acceptodds
Under review as a conference paper at ICLR 2027

Real-Time Neural Vehicle Routing under Delivery Commitments

Abstract

Vehicle routing in the real world is inherently dynamic, where travel times fluctuate and orders are cancelled or released mid-operation, demanding real-time optimization. Existing Neural Vehicle Routing (NVR) methods can infer high-quality routes within milliseconds, yet they remain static and cannot respect delivery commitments under dynamic events. In this paper, we propose a Real-Time NVR (RT-NVR) approach, which extends NVR from static optimization to online decision-making. During training, a View-Comparative Policy Optimization (VCPO) method is designed to train an encoder-decoder model with state broadcasting, where the performance gap between geometrically equivalent views is turned into a learning signal via a contrast-adaptive baseline. In addition, a Spatiotemporal Compatibility-Guided Sampling (SCGS) strategy is devised to inject spatiotemporal priors into decoding, modulating logits toward promising actions. During real-time deployment, a Delivery Commitment Set-based Masking (DCSM) mechanism derives the vehicle load peak and maximum slack time of the delivery commitment set for a vehicle, establishing a provably safe decision boundary. Extensive experiments demonstrate that VCPO outperforms state-of-the-art training methods, and that the real-time approach reduces travel cost, fleet size, and time delay by 19.16%, 42.14%, and 96.73%, respectively, over heuristic methods. Statistical analyses further reveal that state broadcasting carries the most salient routing knowledge, and that the model parameters evolve along a low-dimensional manifold whose top 10 principal components capture 90.6% of the variance.

Primary area:
optimization
Keywords:
Deep Reinforcement Learning, Vehicle Routing Problem, Neural Combinatorial Optimization, Real-Time Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_AUZzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RT-NVR (Real-Time Neural Vehicle Routing), which extends neural vehicle routing methods from static optimization to real-time online decision-making. When dynamic events occur, the model solves the routing plan under the new state while maintaining the delivery commitment set originally assigned to each vehicle, through the Delivery Commitment Set-based Masking (DCSM) . The method proposes VCPO, which converts the performance gap between geometrically equivalent views into a learning signal via a contrast-adaptive baseline(CAB). Spatiotemporal Compatibility-Guided Sampling (SCGS) is used to modulate the decoding process, yet its core appears to merely introduce a heuristic term. Extensive experiments are finally conducted to verify the performance of the method and its planning capability when tackling dynamic problems.

Critical strengths
  1. This paper investigates a dynamic vehicle routing problem with delivery commitments. Compared with conventional static problems, this problem has greater practical significance and is more challenging.

  2. The experiments are relatively comprehensive, with sufficient tests conducted for different   probabilities fcf_c of cancellation event and the factor fif_i for new order adding. Compared with other neural solvers, VCPO achieves overall better performance.

Critical weaknesses
  1. Methodological innovation is relatively limited, consisting mainly of incremental improvements to existing methods. For example, the proposed VCPO is a minor modification built upon existing multi-view methods(e.g., sym-nco), where the baseline is changed from the average baseline to baselines of high and low views. Spatiotemporal Compatibility-Guided Sampling (SCGS) can also be regarded as a heuristic term added on logits, with limited innovation.

  2. There are some doubts regarding the description of the method. The approach decomposes the dynamic problem into multiple VRP-FSDC tasks, and re-plans the complete solution when dynamic events occurs. However, upon the occurrence of an event, multiple vehicles may have different positions and intermediate states, and the form of the solution will differ significantly from that of the standard VRP problem in Section 3.1. However, this method seems not to consider the state information of multiple vehicles, and the solution form is not clearly defined. Besides, regarding state broadcasting, the paper does not elaborate on this mechanism. It merely concatenates the current embedding eue_u with vehicle states.

  3. The baselines used in the experiments are outdated (2017, 2021, 2024) and do not include methods specifically designed for dynamic tasks. In addition, my another concern is that unlike common fixed VRP test datasets, the test instances described in the experiments are random, meaning they are not consistent across all methods, which may introduce a certain degree of unfairness. Ablation studies for the core component SCGS are also missing. Although Figure 3 presents results for the REIN method without SCGS, the performance of VCPO with SCGS removed is not provided.

Questions for authors
  1. As described in W1, how does the method handle the states of multiple vehicles running in parallel in dynamic problems? Sequential solutions generated by traditional methods such as POMO seem unable to represent the solutions to this multi-vehicle multi-origin problem.

  2. Please explain what state broadcast is(see W1).

  3. Do the encoder and decoder in the model rerun after every dynamic event? If there are no order cancellations or new orders, will replanning still be triggered when a vehicle arrival event occurs?

  4. It is noted that the setting of this paper is training in a static environment and testing in a dynamic environment. Would dedicated training for dynamic environments enable better capture of dynamic characteristics?

  5. As described in W3, random test instances are inconsistent across different algorithms, which may lead to certain unfairness. Especially for order cancellation events, different algorithms may select different routes, resulting in different candidate sets of customers for cancellation operations. Even if the random seed is fixed, the above problem seems unsolvable.

  6. What exactly does the Delay metric in Table 2 refer to? It does not seem to be explained in the paper.

  7. As described in W3, how does the performance of VCPO change after removing SCGS? In addition, in many sets of experiments, as fif_i increases, for example, at 0.4 and 0.5, why does the VCPO algorithm no longer show advantages over other algorithms?

7 comments
AuthorsOpenReview ↗

Reply to Reviewer AUZz

Thanks for your positive comments.

We appreciate your recognition and your comments on methodological novelty, multi-vehicle states and state broadcasting, baselines and instance randomness, the missing SCGS ablation, the Delay metric, and the high-event-frequency behaviour were very helpful; the corresponding clarifications are now given in Section 4.1 (solution construction and state broadcasting), Section 4.3 (replanning triggers), Section 5.2 (SCGS ablation, Delay definition, high fif_{\mathrm{i}} analysis) and Section 6, while the points on novelty and baselines are addressed in our response below. Thank you again for your careful review.

Responses to Weakness 1:

Our contribution is to make NVR deployable in real-time operations, not to propose another static VRP architecture. We have rewritten our conctributions in both abstract and introduction. Please see the revised manuscript.

Responses to Weakness 2:

The approach does not decompose the problem. Decoding is performed vehicle by vehicle rather than for multiple vehicles simultaneously. To show this part clearly, we have added the “Solution Construction” paragraph in Section 4.1 on page 5 of the revised manuscript.

We have also further explained state broadcasting in the “Decoder” paragraph in Section 4.1 on page 4 of the revised manuscript.

Responses to Weakness 3:

We did not include methods designed for dynamic tasks because they do not consider delivery commitment constraints and give no safe decision boundary.

Fairness is guaranteed, because the event flow of order releases and cancellations is determined before execution rather than by the routes an algorithm takes. All algorithms are evaluated on the same event flows. The procedure is implemented in realtime_instances.py and realtime_inference.py of the supplementary material.

Following your suggestions, we have reported the results of VCPO with SCGS removed and added related discussions in the revision. Please see the “Training performance comparison” paragraph in Section 5.2 on page 7 of the revised manuscript.

Responses to Question 1:

When a vehicle returns to the depot, the vehicle state is reset and the routing of the next vehicle starts, which is the same autoregressive decoding used in POMO and standard NVR methods. The solution has the same form as that of the static problem, and the difference lies only in how it is constructed, through the alternation of real-time optimization and deployment.

To show this part clearly, we have added the “Solution Construction” paragraph in Section 4.1 on page 5 of the revised manuscript.

Responses to Question 2:

Following your suggestions, we have further explained state broadcasting in the “Decoder” paragraph in Section 4.1 on page 4 of the revised manuscript.

Responses to Question 3:

We have added “At each replanning, the encoder is run once, and the decoder then constructs the solution autoregressively.” in the “Planning Level” paragraph in Section 4.3 on page 6 of the revised manuscript.

Only three dynamic events trigger the planning level, namely a released pickup order, a cancelled order, and an arrival later than the customer's latest arrival time. When none of these three dynamic events occurs, a vehicle arrival alone does not trigger replanning.

Responses to Question 4:

If dynamic scenarios contained predictable patterns, training in a dynamic environment would indeed lead to better performance. Order releases and cancellations, however, are difficult to predict reliably, so training directly in a dynamic environment would introduce noise and make it hard for the model to learn the basic routing strategy.

Training in a static environment can be regarded as the first step. The next step is to study how to fine-tune the model efficiently in real-time scenarios, where improved decision flexibility leaves low-cost room for responding to unpredictable dynamic events. This step is stated as future work in the Conclusion on page 10.

Responses to Question 5:

Please see Responses to Weakness 3.

Responses to Question 6:

We have rewritten “…, Delay denotes the average total delay time over served customers, i.e., the total excess of realized arrival times over latest arrival times, …; all metrics are averaged over instances and event-frequency combinations.” in the revision. Please see the “Real-time performance comparison with heuristic methods” paragraph in Section 5.2 on pages 7-8 of the revised manuscript.

Responses to Question 7:

Following your suggestions, we have reported the results of VCPO with SCGS removed in Figure 3 and Figure 4.

The performance degradation at fi=0.4f_{\mathrm{i}}=0.4 and fi=0.5f_{\mathrm{i}}=0.5 arises because greedy replanning may restrict future decision flexibility, which is a direction for future research and may be handled through fine-tuning. VCPO is not the only method affected, as ITAB and ITDA also degrade in performance at fif_{\mathrm{i}} of 0.4 or higher.

Thank you for the author's reply. Some of my concerns have been addressed, but I still have the following questions:

  1. Although a static VRP problem can naturally be regarded as multiple trips of a single vehicle, multiple vehicles are operating simultaneously in the dynamic VRP. When a dynamic event occurs, each vehicle may be at a different location and in a distinct state. How is replanning carried out at this point? It seems that such replanning cannot be treated as multiple trips of a single vehicle as in the static problem.

  2. According to the author, the event stream is fixed. In that case, I can understand that order release is fair. However, with regard to order cancellation: if the order has already been visited before it is cancelled, will this cancellation event be ignored?

AuthorsOpenReview ↗

Reply to Reviewer AUZz

Thanks for your comments.

  1. Multiple vehicles are en route at the same time, and replanning accounts for that. Specifically, the decoder processes the vehicles one by one in index order; for each vehicle, it starts from that vehicle's current position and real-time state, applies DCSM over its delivery commitments, and autoregressively constructs that vehicle's remaining route only, and then moves on to the next vehicle until all vehicles are routed and no customer remains unrouted. Consequently, the static multi-trip form is merely the special case in which all vehicles start from the depot.

  2. Cancellation is only valid under a precondition: if the customer has already been served, is being served, or has already been reserved for service, the order can no longer be cancelled, and the corresponding event is not generated, i.e., it is ignored. This matches practice, since a customer cannot withdraw an order that is already on the vehicle or has already been delivered.

Thank you again for your careful review.

Reviewer_8MPkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies real-time vehicle routing under dynamic events and formulates each replanning step as a VRP-FSDC with fixed start nodes and delivery commitments. Offline, RT-NVR trains an encoder-decoder model with VCPO, which uses the performance gap between equivalent views as a baseline, and with SCGS, which adds spatiotemporal priors to sampling. Online, DCSM masks infeasible actions using the vehicle load peak and maximum slack time. The method is compared with several training methods and heuristic real-time algorithms.

Critical strengths
  1. Bringing delivery commitments into neural routing makes the setting closer to real logistics operations.
  2. DCSM gives a clear feasibility boundary based on load peak and slack time, with derivations in the appendix.
Critical weaknesses
  1. In Table 1, gaps between training methods are around 0.1 while standard deviations are 1 to 2.5. VCPO also wins only 2 of 10 settings with fi of 0.4 or higher.
  2. The real-time baselines are simple heuristics. Re-optimization with PyVRP under the same time budget and existing DRL methods for dynamic routing are missing.
  3. The number of customers and the fleet settings are not given in the main text, so scalability is unclear.
  4. The parameter statistics and PCA in Section 5 add little to the main claims. Ablations of SCGS and DCSM in the real-time setting would be more useful.
Questions for authors
  1. What are the number of customers, the maximum fleet size, and the vehicle capacity used in the experiments?
  2. How does PyVRP re-optimization perform with a time limit of 0.1 or 1 second?
  3. How well does the method generalize to larger instances?
1 comment
AuthorsOpenReview ↗

Reply to Reviewer 8MPk

Thanks for your positive comments.

We appreciate your recognition of bringing delivery commitments into neural routing and of the feasibility boundary that DCSM derives from the load peak and maximum slack time, and your comments on the Table 1 statistics, the real-time baselines, the instance and fleet settings, and the additional ablations were very helpful; the settings are now given in Section 5.1, the SCGS and DCSM ablations are reported in Table 4, and the Table 1 comparison is clarified in Section 5.2 and Section 6, while the points on the baselines, the PyVRP time limit, and larger instances are addressed in our response below. Thank you again for your careful review.

Responses to Weakness 1:

Each value in Table 1 is the average travel cost over 100 random instances instead of 100 independent runs. Therefore, the high standard deviations stem from the scale differences between random instances rather than algorithm behavior. The optimal travel costs of different instances differ numerically.

We acknowledge the limitation of offline training and online deployment. ITAB and ITDA also degrade in performance with fi of 0.4 or higher. The reason is that greedy replanning may restrict future decision flexibility, which can be a future research direction and may be handled through fine-tuning. To show this part clearly, we have written “... The underlying reason is that greedily replanning may restrict future decision flexibility.” in the revision. Please see the “Real-time performance comparison with NVR methods” paragraph in Section 5.2 on page 7 of the revised manuscript.

Responses to Weakness 2:

PyVRP targets standard VRP setting, where customers are interchangeable across vehicles, whereas VRP-FSDC fixes each vehicle’s start node and committed deliveries, which PyVRP cannot handle. Enforcing delivery commitments means rewriting these operators, after which the solver is no longer PyVRP and the comparison is no longer across standard methods. Existing DRL methods for dynamic routing also share this limitation.

Responses to Weakness 3:

To clarify the number of customers and the fleet settings, we have added “Each instance contains N=50N=50 customers and a depot, with a vehicle capacity of Q=1500Q=1500 kg and a service time of si=5s_i=5 min. The maximum fleet size is set to Kmax⁡=10K_{\max}=10.” in the revision. Please see the “Real-Time Instance Generation” paragraph in Section 5.1 on page 6 of the revised manuscript.

Responses to Weakness 4:

Following your suggestions, we have conducted ablations of SCGS and DCSM and added “In addition, the ablations of DCSM and SCGS are reported in Table 4, where ...” in the revision. Please see the “Ablation Analysis” paragraph in Section 5.2 on page 9 of the revised manuscript.

Responses to question 1:

To clarify the number of customers, the maximum fleet size, and the vehicle capacity used in the experiments, we have added “Each instance contains N=50N=50 customers and a depot, with a vehicle capacity of Q=1500Q=1500 kg and a service time of si=5s_i=5 min. The maximum fleet size is set to Kmax⁡=10K_{\max}=10.” in the revision. Please see the “Real-Time Instance Generation” paragraph in Section 5.1 on page 6 of the revised manuscript.

Responses to question 2:

A direct comparison is difficult to establish, because PyVRP cannot handle the constraint of delivery commitments. PyVRP targets the standard VRP setting, whereas VRP-FSDC fixes each vehicle’s start node and committed deliveries, which PyVRP does not represent and model. Its search operators move customers between routes, so a committed delivery can be reassigned or dropped. Enforcing delivery commitments means rewriting these operators, after which the solver is no longer PyVRP and the comparison is no longer across standard methods.

Responses to question 3:

The architecture places no constraint on the number of customers, as the encoder and decoder operate on a variable-size node set through attention and none of their dimensions depends on NN, so a model trained at N=50N=50 can be applied to other scales without architectural change. The decision cost also grows slowly with NN. At each step the candidate set is Ck={b1}∪Pk\mathcal{C}_k=\{b_1\}\cup\mathcal{P}_k, and the capacity test for each candidate is O(1)O(1). Performance, however, does degrade at larger scales, since the density of attention shifts away from the training distribution. This is a common limitation of heavy-encoder light-decoder architectures. A light-encoder heavy-decoder design can mitigate it, but its rollout efficiency is low and training is considerably more expensive. Enhanced reinforcement learning and self-improving learning have been explored for this purpose, yet an efficient training method for light-encoder heavy-decoder models on combinatorial optimization problems with complex constraints is still missing. Cross-scale generalization is stated as future work in the Conclusion on page 10.

Reviewer_x1sMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the RT-NVR framework, which extends neural vehicle routing from static optimization to real-time scenarios with delivery commitments. On synthetic instances, it outperforms heuristic baselines in travel cost, fleet size, and delay. The authors also analyze the state-broadcasting weights and the evolution trajectory of the model parameters.

Critical strengths
  1. This paper formalizes the constraint in real-time Vehicle Routing Problems (VRP)—where vehicles are already in transit, and their loaded cargo cannot be reassigned—as VRP-FSDC. To address this, the proposed DCSM prunes the candidate set by leveraging the peak load and maximum slack time of the delivery commitment set; this eliminates the need to repeatedly re-evaluate the entire subsequent route for every candidate. Furthermore, by integrating an execution-planning event-driven framework, the approach enables a neural model to perform global planning while remaining within safe operational boundaries.
Critical weaknesses
  1. The candidate set Ck={b1}∪Pk\mathcal{C}_k = \{b_1\} \cup \mathcal{P}_k only permits visiting the next committed delivery b1b_1 or a feasible pickup customer, while the delivery commitment set Bk\mathcal{B}_k is derived from the initial static schedule produced by PyVRP. Consequently, the \emph{order} of committed deliveries is frozen throughout the entire operation, and the neural policy effectively only decides which pickups to insert, when to return to b1b_1, and whether to dispatch a new vehicle. However, the VRP-FSDC definition only requires that the \emph{vehicle assignment} of committed deliveries remain unchanged; it does not require their visiting order to be fixed. The proposed masking is therefore strictly more restrictive than the problem itself, and the achievable solution quality is bounded by the delivery sequence inherited from the initial static plan. This is also at odds with the claim of far-sighted global planning under the safe decision boundary. The authors should quantify the cost of this restriction, for example by comparing against (i) a variant that allows reordering of deliveries within the DCS (while preserving feasibility), and (ii) a strong event-triggered re-optimization baseline (e.g., PyVRP/HGS) operating under the same fixed-start and delivery-commitment constraints.
Questions for authors
  1. Does planning use expected or already-observed travel times? Under which assumptions does the guarantee hold, and can the authors report the empirical constraint-violation rate and decompose the residual delay of DCSM-NVR into travel-time noise versus the planning rule itself?

  2. Since the candidate set Ck={b1}∪Pk\mathcal{C}_k=\{b_1\}\cup\mathcal{P}_k fixes the order of committed deliveries inherited from the initial PyVRP schedule, how much solution quality is lost relative to a variant that allows reordering within the DCS, and relative to an event-triggered PyVRP/HGS re-optimization baseline under the same fixed-start and delivery-commitment constraints?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer x1sM

Thanks for your positive comments.

We appreciate your recognition of the VRP-FSDC formulation and of how DCSM prunes the candidate set, and your comments on the travel-time assumption, the violation metric, the frozen delivery order, and the PyVRP/HGS baseline were very helpful; the first two have been addressed in the revision by clarifying the forecast setting and by reporting the CVR in Table 2 together with the stochastic-travel ablation in Table 3, while the remaining two are addressed in our response below. Thank you again for your careful review.

Responses to Weaknesses:

The VRP-FSDC definition fixes only the vehicle assignment, so the candidate set in DCSM is stricter than the problem itself. The frozen order is a design choice. At each step the model selects one customer from a candidate set. If the model were also allowed to reorder committed deliveries, then before visiting a committed delivery other than the first in BkB_k, the mask would have to decide whether the remaining deliveries in BkB_k can still be served feasibly; this is a TSPTW-type problem and is NP-complete, so an NP-complete test would be embedded in every per-step decision, which cannot be checked exactly within the millisecond budget. The frozen order keeps BkB_k feasible (Appendix A.4.1), and the computational complexity is O(1)O(1) per step. Therefore, fixing the visiting order of a DCS is a compromise for real-time operation.

As for “far-sighted global planning”, it refers to the following behavior: under the constraints of vehicle load peak and maximum slack time, the model judges how inserting a pickup affects all subsequent committed deliveries and plans globally over this chain effect, which can suppress the chain propagation of violations. It does not refer to reordering the DCS.

Finally, PyVRP/HGS has no modeling of fixed start nodes or per-vehicle committed delivery sets, and its search operators do not preserve delivery commitments; running it under our constraints means rewriting these operators, and would no longer give a fair comparison across methods.

Responses to Questions 1:

To clarify the definition and the assumption of travel time, we have added “The forecast travel time is set to 2Dij2D_{ij} and is assumed to be available at planning time, whereas the realized (observed) travel time is τij=max⁡{Dij, N(2Dij, 0.2Dij)}\tau_{ij} = \max\{D_{ij},\, \mathcal{N}(2D_{ij},\, 0.2D_{ij})\}. The safety guarantee of DCSM is conditional on this forecast: it holds when realized travel times do not exceed the forecast values.” in the revision. Please see the “Real-Time Instance Generation” paragraph in Section 5.1 on page 6 of the revised manuscript.

Following your suggestions, we have reported the empirical constraint-violation rate (CVR) and decomposed the residual delay of DCSM-NVR into travel-time noise versus the planning rule itself. Details are as follows.

We have reported the CVR row in Table 2 and added “CVR denotes the average constraint-violation rate, i.e., the fraction of served customers whose time window is violated (other constraints are strictly satisfied during the construction process); all metrics are averaged over instances and event-frequency combinations.” in the revision. Please see the “Real-time performance comparison with heuristic methods” paragraph in Section 5.2 on page 8 of the revised manuscript.

We have also reported the effect of stochastic time on Delay and CVR in Table 3 and added “Table 3 isolates stochastic travel time, where ''Sto.'' and ''Det.'' denote the stochastic and the deterministic setting (forecast travel time equals realized travel time). Both Delay and CVR vanish under deterministic travel time, indicating that the residual delay and violations stem from travel-time uncertainty and that the planning rule of DCSM-NVR is strictly feasible in the deterministic case.” in the revision. Please see the “Real-time performance comparison with heuristic methods” paragraph in Section 5.2 on page 9 of the revised manuscript.

Responses to Question 2:

In principle, reordering cannot degrade the delivery order, since the frozen order is one of the feasible orders of BkB_k. Quantifying the resulting improvement would require computation time far beyond the real-time budget, because obtaining the best feasible reorder involves solving the TSPTW-type sequencing problem at every step, and any such figure would therefore be an offline bound rather than a value attainable online. The gap is thus measured against an offline optimum, not against a method that can run online. Consequently, the frozen order forgoes the improvement that reordering may offer, and the inherited order remains feasible, since it comes from the PyVRP solution on the same instance.

PyVRP/HGS cannot handle fixed start nodes or per-vehicle committed delivery sets; running it under our constraints means rewriting these operators, and would no longer give a fair comparison across methods.

Thank you to the authors for their response. I have no further questions. However, I am not fully confident in my current score, and I respect the opinions of the other reviewers. Thank you.

AuthorsOpenReview ↗

Reply to Reviewer x1sM

We sincerely thank you for your time and effort in reviewing our manuscript, and we hope that our response has addressed the concerns you raised earlier; we fully respect the score you have given as well as the opinions of the other reviewers.

Reviewer_xAJ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for real-time vehicle routing under delivery commitments. It introduces VCPO for policy training, SCGS for spatiotemporal-guided sampling, and DCSM for feasibility-constrained online replanning.

Critical strengths

The paper addresses a meaningful and practically relevant problem by extending neural vehicle routing to dynamic scenarios with delivery commitments.

Critical weaknesses
  1. The experiments focus primarily on comparisons of training strategies in Table 1 and heuristic methods in Table 2, without evaluating the framework against existing complete neural or reinforcement learning-based solvers for dynamic VRP. Such end-to-end comparisons would better demonstrate the competitiveness of the proposed framework.

  2. DCSM constructs a delivery commitment set and applies vehicle load-peak and maximum time-slack checks, which appear to be common ideas in vehicle routing with limited novelty. The authors should further clarify the specific methodological novelty of DCSM. In Table 2, comparisons with SVI and SVI-II change both the routing strategy and the feasibility-handling mechanism. Therefore, these comparisons do not isolate the contribution of DCSM itself. A comparison with an SVI-based mechanism under the same neural routing model would provide stronger evidence.

Questions for authors

DCSM constructs a delivery commitment set and applies vehicle load-peak and maximum time-slack checks, which appear to be common ideas in vehicle routing with limited novelty. The authors should further clarify the specific methodological novelty of DCSM. In Table 2, comparisons with SVI and SVI-II change both the routing strategy and the feasibility-handling mechanism, while DCSM-NN and DCSM-ILS also use DCSM. Therefore, these comparisons do not isolate the contribution of DCSM itself. A comparison with an SVI-based mechanism under the same neural routing model would provide stronger evidence.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer xAJ8

Thanks for your positive comments.

We appreciate your recognition of the practical significance of real-time vehicle routing in honoring delivery commitments, and your suggestions on the end-to-end comparison and on isolating the contribution of DCSM were very helpful; both have been addressed in the revision by clarifying why existing dynamic VRP solvers cannot be directly adapted, by clarifying that the novelty of DCSM lies in suppressing the cascading propagation of violations, and by adding a component ablation that isolates DCSM, and we hope the revised manuscript conveys the contribution of the framework more clearly. Thank you again for your time and for this thorough and constructive review.

Responses to Critical Weaknesses:

We agree that comparing against existing complete neural or reinforcement learning-based solvers for dynamic VRP would be valuable. To our knowledge, however, the published solvers of this kind re-optimize the whole route from scratch at each decision point and assume neither fixed start nodes nor a per-vehicle committed delivery set; adapting them would require redefining their state representation, action space, and feasibility handling, so the resulting models would no longer be the original methods and the comparison would be hard to attribute to the proposed framework. We have therefore evaluated the framework end-to-end within a controlled real-time setting. Table 2 shows that, under the same real-time framework, DCSM-NVR outperforms other methods because its decisions are based on delivery commitments and globally planned, while the others rely on local insertion or local search. Our goal is to make NVR deployable in real-time operations, not to propose another static VRP architecture.

Responses to Questions:

Following your suggestions, we have further clarified the specific methodological novelty of DCSM and conducted an ablations analysis to isolate the contribution of DCSM. Detailed explanation and revision are as follows.

In real-time operation, feasibility is not a static property: a visit that is feasible at the current moment can still push later committed deliveries out of their capacity or time-window constraints, and once this happens the violation propagates along the remaining route and degrades every later decision. What existing approaches lack is therefore a way to judge a candidate by its effect on the entire remaining commitment rather than on the current step alone. To this end, DCSM derives the vehicle load peak and maximum slack time from the delivery commitment set, which compress this route-level chain effect into two scalar tests applied at every decoding step, suppressing the cascading propagation of violations while keeping the decision boundary provably safe. We have accordingly rewritten the abstract, the contribution, and the planning level of the method:

  1. We have rewritten “During real-time deployment, a Delivery Commitment Set-based Masking (DCSM) mechanism derives the vehicle load peak and maximum slack time to suppress the cascading propagation of violations, establishing a provably safe decision boundary.” in the revision. Please see Abstract on page 1 of the revised manuscript.

  2. We have rewritten “For each vehicle, a delivery commitment set (DCS) is constructed. At each decoding step, the peak load and maximum slack time of that vehicle are derived analytically to mask candidate actions that would otherwise trigger cascading violations.” in the revision. Please see Introduction on page 2 of the revised manuscript.

  3. We have also added “At each step, a candidate is checked not only for its own local feasibility but also for its impact on subsequent operations, since a locally feasible visit may violate the capacity or time-window constraints of later committed deliveries and cause violations to propagate in a chain. A Delivery Commitment Set-based Masking (DCSM) strategy compresses this route-level condition into two scalar tests and uses them to construct the candidate set Ck\mathcal{C}_k for the next visit:” in the revision. Please see the “Planning Level” Paragraph on page 6 of the revised manuscript.

Following your suggestions, we have added “In addition, the ablations of DCSM and SCGS are reported in Table 4, where woDCSM denotes the variant without DCSM, which only masks capacity- or time-window-infeasible pickup customers, capacity-infeasible committed customers, and uncommitted customers, while woSCGS denotes the variant whose sampling is not guided by SCGS. The absence of DCSM leads to a significant increase in Delay and CVR, while the absence of SCGS slightly increases cost and VN. This demonstrates that DCSM is crucial for safe decision-making, while SCGS is beneficial for performance improvement.” in the revision. Please see the “Ablation Analysis” paragraph in Section 5.2 on page 9 of the revised manuscript.

Thank you for the author's detailed response and revisions. I appreciate the additional experiments, but I still have following concerns:

  1. Traditional insertion methods also consider the impact of new visits on downstream constraints. The new ablation only compares DCSM with a simple mask. I would be more interested in combining established insertion-based methods with the same neural routing model. It will be better to distinguish how much of the performance improvement comes from the neural network and how much from DCSM.

  2. I understand that existing learning-based dynamic VRP methods do not address exactly the same problem setting. However, could some representative methods be adapted with relatively minor modifications for meaningful comparisons? Further clarification on this possibility would be helpful.

  3. In Table 1, VCPO becomes less competitive as fif_i increases. Although the section provides a possible explanation, I would appreciate a more detailed analysis of why VCPO is more affected than the other training methods.

AuthorsOpenReview ↗

Reply to Reviewer xAJ8

Thanks for your comments.

  1. Insertion and the neural autoregressive constructor are two distinct construction paradigms rather than two components that can be combined, so combining them within one model does not yield a controlled comparison; the ablation instead holds the network, the training signal and the decoder fixed and varies only the masking rule, which is what attributes the reported gap to DCSM. If the intent is to bring the feasibility check of insertion-based methods into the decoder, then under the frozen delivery order that check and DCSM define the same feasible action set, the load-peak and slack-time conditions being necessary and sufficient.

  2. Methods of this family assume that vehicles start from the depot with no delivery commitment set yet formed, whereas in our problem each vehicle is already in transit with its own real-time state and delivery commitment set, and handling this state and these commitments, through real-time state conditioning and the DCSM feasibility boundary, is this paper's core contribution. Supplying that contribution to another architecture would place our solution inside their model, so the difference could not be credited to the original method, which makes such a comparison unfair to those methods and uninformative for ours.

  3. The data do not support the conclusion that VCPO is more affected than the other training methods. At fi≥0.4f_{\mathrm{i}}\ge 0.4 VCPO wins 2 of the 10 combinations and POMO 1, so it does not fall behind them; what disappears at this rate is its margin over ITAB and ITDA. The reason given in Section 5.2 is that greedy replanning may restrict future decision flexibility, a cost incurred at each replanning event rather than one attached to VCPO's training signal.

Thank you for the detailed clarifications and additional experiments.

  1. My original concern was that DCSM uses hard constraints based on vehicle load peaks and time slack, which appear to be well-established in classical insertion methods. My suggestion for further comparisons was intended to better assess DCSM's independent contribution beyond the neural routing model. I appreciate that the new ablation demonstrates DCSM's effectiveness over a local mask, and I understand that insertion-based and neural autoregressive methods may not be directly comparable. However, given the authors' clarification that classical insertion feasibility checks and DCSM yield the same feasible action set under a fixed delivery order, this seems to reinforce my original concern about its methodological novelty to some extent. Could DCSM be viewed primarily as an adaptation of established constraint-handling techniques to neural decoding? I recognize the practical value of such an adaptation, but since DCSM is presented as one of the paper's core contributions, I believe this distinction is worth clarifying.

  2. Regarding VCPO, my question was mainly about its diminishing advantage over ITDA and ITAB, rather than POMO. In Table 1, VCPO achieves the best results in 13 of 15 settings at low (f_i), but only 2 of 10 at high (f_i), where ITDA achieves the best results in 5 settings. I acknowledge VCPO's overall effectiveness and would appreciate further insight into why its relative advantage appears to diminish as (f_i) increases.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.