acceptodds
Under review as a conference paper at ICLR 2027

PACT: Predictive Action Coordination with Traffic Awareness for Lifelong MAPF

Abstract

Learning-based policies offer scalable decentralized multi-agent path finding (MAPF) but struggle to maintain throughput in crowded lifelong settings. Two information gaps cause this: uncertainty about neighbors’ concurrent decisions and a lack of traffic awareness beyond the local view. Relying solely on local conflict resolution or global traffic guidance fails to simultaneously resolve incompatible actions and prevent bottlenecks. We propose PACT (Predictive Action Coordination with Traffic Awareness) to bridge both gaps. PACT avoids exchanging action proposals by using the Neighbor Action Predictor (NAP) to infer neighboring agents’ decisions from partial observations for local coordination. Complementarily, coarse sector-level occupancy information provides long-range awareness of congestion beyond each agent’s local field of view. This enables agents to coordinate immediate interactions and bypass emerging bottlenecks without peer-to-peer communication. Across diverse MAPF benchmarks, PACT achieves over a 13× macro-average per-map throughput ratio relative to MAPF-GPT in highly congested settings. Furthermore, it reaches over 90% of the throughput of the centralized RHCR-PBS planner. These results demonstrate that combining inferred intentions with lightweight global traffic awareness substantially narrows the performance gap between learned policies and centralized planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-Agent Path Finding, Multi-Agent Learning, Imitation Learning, Decentralized Coordination

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Rbx6OpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The article concerns the application of learning-based policies to multi-agent path-finding (MAPF). It proposes PACT (Predictive Action Coordination with Traffic Awareness) to bridge 2 gaps: uncertainty about neighbors’ concurrent decisions and a lack of traffic awareness beyond the local view.

PACT avoids exchanging action proposals by using the Neighbor Action Predictor (NAP) to infer neighboring agents’ decisions from partial observations for local coordination. Complementarily, coarse sector-level occupancy information provides long-range awareness of congestion beyond each agent’s local field of view. This enables agents to coordinate immediate interactions and bypass emerging bottlenecks without peer-to-peer communication. Across diverse MAPF benchmarks, PACT achieves over a 13× macro-average per-map throughput ratio relative to MAPF-GPT in highly congested settings. Furthermore, it reaches over 90% of the throughput of the centralized RHCR-PBS planner.

Critical strengths

The paper is well-written, and the results are also very good against strong baselines. I also appreciate that the authors conducted a very good ablation study and that the article is followed by a comprehensive appendix.

Critical weaknesses

I did not identify significant weaknesses.

Questions for authors

I do not have any specific questions to authors.

1 comment

Section, figure, and numbered-table references follow the revised manuscript. Lettered tables and bibliography numbers belong to this response.


We thank the reviewer for the positive assessment of the paper’s clarity, baselines, ablations, and appendix.

Correction. The >13× ratio cited in the summary incorrectly described a Warehouse-specific gain as a five-map macro-average; the corrected values are 14.36× on Warehouse and 5.80× across maps. The revision now reports 90.5% of centralized RHCR-PBS throughput as the headline result (Table 5) and clarifies that Traffic Guidance uses a shared sector-occupancy service.

Main revisions. Following other reviewers, we strengthened the evidence for the main claims:

  • Matched ablations show that NAP-based coordination contributes beyond policy adaptation and TG, and PACT exceeds CS-PIBT-shielded variants without global agent-level information (Fig. 6; Table 7).
  • With added standalone PIBT and PRIMAL2 baselines, PACT achieves the highest throughput among decentralized and communication-based methods on all five maps (Table 5).
  • Median per-robot decision latency is about 3 ms, whereas RHCR-PBS fails at 1024 agents (Figs. 1 and 11).
Reviewer_UCt2OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PACT (Predictive Action Coordination with Traffic Awareness) to address uncertainty about neighboring agents’ actions and the lack of visibility into distant congestion in lifelong multi-agent path finding. PACT combines neighbor action prediction, confidence-aware selective recovery, and traffic guidance to support complementary local coordination and congestion-aware routing. The method reuses the pretrained policy’s fixed-length input interface, and extensive component ablations and mechanism analyses examine the contributions of its modules. Experiments across five benchmark environments—Warehouse, Random, Maze, Empty Room, and Berlin—demonstrate substantial throughput improvements and reduced waiting compared with MAPF-GPT and its CS-PIBT-enhanced variants. Furthermore, within the same framework, neighbor action prediction yields 4.0%–22.7% higher throughput than greedy action estimates, while low-precision traffic summaries maintain performance close to that achieved with full-precision information. These findings further support the method’s effectiveness and potential practical value.

Critical strengths
  1. Clear motivation and complementary design: The paper addresses local action uncertainty and distant congestion in dense lifelong multi-agent path finding through complementary prediction-guided coordination and traffic guidance.
  2. Targeted recovery and interface reuse: Confidence- and priority-guided selective recovery revises potentially conflicting actions only when necessary, while preserving the pretrained policy’s fixed-length input interface.
  3. Substantial experimental analysis: Experiments cover diverse map structures and include component ablations, Greedy/NAP/oracle comparisons, and conflict-rate analyses to examine the proposed mechanisms.
  4. Evidence of effectiveness and practical potential: NAP improves throughput by 4.0%–22.7% over greedy action estimates within the same framework. Low-precision traffic summaries also maintain performance close to that achieved with full-precision information.
Critical weaknesses
  1. Inconsistencies in the headline results: The abstract’s claim of a macro-averaged throughput ratio exceeding 13× over MAPF-GPT is difficult to verify from Figure 5. The Warehouse improvement ratios reported in the main text and Figure 1 also differ.
  2. Unclear information-access assumptions: The method relies on shared global sector-occupancy information. Its decentralization assumptions and the fairness of information access across competing methods require further clarification.
  3. Incomplete attribution of performance gains: The existing ablations do not fully disentangle the contributions of additional policy adaptation, neighbor prediction, and recovery.
  4. Computational overhead at scale: With 1,024 agents, runtime reaches approximately 0.86–1.27 seconds per environment step. Traffic Guidance accounts for the largest share of computation, potentially limiting real-time applicability.
Questions for authors
  1. Calculation of the headline result: Please provide the per-map throughput values at the highest-density settings across the five benchmark environments and explain how the macro-averaged throughput ratio is calculated. How do these results yield the claimed ratio of over 13×? Why does the main text report 13.92× for Warehouse, whereas Figure 1 reports 14.36×?
  2. Attribution of performance gains: Could you include a control using the same training data, LoRA configuration, and training budget but without NAP, to distinguish the benefits of additional policy adaptation from those of neighbor prediction itself?
  3. Information access and communication costs: Do the baselines receive equivalent global occupancy information? Are the costs of collecting and distributing occupancy summaries through the shared service included in the reported runtime? How would delayed, missing, or noisy occupancy information affect performance?
  4. Scalability and real-time performance: With 1,024 agents, Traffic Guidance accounts for approximately 50%–61% of total runtime. Could you report performance under a fixed decision-time budget and discuss ways to reduce the cost of computing map-wide cost-to-go fields for individual agents?
  5. Long-term reliability: Beyond aggregate throughput, could you report per-agent task completion counts, maximum consecutive waiting times, and task-completion-time distributions to assess whether high throughput comes at the expense of prolonged blocking for a subset of agents?
  6. Differences between predicted and executed actions: NAP predicts neighbors’ initial actions, but their subsequent recovery steps may change their final actions. How often do these changes invalidate predictions, and what impact do they have on conflict rates and throughput?
8 comments

Section, figure, and numbered-table references follow the revised manuscript. Lettered tables and bibliography numbers belong to this response.


[W1, Q1] Correction of Headline Results

We are grateful to the reviewer for identifying these inconsistencies and sincerely apologize for the errors. (1) Warehouse ratio. The earlier 13.92× value incorrectly referenced a previous result. The correct ratio is 14.36× (5.485 / 0.382). We now use it consistently in Sec. 1 and Fig. 1.

(2) Macro-average. The Warehouse-specific gain was incorrectly labeled a five-map macro-average. The correct mean of per-map ratios over MAPF-GPT is 5.80×, not over 13×. We corrected the Abstract, Sec. 1, and Sec. 5.2 and now emphasize 90.5% of centralized RHCR-PBS throughput across the five highest-density settings.

Table A: Throughput at the highest densities (means; full results with 95% CIs in Table 5, Appendix B.1)

BenchmarkMAPF-GPTPACTRHCR-PBSPACT / MAPF-GPTPACT / RHCR-PBS
Warehouse (192)0.3825.4855.50214.36×0.997
Random (64)2.0514.1074.7012.00×0.874
Maze (64)0.3892.4903.0656.40×0.813
Empty Room (64)0.7472.6263.0123.52×0.872
Berlin (256)1.9735.3665.5372.72×0.969
Macro-average———5.80×0.905

Each macro-average gives equal weight to the five maps: RB=15∑mTP‾PACT,m/TP‾B,mR_B = \frac{1}{5}\sum_{m} \overline{\mathrm{TP}}_{\mathrm{PACT},m} / \overline{\mathrm{TP}}_{B,m}, rather than dividing pooled throughputs. Reported relative-TP values may differ slightly from ratios of rounded means. Throughput columns are not averaged because maps differ in scale.

[A for W1, Q1] The corrected headline is 90.5% of RHCR-PBS throughput; the five-map ratio over MAPF-GPT is 5.80×.

[W2, Q3] Information Access and Communication Cost

We thank the reviewer for raising the important issues of information-access fairness and communication cost, which led us to clarify PACT's decentralization assumptions and to evaluate its robustness to delayed, dropped, and noisy occupancy information.

(1) Decentralization and information access. PACT selects actions without exchanging action proposals or accessing jointly collected agent-level action preferences. The static map and coordination keys are public. Its only shared nonlocal dynamic-state information is anonymous sector occupancy, as clarified in the Abstract and Appendix A.1; action selection is decentralized, but occupancy collection is shared.

Table B: Runtime information requirements

MethodLocal FoVShared Sector SummaryGlobal Agent-Level Information
PACT (Ours)OOX
MAPF-GPT / Follower / PRIMAL2OXX
MAPF-GPT + CS-PIBTOXO
MAPF-GPT + CS-PIBT + TGOOO
SILLM-DGOXO
RHCR-PBS / LLLG / PIBTXXO

O marks an explicitly used interface; X means that interface is not required. Global agent-level information denotes nonlocal individual states or jointly available action preferences. In the evaluated implementations, this information supports centralized planning or shielding; it does not imply that every possible implementation must be centralized.

Information access is not identical across all baselines. They retain their original interfaces. MAPF-GPT + CS-PIBT + TG receives the same sector summaries as PACT and additionally uses joint preferences for shielding. The matched Policy + TG control further distinguishes predictive coordination from the benefit of shared traffic information.

(2) Communication overhead. Runtime includes TG computation but excludes occupancy collection and network communication. We estimated aggregate payload using measured refresh frequencies (Table 12, Appendix B.7). Uplink is each agent's sector-transition report to the shared service; downlink is the sector-count snapshot returned to an agent upon refresh. Neither contains action proposals.

Table C: Estimated TG communication payload (Table 12, Appendix B.7)

BenchmarkEncodingEvents/StepUplink (kb/step)Downlink (kb/step)Total (kb/step)
Warehouse (192)Original (5-bit)40.760.6513.0413.70
2-bit40.550.655.195.84
Maze (64)Original (4-bit)21.960.355.625.97
2-bit22.040.352.823.17

Payload is summed across agents under the stated refresh model, not measured network airtime. kb denotes kilobits. Totals exclude protocol overhead and retransmissions; rounding may affect displayed sums.

Two-bit summaries reduce estimated payload by 57.4% and 46.9%, respectively, with comparable relative throughput (0.907 vs. 0.905; Table 10, Appendix B.7). End-to-end communication latency and service processing remain unmeasured.

(3) Imperfect traffic information. We tested delayed snapshots (2× the mean refresh interval), dropped updates (50%), and noisy occupancy (25% agent misassignment; Table 13, Appendix B.7).

Table D: Throughput under degraded TG snapshots

BenchmarkOriginalDelayDropNoise
Warehouse5.4855.2305.4585.266
Random4.1074.1434.1494.122
Maze2.4902.4712.5012.494
Empty Room2.6262.5982.6302.624
Berlin5.3665.3635.3675.357

Mean throughput; 95% confidence intervals are reported in Appendix B.7.

Mean throughput remains at least 95.3% of the original across all 15 degraded-information settings. The largest reduction is 4.65% under Warehouse delay. These results support robustness to the tested perturbations, not arbitrary network failures or service outages.

[A for W2, Q3] PACT uses shared aggregate occupancy rather than global agent-level action coordination. Quantization reduces estimated payload, and mean throughput retains at least 95.3% under the tested delays, losses, and noise.

[W3, Q2] Disentangling Policy Adaptation and NAP

We appreciate the reviewer's suggestion of a matched control without NAP; this experiment allowed us to separate the benefit of additional policy adaptation from that of predictive coordination.

(1) Matched ablation. We used the same policy-adaptation data, LoRA configuration, and training budget to train Policy without NAP or TG, then evaluated NAP and TG under matched conditions (Table 7, Appendix B.3; summary in Fig. 6, Sec. 5.3). This matches policy adaptation, not the separate predictor-distillation workload.

Table E: Matched component ablation (Throughput ↑)

BenchmarkPolicyPolicy + NAPPolicy + TGOurs
Warehouse (192)0.5250.9442.1685.485
Random (64)1.8443.7433.6504.107
Maze (64)0.5251.7442.3192.490
Empty Room (64)0.6012.2702.4532.626
Berlin (256)2.1995.3135.1585.366
Warehouse 20-40-10-2-1 (1024)4.6254.9674.8485.038
den520d (1024)2.8964.4724.5885.006

Mean throughput, wait ratios, runtime, and CS-PIBT comparisons are provided in Table 7 (Appendix B.3).

Adaptation alone gives limited improvement: macro-averaged relative throughput is 0.251 versus 0.247 for MAPF-GPT (Fig. 6). Policy + NAP exceeds Policy in all seven settings, and full PACT exceeds Policy + TG, including 2.53× on Warehouse (192). The latter difference includes predictive coordination and selective recovery; it is not the predictor's isolated effect. The recovery ablation is reported in Fig. 12.

(2) Learned predictor. Table 1 (Sec. 5.3) holds the coordination framework fixed while replacing Greedy, NAP, and oracle information. In this dedicated comparison, NAP improves throughput by 4.0–22.7% over Greedy and reaches 97.6–99.8% of oracle throughput, supporting the value of learned action estimates.

[A for W3, Q2] Matched ablations show gains beyond adaptation and TG; Greedy/NAP/oracle comparisons assess the learned predictor separately.

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

The paper targets lifelong multi-agent path finding with learned decentralized policies, which lose throughput in crowded settings. It identifies two information gaps: an agent cannot see what its neighbors are about to do, and it cannot see congestion beyond its local window. PACT addresses both on top of a frozen MAPF-GPT backbone. A Neighbor Action Predictor reconstructs a higher-priority neighbor’s observation from the ego agent’s own view and predicts that neighbor’s initial action, which then conditions the ego decision and, when confident, removes conflicting actions from the admissible set. Separately, sector-level occupancy counts are converted into a congestion-aware cost-to-go field that replaces the policy’s existing routing tokens. Three LoRA adaptation stages, supervised by CS-PIBT, fit the policy to these inputs. Experiments span five standard-scale and two large-scale maps, with ablations, sensitivity sweeps, and a runtime breakdown.

Critical strengths

The oracle study cleanly bounds prediction error. Substituting true neighbor actions for predicted ones changes throughput very little, which tells the reader how much headroom the prediction component has. Results like this are informative whichever way they come out.

The conflict-reduction measurement is the strongest evidence in the paper. Conflicts in final requests drop to roughly one tenth of the backbone’s rate, and collision-handler rejections to about one twentieth. This measures the quantity the method is actually trying to affect, rather than a ratio against a collapsing baseline.

Critical weaknesses
  1. Weak-to-none technical novelty. Every mechanism is adapted from prior work the paper itself cites including agent modeling, priority determination, and guidance graphs for lifelong MAPF. The one genuinely new mechanism is the keyed sampling scheme that aligns an observer’s random draw with the target’s. The paper’s own ablation shows it has no measurable effect: separating the draws, or making them observer-specific, performs the same or marginally better.

Meanwhile the ablation ordering is awkward for the paper’s framing. Coordination tokens alone provide little benefit. Traffic Guidance alone provides a large jump. So the component contributing most is the one closest to existing work, and the component the title focuses on the most contributes least on its own.

Related: selective recovery, offered as a lighter substitute for runtime shielding, is shown to be worse than the thing it replaces. The shielded variant achieves higher relative throughput and lower waiting. The justification would be computational, but recovery occupies a fraction of a percent of per-step runtime, so almost nothing is saved.

  1. The predictor is weakest precisely where it is needed. Top-1 agreement with the true policy is high across all neighbor pairs but falls sharply on contended pairs, meaning the cases where two agents’ intended actions actually conflict. Those are the only cases where prediction changes anything. After confidence gating, a substantial fraction of retained hypotheses are false positives and a substantial fraction of real conflicts are missed. Gating trades recall for precision rather than improving either much.

So the central mechanism degrades on the interaction-critical subset and contributes least in the ablation. Taken together, these suggest the method may be working mainly through congestion-aware routing and policy adaptation, with prediction playing a supporting role.

  1. The runtime argument has some discrepancies. The runtime comparison in fig. 8 is on a single standard-scale map, where PACT reaches high throughput at tens of milliseconds per step against several hundred for the centralized reference. The large-scale breakdown, in the appendix in fig. 10, reports per-step times approaching and exceeding one second at the largest agent count, with the congestion field accounting for the majority. At that large scale, it is not clear how well it performs comapred with the centralied planner but even taken on its, own, >1s is a weak result for a distributed method.

  2. The decentralized comparison is absent, and the obvious runtime-oriented baselines are missing. Of the evaluated methods, four are the same backbone with different wrappers, one of which is the paper’s own adaptation teacher. Only two are independent decentralized baselines, and one of those runs in a different environment version, which the paper notes without addressing.

More importantly, the canonical methods that trade optimality for runtime, priority-inheritance planning and its search-based successors, do not appear as standalone baselines. The shield appears only as a wrapper on the backbone, which is not the same experiment. The paper’s entire positioning is quality at lower cost, and the established methods making that trade are absent.

  1. The paper is positioned narrowly within learned MAPF and does not engage the broader motion-planning literature. The related work covers learned MAPF, runtime shielding, and congestion-aware lifelong planning, and covers them accurately. But spatially constrained multi-agent motion planning is a much larger field, and two neighboring bodies of work are relevant and absent including reactive and optimization-based methods (ORCA, multi-agent MPC etc.), reinforcement learning approaches, etc. Finally, the congestion mechanism is a decentralized form of congestion pricing on a graph, a long-standing idea in network flow and traffic assignment, and the connection is missing.

  2. Smaller points. The main improvement of 13X is a macro-average of per-map ratios at the density where the backbone collapses, so it describes the baseline’s failure mode more than the method’s capability. The relative throughput against the centralized reference is the informative number and should be reported as such.

The framing emphasizes the absence of peer-to-peer communication, but the congestion mechanism requires a global broadcast of per-sector counts. The abstract should be upfront about this. The paper also notes it never measured the cost of collecting, distributing, or refreshing that information.

Questions for authors
  1. Can you run an ablation that isolates learned coordination from classical coordination at matched components, so the gain over backbone plus guidance plus shielding is attributed?
  2. Where is the runtime crossover against the centralized reference, and what do the per-step numbers look like under parallel per-agent execution?
  3. Why are priority-inheritance planning and its search-based successors not standalone baselines, given the paper’s runtime positioning?
  4. What arbitrates residual conflicts outside a simulator with soft collision handling, and what does that cost?
  5. Given that the predictor degrades on contended pairs and contributes least in the ablation, what does the full method achieve with the predictor removed but adaptation and guidance retained?
7 comments

Section, figure, and numbered-table references follow the revised manuscript. Lettered tables and bibliography numbers belong to this response.


[W1] Technical Novelty and Keyed Sampling

We appreciate the reviewer's careful assessment of PACT's novelty relative to prior work and of the role of keyed sampling, which helped us state our contribution more precisely.

(1) Technical contribution. PACT combines observer-constrained neighbor-action prediction with confidence- and priority-guided conflict restriction and selective recovery for decentralized lifelong MAPF. The contribution is the coordination mechanism, not its individual established ingredients. Predictions guide local decisions without exchanging action proposals or performing joint-action search. Traffic Guidance (TG) supplies complementary long-range congestion information, enabling throughput approaching centralized planning without runtime joint-action shielding.

(2) Contributions of NAP and TG. We conducted matched ablations under the same policy-adaptation training conditions to separate predictive coordination and TG from additional adaptation (Fig. 6, Sec. 5.3; Table 6, Appendix B.3).

Table A: Matched ablation at two densities (mean ± 95% CI; Table 6, Appendix B.3)

BenchmarkNPolicyPolicy + NAPPolicy + TGOurs
Warehouse1280.656 ± 0.0711.350 ± 0.1733.504 ± 0.2714.034 ± 0.023
1920.525 ± 0.0420.944 ± 0.0632.168 ± 0.3625.485 ± 0.027
Random322.102 ± 0.1562.420 ± 0.1942.260 ± 0.1562.405 ± 0.191
641.844 ± 0.4683.743 ± 0.3593.650 ± 0.2704.107 ± 0.196
Maze321.469 ± 0.1421.652 ± 0.0591.574 ± 0.0651.660 ± 0.064
640.525 ± 0.1011.744 ± 0.3052.319 ± 0.0892.490 ± 0.094
Empty Room321.408 ± 0.0171.574 ± 0.0131.479 ± 0.0131.566 ± 0.010
640.601 ± 0.1552.270 ± 0.2652.453 ± 0.0342.626 ± 0.024
Berlin1282.613 ± 0.0652.743 ± 0.0552.604 ± 0.0532.745 ± 0.054
2562.199 ± 0.1655.313 ± 0.1725.158 ± 0.0715.366 ± 0.074

Policy + NAP achieves higher mean throughput than Policy in all ten settings and than Policy + TG in six; full PACT exceeds Policy + TG in all ten, including a 2.53× gain on Warehouse (192). Both comparisons include predictive coordination and selective recovery, not the predictor alone; the coordination-token-only variant in Fig. 12 uses NAP predictions only as tokens, without prediction-guided restriction or selective recovery, whereas the complete mechanism raises relative throughput from 0.251 to 0.650 (Fig. 6). In the dedicated predictor comparison, NAP improves throughput by 4.0–22.7% over Greedy and reaches 97.6–99.8% of oracle throughput (Table 1, Sec. 5.3). The components provide complementary benefits.

(3) Keyed sampling and selective recovery. Keyed sampling is a consistency mechanism, not our primary contribution: it aligns samples when distributions and decoding conditions match. Keyed and independent sampling perform comparably (0.905 vs. 0.906), while GT-Independent reaches 0.917 (Table 14, Appendix B.8). Selective recovery avoids joint-action shielding and its jointly available action preferences; its small amortized overhead does not establish a runtime saving over CS-PIBT, as discussed below, but under matched components PACT still exceeds both shielded policies in mean throughput (Table 7), and shielding PACT's initial decisions further improves throughput when global information is available (Fig. 14).

[A for W1] PACT's contribution is observer-constrained prediction integrated with confidence- and priority-guided coordination, supported by matched ablations beyond policy adaptation and TG.

[W2, Q1, Q5] NAP Contribution and Comparison with Runtime Shielding

We thank the reviewer for the incisive analysis of NAP's contribution, its reliability on contended pairs, and its comparison with runtime shielding; this feedback motivated the matched ablations that now attribute the performance gains more clearly.

(1) Contribution beyond adaptation (Q5). Matched ablations cover five standard-scale and two 1024-agent settings (Table 7, Appendix B.3). Policy + NAP exceeds Policy, and full PACT exceeds Policy + TG in mean throughput across all seven settings, including a 2.53× gain on Warehouse (192). This includes selective recovery. Table 1 separately compares Greedy, NAP, and oracle information within the same framework.

Table B: Component-wise performance and runtime (means; Table 7, Appendix B.3)

Standard scaleLarge scale
WarehouseRandomMazeEmpty RoomBerlinWarehouse†den520d
N19264646425610241024
Throughput ↑
Policy0.5251.8440.5250.6012.1994.6252.896
+ NAP0.9443.7431.7442.2705.3134.9674.472
+ TG2.1683.6502.3192.4535.1584.8484.588
+ CS-PIBT3.4583.8082.3612.4925.3084.7714.963
+ TG + CS-PIBT5.3803.8862.4022.5765.2474.8814.990
Ours5.4854.1072.4902.6265.3665.0385.006
Wait ratio (%) ↓
Policy83.5157.1475.1273.0364.544.6039.87
+ NAP74.2919.0230.9816.411.470.7510.55
+ TG59.8811.609.069.192.791.297.44
+ CS-PIBT6.402.943.243.480.520.780.36
+ TG + CS-PIBT2.062.593.033.510.440.630.27
Ours4.774.915.484.821.120.610.62
All-agent runtime (ms) ↓
Policy25.267.977.978.0238.61121.26122.29
+ NAP53.2517.6117.7117.5769.50426.18505.29
+ TG24.687.907.867.8737.92538.65827.43
+ CS-PIBT24.907.978.077.9438.11126.56130.47
+ TG + CS-PIBT24.798.078.077.9738.32547.51899.97
Ours57.5120.3619.9019.5177.39859.061266.76
Single-agent runtime (ms) ↓
Policy0.800.800.800.810.830.820.80
+ NAP2.151.541.631.511.762.937.36
+ TG0.850.840.840.840.871.251.58
Ours2.932.792.782.803.043.517.90

Runtime is reported per environment step for all agents, or per individual decision. CS-PIBT variants shield joint actions globally, so single-agent runtime is undefined for them. † Warehouse 20-40-10-2-1.

(2) Comparison with classical shielding (Q1). The evaluated CS-PIBT implementation is a centralized runtime shield, jointly using agents' action preferences even with decentralized policies. In contrast, PACT's NAP predicts neighboring actions locally without exchanging preferences; TG receives shared aggregate occupancy. PACT achieves higher mean throughput than both matched CS-PIBT variants across all seven settings, although some differences are small and do not establish statistical superiority. Its all-agent runtime is higher, and shielding generally yields lower waiting. Thus, the benefit is competitive throughput without joint shielding, not computational speedup.

(3) Prediction reliability. Contended-pair Top-1 agreement falls to 65.0–70.3% (Table 8). This measures distribution modes, whereas restrictions use confidence-qualified sampled predictions. NAP also conditions initial decisions before conflict screening. Gating trades recall for precision to limit unnecessary restrictions; residual conflicts still require the environment's handler. The complete framework reduces pre-handler conflicts to 1/4.78 after initial decoding and 1/10.32 after recovery, relative to MAPF-GPT (Fig. 15). These reductions do not isolate NAP alone.

[A for W2, Q1, Q5] Predictive coordination contributes beyond adaptation and TG. PACT achieves competitive throughput without the evaluated centralized shield, at higher computational cost and without a collision-free guarantee.

[W3, Q2] Runtime and Scalability

We are grateful to the reviewer for pointing out the gap between the standard-scale and large-scale runtime results, which led us to clarify the runtime crossover with RHCR-PBS and to report single-agent decision latency separately.

The reported 0.86–1.27 s/step at 1024 agents is the total wall-clock time for all agents on one machine, not individual-agent latency. Our implementation processes agents with batched inference to simulate all agents on one machine; in decentralized deployment, each robot computes only its own action, so per-robot latency determines real-time feasibility.

(1) Centralized comparison and crossover. On Warehouse with 32–192 agents, PACT becomes faster than RHCR-PBS between 64 and 96 agents (Fig. 8). At 192 agents, it achieves comparable throughput at approximately 60 ms/step, versus several hundred milliseconds for RHCR-PBS. The two 1024-agent maps also include centralized LLLG (Fig. 9). RHCR-PBS was also run on these maps but failed because of excessive computational cost (Fig. 1), whereas PACT completed all 1024-agent runs.

(2) Single-agent latency. We added individual decision-latency distributions to distinguish local computation from aggregate runtime (Fig. 11, Appendix B.2).

Table C: Single-agent decision latency (ms) (full distribution in Appendix B.2)

BenchmarkMean ± SDMedianP95Max
Warehouse (192)2.93 ± 0.562.813.905.20
Random (64)2.79 ± 0.442.803.395.05
Maze (64)2.78 ± 0.422.793.374.35
Empty Room (64)2.80 ± 0.482.793.535.32
Berlin (256)3.04 ± 1.082.765.108.33
Warehouse 20-40-10-2-1 (1024)3.51 ± 3.202.6112.1926.07
den520d (1024)7.90 ± 9.212.9425.2568.03

Median latency is 2.61–2.94 ms across the evaluated settings, but large-map tail latency increases substantially: den520d reaches 25.25 ms at P95 and 68.03 ms at the observed maximum. TG dominates aggregate runtime. These local measurements are not measured parallel wall-clock times; end-to-end deployment must also account for synchronization, hardware resources, and communication.

[A for W3, Q2] PACT crosses RHCR-PBS runtime between 64 and 96 agents on Warehouse. Median local computation remains around 3 ms, but parallel execution and end-to-end latency have not been measured.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for AC

Dear AC,

Thank you for your time and for overseeing the review of our submission.

The reviewers provided thoughtful, constructive feedback on the evidence supporting our claims and the practical aspects of PACT. Addressing their concerns has strengthened the paper and clarified its contributions. Below, we summarize the reviewers' key observations and our responses to the main concerns across the three reviews.


Strengths

  • A complementary design addressing uncertainty about neighbors' concurrent decisions and limited traffic awareness beyond the local view (UCt2).
  • Mechanism-level evidence from the oracle study and reductions in conflict and collision-handler rejection rates to roughly 1/10 and 1/20 of the backbone's rates, respectively (FuzX, UCt2).
  • Practical design: selective recovery, reuse of the pretrained policy's fixed-length interface, and robustness to low-precision traffic summaries (UCt2).
  • Strong empirical results, extensive ablations, and a comprehensive appendix (UCt2, Rbx6).

Main Concerns and Responses

1. Attribution of Gains (FuzX, UCt2). In matched ablations, additional policy adaptation alone yields a relative throughput of 0.251 (MAPF-GPT: 0.247). Adding coordination via the Neighbor Action Predictor (NAP) or introducing Traffic Guidance (TG) raises it to 0.650 and 0.735, respectively, while full PACT reaches 0.905 (Fig. 6). PACT also exceeds the mean relative throughput of matched CS-PIBT-shielded variants (0.799 and 0.878) without requiring without requiring global per-agent information.

2. Prediction Reliability (FuzX, UCt2). We acknowledge lower Top-1 agreement on contended pairs (65.0-70.3%). Nevertheless, NAP improves throughput by 4.0-22.7% over Greedy estimates and reaches 97.6-99.8% of the initial-action oracle's throughput (Table 1). Recovery invalidates only 0.71-2.23% of initially correct, gated predictions (Table 9).

3. Information Access and Communication Cost (FuzX, UCt2). The Abstract now states that TG uses anonymous sector-occupancy counts from a shared source. Two-bit summaries reduce the estimated payload by 46.9-57.4% while maintaining comparable performance. PACT also retains at least 95.3% of its throughput under delayed, dropped, or noisy snapshots (Tables 10, 12, and 13).

4. Runtime and Scalability (FuzX, UCt2). The reported 0.86-1.27 s per step at 1024 agents measures total wall-clock time for all agents on one machine. Median single-agent decision latency is about 3 ms (Fig. 11).

5. Per-Agent Reliability and Residual Conflicts (UCt2, FuzX). Every PACT agent completes at least one task in the evaluated 512-step runs, and the longest observed waiting streak is 29 steps (Table 15). Residual conflicts are handled by POGEMA's collision handler, as with MAPF-GPT and Follower.

6. Baselines and Related Work (FuzX). We added standalone PIBT and PRIMAL2 baselines. PACT achieves higher throughput and lower waiting than PIBT in all seven settings (Table 5, Fig. 9). We also expanded Sec. 2 to cover ORCA, distributed MPC, decentralized reinforcement learning, and traffic assignment.

7. Headline Metric (FuzX, UCt2, Rbx6). We apologize for incorrectly labeling a Warehouse-specific gain as a five-map macro-average. The corrected gains are 14.36× on Warehouse and a five-map macro-average of 5.80×. We now report 90.5% of centralized RHCR-PBS throughput as the headline result (Table 5).


These revisions and clarifications are reflected in the updated manuscript. We hope this summary and the revised paper facilitate your assessment of our work and its contributions.

Thank you again for your time and consideration.

Best regards,

Authors of Submission 59976

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.