- 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
-
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.
-
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
-
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?
-
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?
-
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
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.
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+2, 6=2+2+2, 8=2+2+2+2, and 10=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.