acceptodds
Under review as a conference paper at ICLR 2027

PursuitVLA: An Executable Vision–Language–Action Policy for Cross-Map Zero-Shot Pursuit–Evasion Games

Abstract

Pursuit–evasion games require pursuers to coordinate over long adversarial interactions and complex map geometry while producing executable actions. Existing graph-based methods depend on map-specific representations, while vision–language agents are predominantly designed for single-agent tasks, making coordinated and executable multi-agent action generation difficult. We introduce PursuitVLA, an autoregressive vision–language–action policy that generates reasoning traces and executable waypoint tokens from top-down observations and multi-agent states, conditioning each pursuer on previously generated teammate actions. The same role-conditioned policy also operates as an evader. Training first uses dynamic-programming (DP) demonstrations with strategic reasoning and then applies reinforcement fine-tuning with Sequential Action Compensation (SeqAC), which scores candidate actions at intermediate states induced by policy-generated teammate prefixes, allowing later agents to compensate for deviations in earlier teammate actions. Without retraining or map-specific graph reconstruction, PursuitVLA achieves 79.6% capture against DP evaders on eight unseen maps, compared with 63.0% for EPG, the strongest graph-learning baseline. It remains effective under local map perturbations and transfers from 2v1 training to simulated 10v1 pursuit. Without additional training, the same policy controls two pursuers in a closed-loop outdoor deployment, achieving an 80% capture rate (16/20 trials) against a DP-controlled evader, compared with 50% (10/20 trials) for EPG. Our project is available at: https://sites.google.com/view/pursuitagent

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
multi-robot pursuit–evasion, vision–language–action agents

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_txsmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies the zero-shot generalization ability in pursuit-evasion games. To this end, the authors propose PursuitVLA, an autoregressive vision-language-action policy for pursuit-evasion games that generates reasoning traces and executable waypoint tokens from top-down observations and multi-agent states. It includes two training phases. First, it conducts supervised fine-tuning (SFT) using expert demonstrations with strategic reasoning built from dynamic-programming equilibrium trajectories. Then, reinforcement fine-tuning (RFT) applies GRPO with Sequential Action Compensation (SeqAC). Finally, the experimental results show that on eight unseen unseen maps against DP evaders, PursuitVLA outperforms other baselines.

Critical strengths
  1. The motivation of this paper is well stated, and applying VLA in pursuit-evasion games is novel and natural. The method removes the need to construct a map-specific graph and recompute the policy for every new map. Moreover, the closed-loop outdoor deployment on three quadruped robots, without additional training, shows that the approach is practical beyond simulation.

  2. The whole structure of PursuitVLA is complete, and the two training components are well motivated. SFT on expert demonstrations provides a good initialization, and RFT further improves the generalization ability with SeqAC. These ablation study results show that each training component contributes, and SeqAC offers a reusable idea for training other sequentially decoded multi-agent policies.

Critical weaknesses

This training process depends on expert demonstrations which is obtained from DP method and the rewards for RFT also rely on cached DP value functions. Therefore, the approach does not remove the need for a tractable solver at training time, which may affect scalability. Moreover, the main evaluation is conducted against the DP evader or its own evader instead of other adaptive learned evaders or other diverse evaders. It therefore remains unclear how robust PursuitVLA is against stronger or more diverse opponents.

Questions for authors
  1. In some pursuit-evasion game papers, these pursuers commonly make decisions simultaneously and independently. In this paper, these pursuers make decisions sequentially, although actions are executed simultaneously. So each later pursuer makes decisions based on the earlier pursuer’s action, which means that the later pursuer uses more information when making a decision compared with the simultaneous setting. Does this information help coordination, i.e., how much of PursuitVLA’s advantage comes from this centralized coordination? And how would the method apply in settings with limited communication or decentralized execution?

  2. In the main robustness experiment on local map perturbations, why are DP and EPG kept on the graph built from the original map, while PursuitVLA receives the edited map? Is this a fair comparison? I noticed that Appendix B.4 reports results with graphs rebuilt from the perturbed maps, and would the authors consider these results as the main robustness result?

  3. I noticed that both the demonstrations and the RFT rewards rely on DP, and training is limited to 2v1. In the 4v1–10v1 transfer experiments, which evader policy is used, and how is it obtained? How would the training pipeline scale to larger maps or more agents, where DP becomes expensive?

3 comments

Q1. Does sequential teammate-action conditioning improve coordination, and how could the method extend to limited-communication or decentralized execution?

A1. Our method follows the centralized full-state formulation used by EPG [Lu et al., 2025] and introduces autoregressive joint-action generation as the coordination mechanism. The policy observes the global map and multi-agent state, generates pursuer actions sequentially, and executes the completed joint action synchronously. The one-shot ablation uses the same inputs, initialization, action vocabulary, and evaluation protocol, but removes conditioning on previously generated teammate actions. On the eight unseen maps, this reduces capture from 71.8% to 51.6% and increases mean capture steps from 139.0 to 165.0, showing a 20.2-point coordination gain from sequential teammate-action conditioning.

The current paper focuses on centralized coordination with global state access. Under limited communication, the full teammate-action prefix could be replaced by bandwidth-limited messages or learned summaries. Fully decentralized execution would require local-history- or belief-conditioned policies trained through centralized training or policy distillation. These settings impose different information constraints and are left for future work.


Q2. In the main robustness experiment on local map perturbations, why are DP and EPG kept on the graph built from the original map, while PursuitVLA receives the edited map? Is this a fair comparison? Would the authors consider the rebuilt-graph results in Appendix B.4 as the main robustness result?

A2. We agree that the rebuilt-graph protocol in Appendix B.4 is the more appropriate primary comparison for robustness under equally updated map representations. The original experiment was designed to test a different capability: whether PursuitVLA can adapt directly to an edited visual map without rebuilding a map-specific graph or value function. Under this no-reconstruction protocol, PursuitVLA receives the perturbed map, while graph-based methods retain their original representations.

In the complementary updated-representation protocol, the graph-based baselines and the DP evader are recomputed from the perturbed maps, while PursuitVLA receives the same edited visual maps without policy retraining. PursuitVLA achieves 85.9% capture, compared with 60.6% for EPG, 28.7% for MAPPO, and 100.0% for the recomputed DP oracle. We will move this updated-representation comparison to the main robustness discussion and clearly label the original experiment as a separate no-reconstruction evaluation.

The two protocols measure complementary properties: the rebuilt-graph protocol provides the fair comparison of policy performance when map representations are updated, whereas the no-reconstruction protocol measures the practical advantage of adapting directly from the raw edited map without map-specific graph reconstruction.


Q3. In the 4v1–10v1 transfer experiments, which evader policy is used, and how is it obtained?

A3. The 4v1–10v1 experiments use a scalable grouped-DP evader. Following the grouping extension of EPG [Lu et al., 2025], the larger teams are constructed from pairwise groups: 4=2+24=2+2, 6=2+2+26=2+2+2, 8=2+2+2+28=2+2+2+2, and 10=2+2+2+2+210=2+2+2+2+2. Each group uses an exact 2-pursuer DP policy. The evader evaluates the possible groupings using the minimax grouping rule and selects the corresponding worst-case action. The same grouped-DP evader is used when evaluating the 2v1-trained SFT and RFT policies at larger team sizes.

This procedure avoids constructing an exact full joint-state DP table for 4v1–10v1 while providing a scalable DP-based opponent for the transfer experiments.


Q4. How would the training pipeline scale to larger maps or more agents, where DP becomes expensive?

A4. The main scaling challenge is obtaining a dense, state-dependent cost-to-go signal for long-horizon coordination. DP currently provides this signal offline for demonstrations and SeqAC reward construction. For more agents, the grouped-DP construction above provides a scalable approximate supervision signal without enumerating the full joint-state DP space. For larger maps, the same principle could be extended through local or hierarchical graph decomposition and coarse-to-fine DP.

When no tractable solver is available, the SeqAC/RFT framework could instead use a learned critic trained from rollout or self-play data to approximate the DP cost-to-go and provide dense credit for long-horizon action sequences. Developing and validating such a solver-free critic is an important direction for future work.

Summary

Thank you for the positive assessment and constructive questions. We especially appreciate your recognition that “applying VLA in pursuit-evasion games is novel and natural,” that “the whole structure of PursuitVLA is complete,” and that “SeqAC offers a reusable idea for training other sequentially decoded multi-agent policies.”

Across the reviews, PursuitVLA was recognized for:

  • Novel contribution: (txsm: applying VLA to pursuit-evasion games is “novel and natural”; SeqAC “offers a reusable idea for training other sequentially decoded multi-agent policies”).
  • Empirical evidence: (Tvw6: “strong main result on unseen maps under matched comparisons” and “a simple method validated in a closed loop on real robots”), (Wj3N: comparisons include “a broad range of baselines” and generalization is evaluated “along several axes”).
  • Clear presentation: (Tvw6 and Wj3N: “Clear and easy to follow”).

To address your concerns, we clarified the centralized coordination protocol, added the rebuilt-graph robustness comparison, specified the grouped-DP evader for larger teams, and separated solver-free inference from solver-free training. We hope these clarifications address the remaining questions and make the method’s scope and limitations clearer.

Thanks for the detailed responses. The responses address most of my concerns, and I keep my original positive score.

Reviewer_Wj3NOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PursuitVLA, an autoregressive vision-language-action (VLA) policy for pursuit-evasion games. The policy is trained with SFT followed by reinforcement fine-tuning. To address the teammate-action mismatch between SFT and inference, the authors introduce SeqAC in the reinforcement finetuning stage to train later pursuers to compensate for earlier teammates' deviations. The paper reports generalization to unseen maps and larger team sizes, as well as transfer to an outdoor robot deployment without additional training.

Critical strengths
  • The method is compared with a broad range of baselines, including geometric planners, graph-based DP and RL methods, multi-agent RL, and prompted frontier VLMs.
  • Generalization is evaluated along several axes. The paper reports transfer to unseen maps and larger team sizes, as well as robustness to local map perturbations.
  • The simulation-trained policy is deployed on real outdoor robots without additional training, and the paper honestly reports failure cases.
Critical weaknesses
  • The ablation studies appear to be conducted only on training maps. How each component contributes to cross-map generalization is unclear.
  • The team-size transfer experiment does not have baseline comparisons. It is hard to tell whether the improvement comes from better coordination or simply having more pursuers.
  • The method relies on a dynamic-programming solver for both training data and rewards, which may limit its applicability to settings where such solvers are unavailable.
Questions for authors
  • Could the authors report the ablation results on unseen maps to show how each component contributes to generalization?
  • How do the baselines perform when transferred to larger team sizes, and what are the corresponding capture rates?
  • How might the approach be extended to settings where a DP solver is intractable, such as larger teams or partial observability?
3 comments

Q1. Could the authors report the ablation results on unseen maps to show how each component contributes to generalization?

A1. The component gains persist on unseen maps, including an 11.6-point capture-rate improvement from SeqAC. We kept the ablation settings unchanged and evaluated every variant on the eight unseen maps using the same DP evader, matched initial conditions, execution stack, and 200-step horizon; uncaptured trials are assigned 200 steps.

VariantCapture rate ↑Capture steps ↓
SFT design
SFT w/o reasoning40.8%179.5
SFT one-shot joint generation51.6%165.0
SFT full71.8%139.0
Imitation-learning variants
Expert-only; expert action64.5%155.0
All prefixes; fixed expert action57.5%171.0
All prefixes; conditioned action69.2%149.0
RFT design
RFT w/o SeqAC68.0%157.5
RFT w/o action tokens70.4%135.8
VLM + MLP72.3%143.8
Full RFT / PursuitVLA79.6%120.5

The full RFT model improves over RFT without SeqAC by 11.6 percentage points and 37.0 capture steps. These results show that the component contributions observed on the training maps remain present on unseen maps.


Q2. How do the baselines perform when transferred to larger team sizes, and what are the corresponding capture rates?

A2. The 2v1-trained PursuitVLA policies retain high capture rates when directly transferred to larger teams, with RFT providing the stronger scaling behavior.

In the original manuscript, the team-size experiment compared the 2v1-trained SFT and RFT policies. SFT was chosen as the matched baseline because it shares the same policy architecture, map and state inputs, action vocabulary, 2v1 training data, initial conditions, and team-size evaluation protocol as RFT. The main difference is that SFT does not undergo SeqAC-based reinforcement fine-tuning. This comparison isolates the benefit of learning to coordinate under generated teammate actions, beyond the effect of simply adding pursuers.

To address the reviewer's request for an external baseline, we additionally evaluated Voronoi-Guided under the same team-size protocol. The SFT and RFT policies trained in 2v1 are applied directly to 4v1, 6v1, 8v1, and 10v1 without retraining or changing the policy architecture or action vocabulary; all team sizes use the same 12 training maps and 100 matched initial conditions per map.

Method2v14v16v18v110v1
Voronoi-Guided54.5%82.2%93.4%97.6%99.0%
SFT82.8%91.2%98.7%99.1%99.8%
RFT / PursuitVLA90.5%97.3%99.5%99.4%99.7%

The corresponding mean capture steps are:

Method2v14v16v18v110v1
Voronoi-Guided152.5132.6119.8105.299.7
SFT126.5122.1118.0116.6115.9
RFT / PursuitVLA110.2101.389.980.775.9

All three methods benefit from adding pursuers in terms of capture rate. RFT achieves the highest capture rate from 2v1 through 8v1, while SFT is marginally higher at 10v1; Voronoi-Guided also approaches near-complete capture as the team grows.

The efficiency trends are different. RFT has the fewest capture steps at every team size and continues to improve as more pursuers are added. SFT and Voronoi-Guided also become more successful with larger teams, but their capture-step reductions are smaller. Thus, adding pursuers improves the eventual capture probability for all methods, while RFT makes the most effective use of them to accelerate capture.

Q3. How might the approach be extended to settings where a DP solver is intractable, such as larger teams or partial observability?

A3. The current training pipeline relies on a tractable DP solver, but this dependence is confined to offline supervision and reward construction; the deployed PursuitVLA policy does not query DP at inference time. DP is used to generate demonstrations and cache training-time value signals on the training maps. At deployment, PursuitVLA maps the current visual map and agent states directly to executable waypoint actions without running DP, reconstructing a map-specific graph, or querying a value table.

For larger teams, a natural extension is to use the grouped-DP decomposition adopted by EPG [Lu et al., 2025]. Pursuers can be divided into small groups, such as pairwise subteams, and the corresponding local DP problems can be combined through a minimax grouping rule. This replaces the intractable full joint-state DP oracle with a scalable approximate oracle for demonstrations and reward construction. The resulting supervision would not represent an exact full-team equilibrium, but would provide a tractable training signal as the team size increases.

For partial observability, the policy can be extended to condition on observation histories or learned belief states, while the training teacher is replaced by an observation-aware or robust pursuit solver. A perfect-information DP policy could still be used as privileged supervision during training, provided that the student receives only partial observations and the objective accounts for the resulting information mismatch. Recent work, R2PS [ICLR 2026], studies worst-case robust real-time pursuit under partial observability and provides a relevant direction for constructing observation-aware or robust teachers.

When no tractable solver is available even for supervision, DP-based rewards can be replaced by a learned critic, rollout-based value estimates, self-play, or model-based search. These solver-free training extensions are beyond the current experiments and constitute an important direction for future work. Thus, the current paper establishes solver-free inference, while scalable solver-free training remains open.

Reference.

R2PS: Worst-Case Robust Real-Time Pursuit Strategies under Partial Observability. International Conference on Learning Representations (ICLR), 2026.


Summary

Thank you for the positive assessment and constructive questions. We especially appreciate your recognition that PursuitVLA is compared with “a broad range of baselines,” that generalization is evaluated “along several axes,” and that the policy is deployed on “real outdoor robots without additional training.”

Across the reviews, PursuitVLA was recognized for:

  • Novel contribution: (txsm: applying VLA to pursuit-evasion games is “novel and natural”; SeqAC “offers a reusable idea for training other sequentially decoded multi-agent policies”).
  • Empirical evidence: (Tvw6: “strong main result on unseen maps under matched comparisons” and “a simple method validated in a closed loop on real robots”), (Wj3N: “a broad range of baselines” and generalization evaluated “along several axes”).
  • Clear presentation: (Tvw6 and Wj3N: “Clear and easy to follow”).

To address your concerns, we added the unseen-map ablations, the Voronoi-Guided larger-team comparison, and capture-rate results across team sizes. We hope these additions clarify how the components contribute to generalization and how the learned policy benefits from larger teams.

Thank you for the responses. My concerns have been addressed. I will maintain my positive score.

Reviewer_Tvw6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PursuitVLA, a vision-language-action policy for multi-agent pursuit-evasion. Given a global top-down traversability map and agent states, the policy generates reasoning traces and executable waypoints, conditioning later pursuers on actions already generated for earlier teammates. Training combines supervised learning from dynamic-programming demonstrations with reinforcement fine-tuning. Sequential Action Compensation evaluates candidate actions at intermediate states induced by teammate actions, allowing later agents to learn compensatory decisions.

Critical strengths

S1. Strong main result on unseen maps under matched comparisons. All methods share the same continuous simulator and low-level controller, each map uses 100 matched initial conditions, and the unseen maps are excluded from codebook construction, data generation, training and model selection (Sec. 4.1; App. B.2, p. 14). On eight unseen maps, PursuitVLA captures DP evaders in 79.6% of trials (120.5 steps on average) vs. 63.0% (149.1 steps) for EPG, and reaches 84.3% against PursuitVLA evaders (Table 6b, p. 27).

S2. A simple method validated in a closed loop on real robots. The waypoint-token codebook with receding-horizon execution of only the first token (App. G.1, p. 20) makes the VLM output directly executable; without further training, the same policy controls two quadruped robots at 2 Hz on a roughly 120 m × 60 m outdoor site, and failure cases are reported (App. K.2, Fig. 18, p. 26).

Critical weaknesses

W1. The contribution of the components to zero-shot cross-map generalization is not tested on unseen maps

The central claim is zero-shot cross-map generalization (Abstract; contributions in Sec. 1), yet all component ablations (Fig. 4c, Table 7c), the team-size study (Table 7b) and the perturbation study (Table 7a) are run on training maps. For example, full RFT's 90.5%/110.2 steps and SFT's 82.8%/126.5 steps coincide with the training-map numbers in Table 6a (pp. 27–28), and the SFT model itself is never evaluated on unseen maps. At the same time, the RFT reward comes from DP values cached on exactly these training maps (Sec. 3.4, p. 5), and the training-map evaluation opponent is the same DP evader, so the SeqAC gain (80.7% → 90.5%) may partly reflect fitting to the training maps. Fig. 3 also shows that, against the DP evader, PursuitVLA drops by 10.9 points from training to unseen maps (90.5 → 79.6) and EPG by 12.0 points (75.0 → 63.0). In other words, the unseen-map lead largely carries over the training-map lead rather than coming from a smaller generalization gap.

W2. The robustness claim relies on a protocol in which the opposing evader also uses stale map information, and the fairer control in the appendix is not reported in the main text

In the main perturbation study (Sec. 4.3, Table 7a), the evader appears to keep the original map's DP graph and value function. There are three pieces of evidence: App. B.4 says that only "for this evaluation" are methods tested against "the same DP evader recomputed from the perturbed map" (p. 15); App. J.2 says DP intentionally retains the original graph during perturbation evaluation (p. 24); and Fig. 14 shows a stale DP evader entering a modified dead end (p. 19). PursuitVLA's increase from 90.5% to 93.3% after perturbation may therefore partly come from a weakened opponent. When the evader is recomputed and the graph baselines are rebuilt (App. B.4, Table 2, p. 15), PursuitVLA drops by 4.6 points (90.5 → 85.9), close to MAPPO (−3.5) and below EPG (−8.8). The main text, however, reports only the no-rebuild protocol and concludes that graph-based methods "are more sensitive to changes in map connectivity" (p. 8).

Questions for authors
  1. What are the capture rates and capture steps of the SFT model and of each ablation variant—especially RFT w/o SeqAC and SFT w/o reasoning—on the eight unseen maps?
  2. Which evader policy is used in the main perturbation study (Table 7a) and in the team-size experiments (4v1–10v1)? The DP state space is large for 10v1—is a DP evader still used there, and if so, how is it computed?
  3. On which maps and data was EPG trained—the same 12 training maps and DP data as PursuitVLA, or the original released model? The text says PursuitVLA beats MAPPO, the geometric planners and the prompted models "on each unseen map" (p. 7) but does not include EPG—is there any unseen map on which EPG beats PursuitVLA? Could you provide a per-map table?
3 comments

Q1. What are the capture rates and capture steps of the SFT model and of each ablation variant—especially RFT w/o SeqAC and SFT w/o reasoning—on the eight unseen maps?

A1. The component contributions observed on the training maps also hold when the evaluation maps are replaced by the eight unseen maps, with SeqAC providing an 11.6-point capture-rate gain over its matched ablation. We agree that the original training-map ablations did not establish whether the components contribute to the central cross-map result. We therefore kept the seen-map ablation settings unchanged and replaced the evaluation maps with the eight unseen maps. The same DP evader, matched initial conditions, execution stack, and 200-step horizon were used; uncaptured trials are assigned 200 steps.

VariantCapture rate ↑Capture steps ↓
SFT design
SFT w/o reasoning40.8%179.5
SFT one-shot joint generation51.6%165.0
SFT full71.8%139.0
Imitation-learning variants
Expert-only; expert action64.5%155.0
All prefixes; fixed expert action57.5%171.0
All prefixes; conditioned action69.2%149.0
RFT design
RFT w/o SeqAC68.0%157.5
RFT w/o action tokens70.4%135.8
VLM + MLP72.3%143.8
Full RFT / PursuitVLA79.6%120.5

For the controlled RFT comparison, removing SeqAC reduces capture from 79.6% to 68.0% and increases mean capture steps from 120.5 to 157.5. The SFT and imitation-learning rows show the same pattern: removing reasoning, replacing autoregressive teammate conditioning with one-shot generation, or keeping fixed expert actions under deviated prefixes all reduces performance. Full RFT reaches 79.6%/120.5 steps, compared with 71.8%/139.0 for full SFT. These results directly show that the component gains remain on the unseen maps.


Q2. Which evader policy is used in the main perturbation study (Table 7a) and in the team-size experiments (4v1–10v1)? The DP state space is large for 10v1—is a DP evader still used there, and if so, how is it computed?

A2. We use DP-based evaders in both settings, while the 4v1–10v1 experiments use the scalable pairwise grouped-DP procedure based on EPG rather than an exact full joint-state DP solution.

In the main perturbation study, the evader is the DP policy computed on the original map and its original graph representation is retained after the local map edit. PursuitVLA receives the edited raw map and directly adapts its actions without rebuilding a map-specific graph. This protocol is intended to test the proposed no-reconstruction setting. We also report the complementary updated-representation experiment in Appendix B.4, where the DP evader is recomputed on the perturbed map and graph-based baselines receive rebuilt graph representations.

For the team-size experiments, following the grouping extension of EPG [Lu et al., 2025], we use two-pursuer groups to construct the larger teams: 4=2+24=2+2, 6=2+2+26=2+2+2, 8=2+2+2+28=2+2+2+2, and 10=2+2+2+2+210=2+2+2+2+2. Each group uses the exact 2-pursuer DP policy, and the same grouped-DP evader is used when evaluating SFT and RFT. The evader considers the possible grouping results and selects its action according to the minimax grouping rule: it chooses the grouping under which at least one subteam has the largest remaining DP capture value, and follows the corresponding DP evader action. This avoids constructing the exponentially large full joint-state DP table.

Q3. Please clarify EPG's training/checkpoint provenance and provide per-map results on the eight unseen maps, including whether EPG outperforms PursuitVLA on any map.

A3. PursuitVLA outperforms the publicly released EPG checkpoint on all eight unseen maps.

EPG was not retrained on PursuitVLA's 12 training maps or DP supervision, and no adaptation was performed on the unseen maps. The released EPG checkpoint was trained by its authors on 100-node and 500-node graph scales. We evaluate this checkpoint in the same zero-shot cross-map setting as the other baselines, which directly addresses the transfer question in this review.

Both methods were evaluated with the same simulator, matched initial conditions, and execution protocol. EPG was given its required map-specific graph inputs, including the graph-distance matrix and other graph-derived features recomputed for each evaluation map. Thus, the EPG comparison does not omit the structured graph information on which its method relies.

Our benchmark maps contain 1,679--2,738 nodes. We also attempted to retrain EPG using the same 12 training maps and DP supervision, but retraining at this graph scale was computationally infeasible: one epoch took several hours, while convergence required a large number of RL optimization updates, and the attempted run did not reach convergence. We therefore use the released EPG checkpoint as the reproducible zero-shot baseline in our evaluation.

The omission of EPG from the per-map statement in the main text was an editorial oversight, which we will correct.

MapEPG capture rate (%)PursuitVLA capture rate (%)EPG capture stepsPursuitVLA capture steps
Map 166.084.0134.3103.6
Map 269.778.7147.1129.2
Map 362.072.2137.0120.8
Map 464.076.8158.9110.7
Map 567.780.3135.9109.0
Map 651.775.0149.7133.7
Map 763.787.7168.6104.1
Map 859.082.1161.3138.9
Average63.079.6149.1120.5

The per-map comparison shows no reversal on any of the eight maps.


Summary

Thank you for the positive assessment and constructive questions. We especially appreciate your recognition of PursuitVLA’s “strong main result on unseen maps under matched comparisons” and “a simple method validated in a closed loop on real robots.”

Across the reviews, PursuitVLA was recognized for:

  • Novel contribution: (txsm: applying VLA to pursuit-evasion games is “novel and natural”; SeqAC “offers a reusable idea for training other sequentially decoded multi-agent policies”).
  • Empirical evidence: (Tvw6: “strong main result on unseen maps under matched comparisons” and “a simple method validated in a closed loop on real robots”), (Wj3N: comparisons include “a broad range of baselines” and generalization is evaluated “along several axes”).
  • Clear presentation: (Tvw6 and Wj3N: “Clear and easy to follow”).

To address your concerns, we added unseen-map ablations, clarified the perturbation and grouped-DP protocols, and reported the EPG checkpoint provenance and per-map results. We hope these additions address the remaining questions and clarify the scope of our generalization claims.

The authors' well-executed additional experiments and clarifications address my concerns thoroughly, and I am pleased to maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.