acceptodds
Under review as a conference paper at ICLR 2027

TEMPORAL HYPERGRAPH JOINT- EMBEDDING PREDICTIVE ARCHITECTURE FOR MULTI-AGENT PATH FINDING

Abstract

Multi-Agent Path Finding (MAPF) becomes difficult in dense environments where multiple agents must coordinate through shared aisles, intersections, and bottlenecks. Existing learned methods typically predict actions from the current interaction state, while pairwise representations can lose information about conflicts involving several agents. We introduce TH-JEPA, a Temporal Hypergraph Joint-Embedding Predictive Architecture that models MAPF as the evolution of higher-order agent-resource interactions. TH-JEPA first constructs a dynamic conflict hypergraph from anticipated shared-resource usage, with persistent resource nodes preserving conflict identity over time. Training proceeds in two stages: masked temporal pretraining learns the structure of conflict evolution without actions, followed by action-conditioned post-training that predicts future agent and resource representations over multiple horizons. To prevent collapse in the shared latent space, we introduce Conflict-Spectral Dispersion Regularization (CSDR), which preserves variation over the conflict hypergraph while discouraging redundant latent directions. At inference, candidate joint-action sequences are rolled forward in latent space and evaluated using receding-horizon planning. We further establish that pairwise projection cannot, in general, preserve higher-order conflict structure and derive sufficient conditions under which CSDR prevents structural collapse. Experiments across dense and out-of-distribution MAPF settings show improved scaling to larger agent populations, stronger performance under higher-cardinality conflicts, and accurate multi-horizon prediction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi Agent, Graph Neural network

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_unvnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces TH-JEPA, a predictive world model for multi-agent path finding (MAPF). It constructs a dynamic agent–resource conflict hypergraph in which persistent navigation resources represent shared bottlenecks and hyperedge memberships evolve over time. Training consists of masked temporal representation pretraining followed by action-conditioned multi-horizon latent prediction. A conflict-spectral dispersion regularizer (CSDR) is introduced to discourage representation collapse, and the resulting model supports receding-horizon joint-action planning. Experiments on POGEMA and additional MAPF maps report improvements in success rate, congestion, and generalization to larger populations and unseen layouts.

Critical strengths
  1. Modeling anticipated contention around persistent resources rather than only current pairwise proximity is a reasonable idea for dense MAPF. The experiments grouping cases by hyperedge cardinality help motivate why higher-order coordination may matter.

  2. The paper evaluates population and topology shifts, multi-step prediction, action shuffling, pairwise-versus-hypergraph representation, and learned-versus-oracle dynamics. These results are substantive and suggest that the combined system can improve dense MAPF planning.

Critical weaknesses
  1. TH-JEPA primarily combines established components: incidence-based hypergraph message passing, masked JEPA-style predictive pretraining, action-conditioned latent dynamics, dispersion regularization, and receding-horizon planning. The main distinct design appears to be persistent-resource hyperedges derived from short nominal routes. While useful, it is not established that this representation provides a qualitatively new capability beyond resource-aware bipartite graphs, enriched pairwise graphs, or graph world models supplied with the same persistent resource identifiers. The Pairwise-JEPA and Random-HG ablations are helpful, but a capacity-matched resource-aware graph baseline would more directly test this claimed contribution.

  2. Theorem 1 formalizes the well-known fact that an unweighted pairwise projection loses hyperedge identity and multiplicity. It does not exclude pairwise models with hyperedge/resource attributes, which are a more relevant alternative. Theorem 2 is a conditional linear-algebra result: if the projected structural energies are bounded below and the off-diagonal terms are sufficiently small, the corresponding Gram matrix is full rank. This does not show that optimizing the finite-weight CSDR loss achieves those conditions, nor does it rule out collapse outside the selected projected subspace. The necessary feasibility condition on the number of projected dimensions relative to the Laplacian rank also deserves discussion. Consequently, the theoretical contribution should be presented more modestly.

Questions for authors
  1. What does the persistent-resource conflict hypergraph capture that a resource-aware bipartite or attributed pairwise world model cannot capture under an equivalent encoder and planning budget? Do the authors have a matched comparison?
  2. Does minimizing CSDR ensure the assumptions of Theorem 2 are satisfied in practice? How is the bound affected when the number of projected dimensions exceeds the rank of the conflict-hypergraph Laplacian, for example in highly disconnected states?
  3. How much of the gain remains when the same candidate generator, collision checks, receding-horizon depth, and readout are applied to a pairwise or alternative learned world model? What is the full per-decision wall-clock cost at 128 agents?
  4. How sensitive is the method to the hand-defined resource inventory and nominal route lookahead? How does it perform when the resource definitions are incomplete or misaligned with actual contention?
3 comments

We appreciate your question and suggestions for the paper. Thank you for asking which gains come from the resource representation and which come from predicting the consequences of actions.

1. Weakness 1 and Question 1 — Contribution beyond a resource-aware graph

A resource-aware bipartite graph can preserve the same agent–resource memberships. Theorem 1 addresses the narrower case of an unweighted agent-only projection. One aisle shared by three agents and three separate pairwise conflicts can produce the same triangle. Explicit resource identities preserve that distinction.

Our contribution is to predict how each persistent resource changes under a candidate action sequence. The original Table 4 reports 85.2% success for a direct policy using the hypergraph encoder and 95.3% for the full predictive planner. This measures the combined benefit of predictive training and planning. Figure 3 further shows that the advantage over Pairwise-JEPA grows from 3.1 percentage points for conflicts involving 3–4 agents to 16.5 points for conflicts involving at least seven.

We added Table 12 in Appendix D.1 for the requested comparison under a shared world-model and planning budget.

Representation in the proposed comparisonI success (%)relative SoC
HMAGAT construction with the shared framework92.11.17
Resource-aware bipartite world model94.01.12
TH-JEPA94.51.11

2. Weakness 2 and Question 2 — What does CSDR guarantee?

Minimizing CSDR encourages variation and penalizes redundant directions. A finite penalty cannot guarantee that the assumptions of Theorem 2 hold. With NN agents and ctc_t connected components, rank⁡(Ct)≤N−ct\operatorname{rank}(C_t)\leq N-c_t. Disconnected states therefore support fewer independent directions. We corrected Section 4.5 and Appendix A.1 to state this limit and restrict the theorem to the projected subspace. In the original ablation, removing CSDR lowered success by 3.9 percentage points.

3. Question 3 — Shared-planner controls and wall-clock cost

The original Table 5 keeps the planner fixed and replaces learned transitions with simulator transitions. Success increases from 95.3% to 97.7%. This isolates the effect of transition error. The separate control in revised Appendix D.1 addresses the effect of changing the representation under the same planner.

Original Appendix H reports 31.9 ms per replanning decision on Dense Warehouse at 128 agents. Its timing description includes hypergraph construction, encoding, latent rollout, and scoring. New Appendix I reports a different quantity: total wall-clock time per successful episode. It therefore complements the original latency figure.

4. Question 4 — Sensitivity to resource definitions and look-ahead

Original Appendix B.2 specifies the extraction rules and the eight-step route look-ahead. If extraction misses an aisle, the model loses the resource token that tracks contention there. If a route estimate misses a detour, the model can associate an agent with the wrong resource. Our held-out Room and Warehouse experiments test transfer across layouts using the same extraction rule.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the new resource-aware bipartite baseline, the clarification of the CSDR theoretical guarantee, and the additional planning and computational comparisons.

I consider the theoretical clarification and computational analysis largely satisfactory. The authors now appropriately acknowledge that Theorem 2 provides conditional sufficient properties rather than a general guarantee against representation collapse, and the reported runtime measurements help clarify the practical cost of the proposed planner.

However, I still have some reservations about the methodological novelty. The new resource-aware bipartite baseline achieves performance close to TH-JEPA under a matched world-model and planning framework. This suggests that preserving persistent resource information accounts for much of the improvement, while the distinct contribution of the proposed hypergraph formulation remains relatively limited.

I have two remaining questions:

  1. Matched comparison and statistical significance: Table 12 reports 95.3% success for TH-JEPA, whereas the accompanying text and rebuttal report 94.5%. Could the authors clarify the correct result and provide the variability across the five seeds? Given the relatively small improvement over the resource-aware bipartite baseline (94.0%), how statistically reliable is this difference?

  2. Representation novelty and resource sensitivity: The authors acknowledge that an incidence-equivalent bipartite encoder can represent the same information as the proposed hypergraph. Could the authors clarify what specific advantage remains attributable to the proposed representation, rather than to persistent resource features and action-conditioned predictive planning? Is there any existing evidence regarding sensitivity to inaccurate resource extraction or different route-lookahead horizons?

Overall, the rebuttal has strengthened my assessment of the paper's technical credibility and experimental rigor. Nevertheless, I remain uncertain whether the methodological contribution is sufficiently distinctive for ICLR. I am therefore maintaining my current rating of 2 (Weak Reject), while recognizing the value of the proposed predictive-planning framework and the authors' substantial efforts to address the review concerns.

Our novelty lies in adapting JEPA to action-conditioned prediction over evolving multi-agent resource interactions. The model learns how joint robot actions change future agent and resource representations, allowing the planner to compare consequences directly in latent space without reconstructing the full environment. Latent-space action planning has not been implemented for this setting.

A bipartite encoder with the same links can represent that bottleneck. What the encoder alone cannot do is predict how contention changes if one group enters and another waits. TH-JEPA rolls both choices forward in latent space before selecting a move. Table 4 shows why this matters: the direct policy using the same hypergraph encoder achieves 85.2% success, while the full predictive planner reaches 95.3%. This supports the value of learning action consequences on persistent resources. A separate gain from the hypergraph form over an incidence-equivalent bipartite encoder remains unmeasured.

For sensitivity, we applied the same extraction rule and eight-step route lookahead to unseen Room and Warehouse maps. This test transfers across layouts.

Reviewer_eRW4OpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TH-JEPA, a Temporal Hypergraph Joint-Embedding Predictive Architecture, that uses a conflict hypergraph to represent agent-resource interactions. It uses a two stage training: masked temporal pretraining and action-conditioned post-training. The paper additionally proposes Conflict-Spectral Dispersion Regularization (CSDR) to mitigate representation collapse and provides theoretical analysis of its structural non-collapse properties.

Critical strengths
  1. The paper makes a reasonable case for jointly modeling higher-order conflicts and their temporal evolution.
  2. TH-JEPA seems to outperform other methods on success rate and makespan. The paper also provides useful ablation experiments.
Critical weaknesses
  1. Poor Evaluation Reports: The paper does not report the runtime and sum of costs comparison of the methods across different maps and different agent counts. Providing graphs of these metrics alongside the success rates is standard in the field. In fact, all of the baseline methods provide these graphs in their respective papers.
  2. Map information: The information about the maps is poor. The paper uses the same name of the maps as the HMAGAT paper, but clearly the maps are different, as seen by Figure 4. This is misleading especially given the lack of proper information about the maps.

Together these make understanding the empirical benefits of the model hard.

Questions for authors
  1. Can you please provide the runtime and sum of costs comparison of the methods across different maps and different agent counts?
  2. Can you please provide information about the maps?
  3. Can you provide results of TH-JEPA and the baseline methods on maps and instances used in HMAGAT?
2 comments

Thank you for asking us to clarify the computational cost and identify the maps used in our comparisons.

1. Weakness 1 and Question 1 — Runtime and sum of costs

The original Appendix H reported decision latency. On Dense Warehouse with 128 agents, TH-JEPA took 31.9 ms per decision, compared with 9.8 ms for HMAGAT and 18.6 ms for MAPF-World. The original Appendix G.1 also reported how planning depth affects solution cost. Increasing the horizon from one to eight steps lowered relative sum of costs from 1.20 to 1.11.

However, we added Table 20 in Appendix I to bring success rate, relative sum of costs, and episode runtime together across four map families and three agent counts. The following excerpt reproduces the entries in the current revision for Dense Warehouse. We report the new cost and episode-time values.

AgentsMethodSuccess (%)Relative SoCSeconds per successful episode
64HMAGAT98.41.080.82
64TH-JEPA99.21.041.72
90HMAGAT94.51.121.41
90TH-JEPA97.71.072.51
128HMAGAT89.11.172.12
128MAPF-World84.41.222.79
128TH-JEPA94.51.113.47

Appendix I defines runtime as mean wall-clock time per successful episode, including inference and planning. Relative SoC uses the LaCAM3 cost on the same instance as its reference. Since both metrics depend on which instances a method solves, the table reports success rate alongside them.

2. Weakness 2 and Questions 2–3 — Map information and the original HMAGAT benchmark

The original Section 5 and Appendix B.1 described the 21,000-instance collection, with 20% random-obstacle layouts and 80% mazes. The 17×1717\times17 to 21×2121\times21 dimensions refer to training grids. We held out Room and Warehouse layouts for evaluation. For each small-map family and population, we evaluated all methods on the same 500 instances with the same starts, goals, and 256-step limit.

We added Table 11 in Appendix C.1 to distinguish the training maps, small evaluation maps, and larger evaluation maps. It also identifies the missing test-grid dimensions and map identifiers. Figure 4 shows how the methods resolve the same conflict on one instance from our suite. The 5.4 percentage-point success advantage over HMAGAT at 128 agents applies to the shared Dense Warehouse instances used in our evaluation. Revised Appendix C.1 makes that scope explicit.

My original concerns still stand.

  1. The main results should be reported using the standard metrics. TH-JEPA uses LaCAM3 as an expert for imitation and LaCAM3 is designed and presented with sum-of-costs (SoC) metric in mind. Each baseline method focuses on SoC. Thus, SoC is an important metric to report. Runtime, success rates and SoC for each configuration (map type and agent count) should be reported for all the methods.

  2. Lack of information about the maps. There should be sample maps provided in the manuscript. The description in the response by the authors is not enough. More information is needed for the random nomenclature, as they seem totally different from the maps used in previous works, and how these maps were generated.

The SoC results are also not impressive. There seems to be a minor improvement in SoC over HMAGAT, while requiring 50-100% more time than it. The SoC results for MAPF-World are not provided for the lower agent counts.

Overall, I believe the paper needs larger changes to the evaluation setup and reporting, which is likely out-of-scope for this discussion period now. Thus, I am more convinced in my rejection recommendation and have updated my score from 2 to 1.

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

TH-JEPA models MAPF via persistent resource-based conflict hypergraphs. It combines masked temporal pretraining, action-conditioned latent prediction, and CSDR regularization, followed by latent receding-horizon planning. The experiments report improved population non-collapse guarantees, yet the implementation details of the planning procedure require clarification.

Critical strengths
  1. The hypergraph associates agents with the resources they intend to occupy and maintains persistent resource identities across time, yielding a native MAPF-friendly structure to forecast the evolution of conflicts.
  2. The method exhibits compelling generalization across agent populations and environment topologies, notably under dense, higher-order conflict settings. Extensive evidence including component ablations, action-shuffling and counterfactual analyses, as well as comparisons against oracle dynamics, strongly corroborates the efficacy of the proposed system.
Critical weaknesses
  1. Theorem 2 requires q independent structural directions. However, rank(L) is at most N-1 because L annihilates the constant vector. Training uses N = 16, 24, or 32 with q = 32 (Table 7). Thus C cannot be positive definite for any individual training state. CSDR may remain useful, but the theorem cannot provide the claimed full-rank guarantee in this setting.
  2. Section 4.4 and Appendix B.8 omit partial-assignment ranking and future action proposals. The risk readout is supervised from trajectory segments, but the stated data are valid expert solutions. The paper does not explain where positive collision examples come from or how risk prediction is trained on off-expert candidates. This limits reproducibility and assessment of the planner.
  3. TH-JEPA combines established ideas in hypergraph-based MAPF, action-conditioned prediction, joint-embedding learning, and receding-horizon planning. Its main novelty appears to be the persistent resource representation and its integration with latent dynamics. However, the ablations do not establish whether this representation improves over existing MAPF hypergraphs using the same world model and planner.
Questions for authors
  1. Is C computed per state or pooled across states? How can q = 32 satisfy Theorem 2 during training?
  2. Please specify candidate generation and readout training, including any invalid or counterfactual training trajectories.
  3. What benefit does the persistent resource representation provide over existing MAPF hypergraphs when the world model and planner are held fixed?
2 comments

Thank you for recognizing the value of tracking shared resources over time.

1. Weakness 1 and Question 1 — Scope of Theorem 2

We compute CtC_t separately for each state. It is correct that rank⁡(Ct)≤rank⁡(Lt)≤N−1\operatorname{rank}(C_t)\leq\operatorname{rank}(L_t)\leq N-1. Even our largest training state has only 32 agents, so it supports at most 31 independent structural directions. We corrected Section 4.5 and Appendix A.1 to state that the full-rank guarantee cannot hold with q=32q=32 during training.

The original Table 4 gives the empirical reason for using CSDR. Removing it reduced success from 95.3% to 91.4% and increased relative sum of costs from 1.11 to 1.18. This ablation supports its use as a regularizer.

2. Weakness 2 and Question 2 — Candidate generation and collision supervision

The original Section 4.4 specified the eight-step horizon, immediate collision checks, and budget of 128 candidates. Revised Appendix B.8 adds the partial-assignment score and future-action proposal rules. Four width-32 searches use different agent orders. Each search treats unassigned agents as waiting while it scores partial assignments by goal progress, waiting, and congestion. Merging the completed searches gives at most 128 sequences.

Appendix B.8 also describes a readout-training procedure that generates off-expert proposals from training states. A proposal receives a positive label if two agents attempt to enter the same cell or swap positions within the horizon. The simulator records the attempt before resolving the conflict. This provides a concrete source of positive labels. The original expert trajectories alone do not establish which positive examples the original readout received.

3. Weakness 3 and Question 3 — Contribution beyond existing MAPF hypergraphs

HMAGAT predicts actions from the current interaction state. TH-JEPA evaluates what happens after different joint actions. For example, it can predict congestion at the same aisle if both groups enter or if one group waits. In our original counterfactual experiment, TH-JEPA identified the lower-cost sequence in 87.6% of 1,000 pairs. Shuffling the actions reduced accuracy to 51.2%. The original representation ablation also showed that replacing resource hyperedges with pairwise interactions reduced success from 95.3% to 88.3%. To address the reviewer's question, we also added a comparison in Appendix D.1 that keeps the world model and planner fixed while changing the resource construction.

I sincerely thank the authors for the careful and candid responses. The correction to Theorem 2 and the added details on candidate generation are much appreciated. One point remains open for me: the clarification that the original expert trajectories do not establish the readout's positive examples leaves me slightly unsure how the reported planner was trained. More broadly, I am still not sure what problem the paper solves beyond existing MAPF hypergraph approaches, or where its essential novelty lies. I therefore keep my score.

Comments on the paper

1 comment

Dear Reviewers,

We are now halfway through the discussion phase. The authors have provided their responses to your reviews. If their responses have not adequately addressed your concerns, please take this opportunity to raise any remaining questions or concerns.

I encourage you to engage with the authors as soon as possible to ensure a productive discussion.

Many thanks,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.