1)
No difficulty head0.900.1285 (8)0.0161 (2)
No edge features0.980.0482 (3)0.0000 (0)
No heterogeneity0.900.0803 (5)0.0080 (1)

These FRRs pool counts across seeds, using 6,225 feasible validation instances and 12,449 feasible test instances. The main result tables instead report run means and sample standard deviations. The ablations' lower test FRRs help explain why their savings gaps cannot be attributed solely to removed components. Table S9 applies only to the original thresholds; Table S12 reports the separate additional operating points discussed in our ablation response. Supplementary Section B also distinguishes deployment threshold selection from the post hoc gate-score scans in Section H.5.

AuthorsOpenReview ↗

Response to MVv1 (4/4): Questions 1–4 — Pipeline and statistics

Part 4/4: Questions 1–4. The numbering below follows your original review.

Question 1. Reference pipeline, constraints, and label generation

We have expanded Section 3 and Supplementary Section H.1 with the following sequence. Planning includes task allocation, and a UAV can execute multiple tasks.

  1. Necessary-condition checks: check start/goal and airspace legality, static connectivity, and resource lower bounds. A failed necessary condition returns Infeasible.
  2. Assignment and ordering: process tasks by deadline, release time, and identifier. Greedy insertion selects the eligible insertion with the smallest increase in static path length, subject to cumulative time and resource lower bounds. Failure returns Infeasible for this fixed procedure, without proving that all possible assignments fail.
  3. Path search: keep assignment and task order fixed. EECBS resolves inter-UAV conflicts by imposing constraints on low-level 3D SIPP searches. Paths must remain within the allowed cells and altitude limits, avoid obstacles and no-fly regions, and satisfy task-window and resource constraints. Transfers between tasks count toward time and resource use. One failed low-level branch does not determine the entire instance's label, and failed search does not trigger reassignment.
  4. Validation and termination: a valid complete plan returns Feasible; a definitive negative pipeline result returns Infeasible. Reaching a time or search limit without a definitive result returns Unresolved.

The default 0.03 s budget covers the complete reference call. EECBS uses w = 1.1 and at most 1,000 high-level expansions; all SIPP calls together are capped at 20,000 state expansions. Offline labeling and online planning of forwarded requests use this same procedure.

Question 2. End-to-end latency and its components

Mean end-to-end latency without screening is 18.84 ms/request, compared with 15.31 ms/request for CONS-GNN. We have added the breakdown to Supplementary Section H.3 and Table S10, and the net reduction to Section 5.2.

MethodPreprocessingGraph constructionInferencePlanningTotal
No screening0.800018.0418.84
CONS-GNN0.801.100.4512.9615.31

All values are milliseconds averaged over all submitted requests. Rejected requests contribute zero planning time. For forwarded requests, planning includes internal pre-checks, assignment, path search, and plan validation. Common preprocessing is timed separately; feature computation is included in preprocessing and graph construction, without double-counting stages.

Graph construction and inference add 1.55 ms/request, while average planning time decreases by 5.08 ms/request. The net saving is therefore 3.53 ms/request, or 18.74%, after screening overhead.

Question 3. Standard deviations in Tables 2 and 3

Tables 2 and 3 report the mean and sample standard deviation over three independent runs, with seeds 37, 41, and 43. Each run includes its own data generation and model training. Given one metric value per run, we average the three values and compute the sample standard deviation with denominator 3 − 1 = 2.

These statistics describe variation between runs, rather than between individual test instances. For example, the three CONS-GNN FRRs are 1/4,000, 1/4,149, and 1/4,300. Their mean is 0.0241%, and their sample standard deviation is approximately 0.0009 percentage points. Section 5.1 now identifies the statistical unit, and Supplementary Section H.2 gives the formula and this example.

Question 4. Class distributions by test set and team size

We have added the class counts to Supplementary Section H.2, Tables S7–S8, with a pointer in Section 5.1.

Test seedFeasibleInfeasibleUnresolvedTotal
374,0004,5001,50010,000
414,1494,3501,50110,000
434,3004,2001,50010,000
Total12,44913,0504,50130,000
UAVsFeasibleInfeasibleUnresolvedTotal
44805401801,200
63606501901,200
82807401801,200
102108501401,200

The feasible proportion decreases as team size increases in these tests. CONS-GNN makes no observed false rejections at the four sizes: 0/480, 0/360, 0/280, and 0/210, respectively.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.