Pith. sign in

REVIEW 5 major objections 6 minor 63 references

PRIMAL3: Pathfinding via Reinforcement and Imitation Multi-Agent Learning - Leveraging LaCAM3

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PRIMAL3 claims a learned decentralized policy that coordinates up to 100,000 agents, reaching 95–99% success at 1,000–10,000 agents and nearly matching search-based solvers on random maps.

desk verdict Solid incremental architecture with credible mid-scale results; the city-scale claim needs a runtime audit before it's taken at face value. read the letter →

arxiv 2608.04905 v1 pith:726VX7M6 submitted 2026-08-05 cs.RO

Chengyang He , Tanishq Duhan , Gadiel Sznaier Camps , Fangyuan Wang , Yuhong Cao , Jiankai Sun , Ge Sun , Mac Schwager
show 1 more author
Guillaume Sartoretti
This is my paper · ORCID
classification cs.RO
keywords multi-agentpathfindingreinforcementlearningimitationdecentralizedpolicygraphcommunicationPIBTactionrefinementLaCAM3expertguidancescalableMAPF
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PRIMAL3 aims to show that a learned, decentralized multi-agent pathfinding policy can scale to city-level problems that were previously the preserve of centralized search-based planners. The paper's central claim is that a policy trained with reinforcement learning plus selectively injected expert guidance reaches success rates of 95–99% on instances with 1,000 to 10,000 agents, completes single instances with 50,000 and 100,000 agents, and on random maps performs almost on par with state-of-the-art search-based solvers such as LNS2 and LaCAM3. The key is to make coordination graph-structured: one graph for agents moving in the same direction and one for agents competing for the same space, together with topology features that expose bottlenecks and dead ends. If the claim holds, learned reactive policies become a practical alternative to centralized replanning for large-scale fleets.

What carries the argument

The load-bearing object is a pair of interaction graphs built from A* reference paths. A directed conflict graph weights opposing, stationary, and differently-oriented path overlaps with a discount factor $\gamma_{\mathrm{conf}}=0.9$, while a following graph weights same-direction overlaps with $\gamma_{\mathrm{foll}}=0.5$; each agent keeps at most one conflict neighbor and one following neighbor, updated only when the previous overlap vanishes. These graphs feed two separate communication branches: multihop gated aggregation for following relations, and masked attention with ego-difference encoding for conflict relations. Training uncertainty is measured by action entropy against a threshold of about 0.464, which gates when LaCAM3 intervenes. Execution uses a priority-aware PIBT shield, where PIBT is priority inheritance with backtracking, an online method that resolves vertex and edge conflicts by propagating priority; here its priority is a weighted combination of persistent age, a learned scalar, and normalized goal distance.

What would settle it

Run PRIMAL3's graph-construction steps (Algorithms 1 and 2) on a 100,000-agent instance and measure wall-clock time and memory per timestep; if they exceed the execution horizon of 2048 steps or available memory, the scaling claim fails.

Watch

Extended reading notes

Core claim

PRIMAL3 claims that decentralized learning-based MAPF can be scaled far beyond previous limits by making both communication and action refinement sensitive to the type of interaction and the map's topology. The policy distinguishes path-compatible interactions, handled by multihop message passing on a following graph, from path-competing interactions, handled by masked attention on a conflict graph. During training, agents whose action distributions have high entropy are identified as uncertain, and for those agents the LaCAM3 planner supplies label-smoothed expert actions; in the late episode, unresolved agents follow a cached LaCAM3 plan. During execution, a priority-aware PIBT module uses persistent waiting times, learned priorities, and goal distances to keep actions collision-free while retaining the policy's fallback preferences. The result is a policy that does not need LaCAM3 at inference yet, on random 32x32 maps, reaches success rates from 1.00 down to 0.92 for 50 to 300 agents and, at ultra-large scale, maintains 95.0–99.0% success from 1,000 to 10,000 agents, with two single-instance demonstrations at 50,000 and 100,000 agents.

Load-bearing premise

The city-scale claim assumes that computing an A* path for every agent and the pairwise path-overlap matrices at each step stays fast enough at 100,000 agents, but the paper reports no runtime or memory scaling, so the practical feasibility is unverified.

Editorial extensions

If this is right

  • A decentralized learned policy can solve one-shot MAPF instances with 100,000 agents, a scale previously unreported for learning-based methods, without invoking LaCAM3 at inference.
  • On random maps, learning-based MAPF can reach success rates close to centralized search-based solvers (LNS2 and LaCAM3), narrowing the long-standing gap between reactive policies and search.
  • Each agent communicates with at most one partner per interaction type, keeping average communication degree at most 2 regardless of team size, which is what makes the graph updates scale-independent.
  • The effective per-agent failure hazard of PRIMAL3 at 1,000–10,000 agents is about 16 times lower than HMAGAT's, meaning the success-rate gap widens predictably with team size rather than from a sudden policy collapse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The entropy-thresholded injection of an expensive search-based expert suggests a general training recipe: use a classical planner as a selective oracle only where the policy is uncertain, which could transfer to other multi-agent coordination tasks beyond MAPF.
  • Editorial inference: The paper does not report per-timestep runtime or the cost of computing pairwise path-overlap matrices; verifying that Algorithms 1 and 2 scale to 100,000 agents within the 2048-step horizon is the natural next test before deployment claims are accepted.
  • Editorial inference: Since the following and conflict graphs are built from reference paths, the approach could extend to lifelong MAPF by recomputing reference paths on goal changes, with the persistent age variable already providing the temporal memory that lifelong settings need.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This paper proposes PRIMAL3, a decentralized learning-based MAPF framework with three main components: a topology-aware dual-graph communication module that separates following and conflict interactions derived from A* reference-path overlaps, a LaCAM3-guided confidence-boosting training scheme that uses policy entropy to trigger selective expert interventions, and a priority-aware PIBT action-refinement shield that preserves the policy's fallback preferences. The evaluation compares against search-based solvers (LaCAM, LaCAM3, LNS2) and learning-based baselines (SYLPH, MAPF-GPT, HMAGAT) on 32x32 random and maze maps, reports success rates of 95-99% for up to 10,000 agents, describes success on single 50,000 and 100,000 agent instances, provides component ablations, and includes two real-robot demonstrations.

Significance. If the results hold, PRIMAL3 would be a meaningful step for learned MAPF: it shows that a decentralized policy with topology-aware communication can substantially narrow the gap to centralized search solvers at moderate scale and pushes the scale envelope of learned policies. The paper has clear strengths: the following/conflict graph separation is well motivated; the entropy-triggered selective use of LaCAM3 is a sensible way to spend expert queries; the priority-aware PIBT shield preserves action preferences; and the effective-failure-load analysis in Section V-C is a thoughtful way to compare policies under the all-agents-success criterion. The real-robot experiments, although small, support deployability. However, the current manuscript does not yet provide enough evidence for the headline city-scale and near-search-parity claims, and one part of the central graph-construction pseudocode is inconsistent with the equations.

major comments (5)
  1. [IV-A.2, Algorithm 1] Algorithm 1 is inconsistent with Eq. (7) and Eq. (8) in the non-opposite, both-nonzero, different-direction case. The else branch on line 15 adds I[r_i(k) = r_j(l)] * gamma_conf^k, but the branch is reached only when r_i(k) != r_j(l), so the indicator is always 0. Eq. (8) assigns rho_conf = 0.5 to this case, and the text and Fig. 2 say that different-direction overlaps should enter the conflict graph. As written, the pseudocode would discard exactly the soft different-direction conflict signal. Please correct the pseudocode and state which behavior was actually implemented.
  2. [IV-A.2 and V-C] The city-scale claims in Section V-C rest on an unverified per-timestep cost. Algorithm 1 enumerates all ordered agent pairs and all overlapping positions between their reference paths, and Algorithm 2 performs a row-wise argmax over the dense matrices A_conf and A_foll. The sparsity argument in Section V-C ('each symmetrized graph contains at most N edges, average degree at most 2') bounds only the final communication graphs, not the dense computation needed to select those edges. At 50,000 or 100,000 agents, the reference-path computation and the pairwise overlap enumeration must be executed at every decision step, yet the paper reports no wall-clock time, no memory usage, and no asymptotic complexity for any scale. Without this information, or an explicit pruned implementation, the feasibility of the headline scaling result is not established.
  3. [V-C, last paragraph] The 50,000- and 100,000-agent results are based on two single instances, with no map side lengths, no instance specification, no repeated runs, and no information on how close the runs were to the horizon limit. A single favorable configuration is not sufficient evidence for the abstract's 'up to city-level 100,000 agents' claim. Please report multiple instances per scale, the exact map sizes, and the number of used timesteps.
  4. [V-D2 / Fig. 8] The decay-factor sweep selects gamma_conf = 0.9 and gamma_foll = 0.5 on the same 32x32 maze 64-agent benchmark that is then used for the component ablation and for the headline 93.5% success rate. The full-model number is therefore the maximum of a tuning sweep on the evaluation set, and the ablated variants in Fig. 7 are compared at parameters chosen to favor the full model. This is a fitting-on-the-test-set concern. Please move the decay-factor selection to a validation set, or show that the ablation conclusions are unchanged over a range of decay settings.
  5. [V-B and Fig. 5] The claim that PRIMAL3 is 'almost on par' with LNS2 and LaCAM3 on random maps is not statistically supported. Fig. 5 reports no confidence intervals or significance tests for the 200 instances, and the comparison mixes time budgets: search-based solvers are allowed 30 seconds of planning time while PRIMAL3's wall-clock inference time is not reported. Please provide binomial confidence intervals for the success rates, report per-instance wall-clock time for all methods, and state the random-map generation protocol so that the comparison is interpretable.
minor comments (6)
  1. [VII] The conclusion says 'scalability to instances with up to 10,000 agents,' but the abstract and Section V-C claim 50,000 and 100,000 agents; please make these statements consistent.
  2. [Fig. 7] The ablation is labeled 'w/o LaCAM* guidance' while the text mostly says LaCAM3; please standardize the terminology.
  3. [IV-C, Eq. (36)] The learned priority term s_prio_theta,i in Eq. (36) is never defined as a network output or given a training objective; please clarify where this quantity comes from.
  4. [IV-B, Eqs. (27) and (33)] Equations (27) and (33) use epsilon_num inside the logarithm; since the action distribution is a softmax output it is already strictly positive, so this additive constant should either be removed or applied to the probability before the log.
  5. [V-C] The exact map side lengths for the 50,000 and 100,000 agent instances are missing; please include them with the other instance details.
  6. [General] A code release or a more detailed reproducibility appendix would strengthen the empirical claims; the current manuscript does not state whether code is available.

Circularity Check

1 steps flagged · score 3.0 of 10

The 93.5% maze-64 success rate is the in-sample maximum of the decay-factor sweep on the same benchmark, so that ablation/number is partly fitted; the central learned-policy pipeline and other benchmarks remain independent.

  1. fitted input called prediction [Section V-D.2 (Decay Factors Analysis), applied in V-D.1 (Components Ablation) and reported in V-B (maze SR at 64 agents); values introduced in Section IV-A.2.]
    "Based on this analysis, we use (γ_conf, γ_foll) = (0.9, 0.5) in the final model. ... The best performance, 93.5%, is achieved when γ_conf = 0.9 and γ_foll = 0.5. ... As shown in Fig 7(a), the full PRIMAL3 model achieves the highest success rate of 93.5%. ... Specifically, its SR reaches 1.00, 1.00, 0.975, and 0.935 for 8, 16, 32, and 64 agents."

    The decay-factor sweep in V-D.2 is performed on 32×32 maze maps with 64 agents, the same benchmark used for the full-model ablation (V-D.1) and for the maze result reported in V-B. The paper selects (γ_conf, γ_foll) = (0.9, 0.5) precisely because that pair yields the maximum success rate (93.5%) on that benchmark, and then reports 93.5% (V-D.1) and SR 0.935 (V-B) as the full PRIMAL3 model's performance on the same benchmark. The reported number is therefore the in-sample optimum of the tuning surface, not an independent estimate of the chosen configuration. Presenting this selected maximum as an ablation-validated result makes the 93.5% figure partly fitted by construction rather than a prediction.

full rationale

The central derivation is not circular: LaCAM3 is an external search-based solver used only to generate training labels and interventions, and at inference PRIMAL3 executes the learned policy without LaCAM3, so the claim that a learned decentralized policy works without the expert is not definitionally forced. The baselines (LaCAM, LNS2, LaCAM3, HMAGAT, MAPF-GPT, SYLPH) are external or prior methods, and success is measured end-to-end. The SYLPH citations for reference-path overlap and learned priority scalars are supporting design choices; the paper supplies its own algorithms and equations and does not invoke a self-citation as a uniqueness or correctness proof. The one genuine circularity is narrow but concrete: the conflict/following decay factors are explicitly tuned on the 32×32 maze / 64-agent benchmark, and the same benchmark's best success rate (93.5%, i.e., SR 0.935 in V-B) is then presented as the full model's ablation and comparison result. That particular number is an in-sample fitted value, not an independent prediction. The random-map, ultra-large-scale (1,000-100,000 agents), and real-robot results are on different configurations and retain independent content, though the ultra-large feasibility also depends on unverified per-step computational cost, which is a correctness concern rather than circularity. Overall score 3 reflects one partly fitted headline number while the central learned-pipeline claim remains independent.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The method relies on a number of hand-set hyperparameters (gamma_conf, gamma_foll, eta, epsilon_E, rho_conf, and unreported weights) and on the assumption that a small-map, small-team policy transfers to ultra-large instances. The decay-factor selection on the evaluation benchmark is a fitting step. No code or precise training recipe is provided, so the contribution rests on many unstated defaults.

free parameters (7)
  • gamma_conf = 0.9
    Temporal discount for conflict overlaps; selected by a sweep on the 32x32 maze 64-agent benchmark (Figure 8).
  • gamma_foll = 0.5
    Temporal discount for following overlaps; selected in the same sweep as gamma_conf.
  • eta = 0.464
    Entropy threshold for uncertainty; derived from a chosen reference distribution, not tuned, but its value determines intervention frequency.
  • epsilon_E = 0.025
    Label smoothing for expert imitation targets (Equation 30); hand-set.
  • w_age, w_prio, w_dist = not reported
    Weights in the PIBT priority initialization (Equation 36); values are not given in the paper, so the contribution of each term is not reproducible.
  • lambda_BC = not reported
    Weight of the behavior-cloning loss (Equation 34); value not reported.
  • rho_conf coefficients = 5, 1, 0.5
    Hand-crafted conflict weights in Equation 8 for stationary, opposite, and other directional mismatches.
assumptions (7)
  • standard math A* computes an optimal reference path for each agent individually in the current grid graph.
    Used in Section IV-A.1 to define reference paths p_i and the following/conflict overlap matrices; assumes the grid graph and movement model are static and known.
  • domain assumption LaCAM3 can produce valid, collision-free joint plans for the training instances (up to 14 agents).
    Used in Section IV-B to provide expert actions for uncertain agents; if LaCAM3 fails or is slow on some training states, the confidence-triggered intervention provides no useful signal.
  • ad hoc to paper The policy trained on maps of size 10 to 30 with 8 to 14 agents generalizes to 32x32 maps with up to 300 agents and to 72x72 to 226x226 maps with 1,000 to 10,000 agents without further training.
    Section V-A training setup and Section V-C evaluation support the scalability claim entirely on this extrapolation; no fine-tuning or domain adaptation is described.
  • ad hoc to paper The entropy threshold eta=0.464, derived from the reference distribution [0.9,0.025,...], correctly identifies states where expert intervention is needed.
    Equations 27 to 29 in Section IV-B.1; no sensitivity analysis for eta is provided.
  • ad hoc to paper The path-overlap classification into following and conflict interactions captures the coordination requirements near bottlenecks and dead ends.
    Section IV-A.2 equations 7 to 8; the weight choices (5 for stationary blocking, 1 for opposite, 0.5 for other) are heuristic and not derived.
  • standard math PIBT with priority inheritance and backtracking maintains collision-free one-step actions.
    Adopted from prior work [38] and used as the refinement module in Section IV-C.
  • ad hoc to paper The decay-factor configuration (gamma_conf=0.9, gamma_foll=0.5) selected on the 32x32 maze 64-agent benchmark is optimal for other map sizes and agent counts.
    Section V-D2 Figure 8 shows the sweep on the evaluation benchmark; assuming these values transfer to all reported settings is a fitting assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PRIMAL3: Pathfinding via Reinforcement and Imitation Multi-Agent Learning - Leveraging LaCAM3." pith.science (2026). https://pith.science/paper/726VX7M6

@misc{pith2026260804905,
  author       = {Pith},
  title        = {Pith review of: PRIMAL3: Pathfinding via Reinforcement and Imitation Multi-Agent Learning - Leveraging LaCAM3},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/726VX7M6}},
  note         = {Machine review of arXiv:2608.04905}
}
read the original abstract

We present PRIMAL3, an ultra-large-scale learning-based framework for multi-agent pathfinding (MAPF) that integrates reinforcement learning, topology-aware communication, LaCAM3-guided training, and PIBT-based action refinement. PRIMAL3 targets failures at topologically critical states, where agents must coordinate decisively around bottlenecks, dead ends, and persistent conflicts. Each agent is represented using features derived from cut vertices, dead-end regions, shortest-path distances, and blocking estimates. Two complementary graphs capture agent interactions: a same-direction following graph propagates multihop context along compatible paths, while a different-direction conflict graph differentiates agents competing for shared space through masked attention and relative features. During training, we propose to let policy entropy identify uncertain agents, for which LaCAM3 provides confidence-triggered action interventions and label-smoothed imitation targets. During execution, a priority-aware PIBT module refines the proposed joint actions using persistent, learned, and distance-aware priorities together with policy-aware fallback preferences while maintaining collision-free execution. The resulting framework combines learned exploration with structured expert guidance without requiring LaCAM3 at inference. Experiments demonstrate that PRIMAL3 substantially outperforms state-of-the-art learning-based baselines and scales to ultra-large instances with up to city-level 100,000 agents. Real-world experiments further demonstrate the feasibility of deploying PRIMAL3 on physical robotic systems and ablation studies validate the individual contributions the components we proposed. Project page: https://marmotlab.github.io/PRIMAL3/

Figures

Figures reproduced from arXiv: 2608.04905 by the authors.

Figure 1
Figure 1. From urban mobility to ultra-large-scale MAPF. Mobile entities and inaccessible city blocks can conceptually be [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Dual-graph construction from reference-path overlaps. Same-direction overlaps are accumulated into a following graph, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Following-graph communication for compatible path overlaps. Left: same-direction path overlaps induce directed [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Conflict-graph communication for competing path overlaps. Left: agents approaching shared space from incompatible [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Comparison results on 32 × 32 random and maze maps across varying numbers of agents. We report success rate (SR), episode length (EL), and arrival rate (AR). Higher SR and AR are better, whereas lower EL is preferred. PRIMAL3 is highlighted in green. 200, 250, and 300 …
Figure 6
Figure 6. Figure 6: Ultra-large-scale evaluation of PRIMAL3 and HMAGAT. (a) Instance-level success rate, where success requires all [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Components ablation on 32×32 maze maps with 64 agents. Panels show (a) success rate, (b) average episode length, and (c) arrival rate. Numbers above the bars in panel (a) indicate success rates in percentage. Bars denote the mean performance, open circles denote indivi…
Figure 8
Figure 8. Figure 8: Success-rate landscape under different decay-factor [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Real-robot evaluation of PRIMAL3. Panels show (above) 8 physical robots performing MAPF on a [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 42 canonical work pages

  1. [1]

    Lifelong multi-agent path finding in large-scale warehouses,

    J. Li, A. Tinka, S. Kiesel, J. W. Durham, T. S. Kumar, and S. Koenig, “Lifelong multi-agent path finding in large-scale warehouses,” inPro- ceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 13, 2021, pp. 11 272–11 281

    2021

  2. [2]

    Mobile robot path planning in dynamic environments through globally guided reinforcement learning,

    B. Wang, Z. Liu, Q. Li, and A. Prorok, “Mobile robot path planning in dynamic environments through globally guided reinforcement learning,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 6932–6939, 2020

    2020

  3. [3]

    P3gasus: Pre-planned path execution graphs for multi-agent systems at ultra-large scale,

    T. Duhan, C. He, and G. Sartoretti, “P3gasus: Pre-planned path execution graphs for multi-agent systems at ultra-large scale,”IEEE Robotics and Automation Letters, vol. 11, no. 2, pp. 1274–1281, 2025

    work page 2025

  4. [4]

    Intersection coordination with priority-based search for autonomous vehicles,

    J. Li, E. Lin, H. L. Vu, S. Koeniget al., “Intersection coordination with priority-based search for autonomous vehicles,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023, pp. 11 578–11 585

    work page 2023

  5. [5]

    Multi-agent pathfinding: Defi- nitions, variants, and benchmarks,

    R. Stern, N. Sturtevant, A. Felner, S. Koenig, H. Ma, T. Walker, J. Li, D. Atzmon, L. Cohen, T. Kumaret al., “Multi-agent pathfinding: Defi- nitions, variants, and benchmarks,” inProceedings of the international symposium on combinatorial search, vol. 10, no. 1, 2019, pp. 151–158

    work page 2019

  6. [6]

    Feasibility study: Moving non-homogeneous teams in congested video game environ- ments,

    H. Ma, J. Yang, L. Cohen, T. Kumar, and S. Koenig, “Feasibility study: Moving non-homogeneous teams in congested video game environ- ments,” inProceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, vol. 13, no. 1, 2017, pp. 270–272

    work page 2017

  7. [7]

    M*: A complete multirobot path planning algorithm with performance bounds,

    G. Wagner and H. Choset, “M*: A complete multirobot path planning algorithm with performance bounds,” in2011 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2011, pp. 3260– 3267

    work page 2011

  8. [8]

    Conflict-based search for optimal multi-agent pathfinding,

    G. Sharon, R. Stern, A. Felner, and N. R. Sturtevant, “Conflict-based search for optimal multi-agent pathfinding,”Artificial intelligence, vol. 219, pp. 40–66, 2015

    2015

Show all 63 references
  1. [9]

    Subdimensional expansion for multirobot path planning,

    G. Wagner and H. Choset, “Subdimensional expansion for multirobot path planning,”Artificial intelligence, vol. 219, pp. 1–24, 2015

    2015

  2. [10]

    Suboptimal variants of the conflict-based search algorithm for the multi-agent pathfinding problem,

    M. Barer, G. Sharon, R. Stern, and A. Felner, “Suboptimal variants of the conflict-based search algorithm for the multi-agent pathfinding problem,” inProceedings of the international symposium on combinatorial Search, vol. 5, no. 1, 2014, pp. 19–27

    2014

  3. [11]

    Eecbs: A bounded-suboptimal search for multi-agent path finding,

    J. Li, W. Ruml, and S. Koenig, “Eecbs: A bounded-suboptimal search for multi-agent path finding,” inProceedings of the AAAI conference on artificial intelligence, vol. 35, no. 14, 2021, pp. 12 353–12 362

    2021

  4. [12]

    Anytime multi-agent path finding via large neighborhood search,

    J. Li, Z. Chen, D. Harabor, P. J. Stuckey, and S. Koenig, “Anytime multi-agent path finding via large neighborhood search,” inInternational joint conference on artificial intelligence 2021. Association for the Advancement of Artificial Intelligence (AAAI), 2021, pp. 4127–4135

    2021

  5. [13]

    Mapf-lns2: Fast repairing for multi-agent path finding via large neighborhood search,

    ——, “Mapf-lns2: Fast repairing for multi-agent path finding via large neighborhood search,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 9, 2022, pp. 10 256–10 265

    2022

  6. [14]

    Lacam: Search-based algorithm for quick multi-agent pathfinding,

    K. Okumura, “Lacam: Search-based algorithm for quick multi-agent pathfinding,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023, pp. 11 655–11 662

    2023

  7. [15]

    Engineering lacam*: Towards real-time, large-scale, and near- optimal multi-agent pathfinding,

    ——, “Engineering lacam*: Towards real-time, large-scale, and near- optimal multi-agent pathfinding,”arXiv preprint arXiv:2308.04292, 2023. JOURNAL OF LATEX CLASS FILES, VOL. *, NO. *, JULY 2026 18

    2023 arXiv

  8. [16]

    Primal: Pathfinding via reinforcement and imitation multi- agent learning,

    G. Sartoretti, J. Kerr, Y . Shi, G. Wagner, T. S. Kumar, S. Koenig, and H. Choset, “Primal: Pathfinding via reinforcement and imitation multi- agent learning,”IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2378–2385, 2019

    2019

  9. [17]

    Mapper: Multi-agent path planning with evolutionary reinforcement learning in mixed dynamic environments,

    Z. Liu, B. Chen, H. Zhou, G. Koushik, M. Hebert, and D. Zhao, “Mapper: Multi-agent path planning with evolutionary reinforcement learning in mixed dynamic environments,” in2020 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 11 748–11 754

    2020

  10. [18]

    Graph neural networks for decentralized multi-robot path planning,

    Q. Li, F. Gama, A. Ribeiro, and A. Prorok, “Graph neural networks for decentralized multi-robot path planning,” in2020 IEEE/RSJ inter- national conference on intelligent robots and systems (IROS). IEEE, 2020, pp. 11 785–11 792

    2020

  11. [19]

    Message-aware graph attention networks for large-scale multi-robot path planning,

    Q. Li, W. Lin, Z. Liu, and A. Prorok, “Message-aware graph attention networks for large-scale multi-robot path planning,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 5533–5540, 2021

    2021

  12. [20]

    Pairwise is not enough: Hypergraph neural networks for multi-agent pathfinding,

    R. Jain, K. Okumura, M. Amir, P. Lio, and A. Prorok, “Pairwise is not enough: Hypergraph neural networks for multi-agent pathfinding,”arXiv preprint arXiv:2602.06733, 2026

    2026 arXiv

  13. [21]

    Alpha: Attention- based long-horizon pathfinding in highly-structured areas,

    C. He, T. Yang, T. Duhan, Y . Wang, and G. Sartoretti, “Alpha: Attention- based long-horizon pathfinding in highly-structured areas,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 14 576–14 582

    2024

  14. [22]

    Mapf-gpt: Imitation learning for multi-agent pathfinding at scale,

    A. Andreychuk, K. Yakovlev, A. Panov, and A. Skrynnik, “Mapf-gpt: Imitation learning for multi-agent pathfinding at scale,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 22, 2025, pp. 23 126–23 134

    2025

  15. [23]

    Advancing learnable multi-agent pathfinding solvers with active fine-tuning,

    ——, “Advancing learnable multi-agent pathfinding solvers with active fine-tuning,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 10 564–10 571

    2025

  16. [24]

    Social behavior as a key to learning-based multi-agent pathfinding dilemmas,

    C. He, T. Duhan, P. Tulsyan, P. Kim, and G. Sartoretti, “Social behavior as a key to learning-based multi-agent pathfinding dilemmas,”Artificial Intelligence, p. 104397, 2025

    2025

  17. [25]

    Improving learnt local mapf policies with heuristic search,

    R. Veerapaneni, Q. Wang, K. Ren, A. Jakobsson, J. Li, and M. Likhachev, “Improving learnt local mapf policies with heuristic search,” inProceedings of the International Conference on Automated Planning and Scheduling, vol. 34, 2024, pp. 597–606

    2024

  18. [26]

    Deploying ten thousand robots: Scalable imitation learning for lifelong multi-agent path finding,

    H. Jiang, Y . Wang, R. Veerapaneni, T. Duhan, G. Sartoretti, and J. Li, “Deploying ten thousand robots: Scalable imitation learning for lifelong multi-agent path finding,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1–7

    2025

  19. [27]

    A review of graph-based multi-agent pathfinding solvers: From classical to beyond classical,

    J. Gao, Y . Li, X. Li, K. Yan, K. Lin, and X. Wu, “A review of graph-based multi-agent pathfinding solvers: From classical to beyond classical,”Knowledge-Based Systems, p. 111121, 2023

    2023

  20. [28]

    Icbs: The improved conflict-based search algorithm for multi-agent pathfinding,

    E. Boyarski, A. Felner, R. Stern, G. Sharon, O. Betzalel, D. Tolpin, and E. Shimony, “Icbs: The improved conflict-based search algorithm for multi-agent pathfinding,” inProceedings of the International Symposium on Combinatorial Search, vol. 6, no. 1, 2015, pp. 223–225

    2015

  21. [29]

    Adding heuristics to conflict-based search for multi-agent path finding,

    A. Felner, J. Li, E. Boyarski, H. Ma, L. Cohen, T. S. Kumar, and S. Koenig, “Adding heuristics to conflict-based search for multi-agent path finding,” inProceedings of the International Conference on Auto- mated Planning and Scheduling, vol. 28, 2018, pp. 83–87

    2018

  22. [30]

    Improved heuristics for multi-agent path finding with conflict-based search

    J. Li, A. Felner, E. Boyarski, H. Ma, and S. Koenig, “Improved heuristics for multi-agent path finding with conflict-based search.” inIJCAI, vol. 2019, 2019, pp. 442–449

    2019

  23. [31]

    Symmetry- breaking constraints for grid-based multi-agent path finding,

    J. Li, D. Harabor, P. J. Stuckey, H. Ma, and S. Koenig, “Symmetry- breaking constraints for grid-based multi-agent path finding,” inPro- ceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01, 2019, pp. 6087–6095

    2019

  24. [32]

    New techniques for pairwise symmetry breaking in multi-agent path finding,

    J. Li, G. Gange, D. Harabor, P. J. Stuckey, H. Ma, and S. Koenig, “New techniques for pairwise symmetry breaking in multi-agent path finding,” inProceedings of the International Conference on Automated Planning and Scheduling, vol. 30, 2020, pp. 193–201

    2020

  25. [33]

    Multi-agent path finding with mutex propagation,

    H. Zhang, J. Li, P. Surynek, T. S. Kumar, and S. Koenig, “Multi-agent path finding with mutex propagation,”Artificial Intelligence, vol. 311, p. 103766, 2022

    2022

  26. [34]

    Pairwise symmetry reasoning for multi-agent path finding search,

    J. Li, D. Harabor, P. J. Stuckey, H. Ma, G. Gange, and S. Koenig, “Pairwise symmetry reasoning for multi-agent path finding search,” Artificial Intelligence, vol. 301, p. 103574, 2021

    2021

  27. [35]

    Dis- joint splitting for multi-agent path finding with conflict-based search,

    J. Li, D. Harabor, P. J. Stuckey, A. Felner, H. Ma, and S. Koenig, “Dis- joint splitting for multi-agent path finding with conflict-based search,” inProceedings of the international conference on automated planning and scheduling, vol. 29, 2019, pp. 279–283

    2019

  28. [36]

    Adaptive-horizon conflict- based search for closed-loop multi-agent path finding,

    J. Li, F. Pecora, R. Zhang, and G. Zardini, “Adaptive-horizon conflict- based search for closed-loop multi-agent path finding,”arXiv preprint arXiv:2602.12024, 2026

    2026 arXiv

  29. [37]

    Cooperative pathfinding,

    D. Silver, “Cooperative pathfinding,” inProceedings of the aaai con- ference on artificial intelligence and interactive digital entertainment, vol. 1, no. 1, 2005, pp. 117–122

    2005

  30. [38]

    Priority inheri- tance with backtracking for iterative multi-agent path finding,

    K. Okumura, M. Machida, X. D ´efago, and Y . Tamura, “Priority inheri- tance with backtracking for iterative multi-agent path finding,”Artificial Intelligence, vol. 310, p. 103752, 2022

    2022

  31. [39]

    Fico: Finite- horizon closed-loop factorization for unified multi-agent path finding,

    J. Li, A. Zanardi, F. Pecora, R. Zhang, and G. Zardini, “Fico: Finite- horizon closed-loop factorization for unified multi-agent path finding,” arXiv preprint arXiv:2511.13961, 2025

    2025

  32. [40]

    Multi-agent path finding via finite- horizon hierarchical factorization,

    J. Li, A. Zanardi, and G. Zardini, “Multi-agent path finding via finite- horizon hierarchical factorization,”arXiv preprint arXiv:2505.07779, 2025

    2025 arXiv

  33. [41]

    Distributed heuristic multi-agent path finding with communication,

    Z. Ma, Y . Luo, and H. Ma, “Distributed heuristic multi-agent path finding with communication,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 8699–8705

    2021

  34. [42]

    Multi-agent path finding with prioritized communication learning,

    W. Li, H. Chen, B. Jin, W. Tan, H. Zha, and X. Wang, “Multi-agent path finding with prioritized communication learning,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 10 695–10 701

    2022

  35. [43]

    Scrimp: Scalable communication for reinforcement-and imitation-learning-based multi- agent pathfinding,

    Y . Wang, B. Xiang, S. Huang, and G. Sartoretti, “Scrimp: Scalable communication for reinforcement-and imitation-learning-based multi- agent pathfinding,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 9301–9308

    2023

  36. [44]

    Learning selective communication for multi- agent path finding,

    Z. Ma, Y . Luo, and J. Pan, “Learning selective communication for multi- agent path finding,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 1455–1462, 2021

    2021

  37. [45]

    Sigma: Sheaf-informed geometric multi-agent pathfinding,

    S. Liao, W. Xia, Y . Cao, W. Dai, C. He, W. Wu, and G. Sartoretti, “Sigma: Sheaf-informed geometric multi-agent pathfinding,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1–7

    2025

  38. [46]

    Subdimensional expansion using attention-based learning for multi-agent path finding,

    L. Virmani, Z. Ren, S. Rathinam, and H. Choset, “Subdimensional expansion using attention-based learning for multi-agent path finding,” arXiv preprint arXiv:2109.14695, 2021

    2021 arXiv

  39. [47]

    Adapting to planning failures in lifelong multi-agent path finding,

    J. Morag, R. Stern, and A. Felner, “Adapting to planning failures in lifelong multi-agent path finding,” inProceedings of the International Symposium on Combinatorial Search, vol. 16, no. 1, 2023, pp. 47–55

    2023

  40. [48]

    Work smarter not harder: Simple imitation learning with cs-pibt outperforms large-scale imitation learning for mapf,

    R. Veerapaneni, A. Jakobsson, K. Ren, S. Kim, J. Li, and M. Likhachev, “Work smarter not harder: Simple imitation learning with cs-pibt outperforms large-scale imitation learning for mapf,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp....

    2025

  41. [49]

    Lifelong path planning with kinematic constraints for multi-agent pickup and de- livery,

    H. Ma, W. H ¨onig, T. S. Kumar, N. Ayanian, and S. Koenig, “Lifelong path planning with kinematic constraints for multi-agent pickup and de- livery,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 7651–7658

    2019

  42. [50]

    Learn to follow: Decentralized lifelong multi-agent pathfinding via planning and learning,

    A. Skrynnik, A. Andreychuk, M. Nesterova, K. Yakovlev, and A. Panov, “Learn to follow: Decentralized lifelong multi-agent pathfinding via planning and learning,”arXiv preprint arXiv:2310.01207, 2023

    2023 arXiv

  43. [51]

    Primal 2: Pathfind- ing via reinforcement and imitation multi-agent learning-lifelong,

    M. Damani, Z. Luo, E. Wenzel, and G. Sartoretti, “Primal 2: Pathfind- ing via reinforcement and imitation multi-agent learning-lifelong,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2666–2673, 2021

    2021

  44. [52]

    Socialmapf: Optimal and efficient multi-agent path finding with strategic agents for social navigation,

    R. Chandra, R. Maligi, A. Anantula, and J. Biswas, “Socialmapf: Optimal and efficient multi-agent path finding with strategic agents for social navigation,”IEEE Robotics and Automation Letters, 2023

    2023

  45. [53]

    Online guidance graph optimization for lifelong multi-agent path finding,

    H. Zang, Y . Zhang, H. Jiang, Z. Chen, D. Harabor, P. J. Stuckey, and J. Li, “Online guidance graph optimization for lifelong multi-agent path finding,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 14, 2025, pp. 14 726–14 735

    2025

  46. [54]

    Learning-guided prioritized planning for lifelong multi-agent path finding in warehouse automation,

    H. Zheng, Y . Ma, B. Araki, J. Chen, and C. Wu, “Learning-guided prioritized planning for lifelong multi-agent path finding in warehouse automation,”Journal of Artificial Intelligence Research, vol. 85, 2026

    2026

  47. [55]

    Optimiza- tion of edge directions and weights for mixed guidance graphs in lifelong multi-agent path finding,

    Y . Zhang, V . Bhatt, M. C. Fontaine, S. Nikolaidis, and J. Li, “Optimiza- tion of edge directions and weights for mixed guidance graphs in lifelong multi-agent path finding,”arXiv preprint arXiv:2602.23468, 2026

    2026

  48. [56]

    Multi-robot mo- tion planning with diffusion models,

    Y . Shaoul, I. Mishani, S. Vats, J. Li, and M. Likhachev, “Multi-robot mo- tion planning with diffusion models,”arXiv preprint arXiv:2410.03072, 2024

    2024 arXiv

  49. [57]

    Collaborative multi-robot non-prehensile manipulation via flow- matching co-generation,

    Y . Shaoul, Z. Chen, M. N. G. Mohamed, F. Pecora, M. Likhachev, and J. Li, “Collaborative multi-robot non-prehensile manipulation via flow- matching co-generation,”arXiv preprint arXiv:2511.10874, 2025

    2025

  50. [58]

    Flow- based task assignment for large-scale online multi-agent pickup and delivery,

    Y . Zhang, Z. Chen, D. Harabor, P. L. Bodic, and P. J. Stuckey, “Flow- based task assignment for large-scale online multi-agent pickup and delivery,”arXiv preprint arXiv:2508.05890, 2025

    2025 arXiv

  51. [59]

    Lifelong multi-agent path finding for online pickup and delivery tasks,

    H. Ma, J. Li, T. Kumar, and S. Koenig, “Lifelong multi-agent path finding for online pickup and delivery tasks,”arXiv preprint arXiv:1705.10868, 2017. JOURNAL OF LATEX CLASS FILES, VOL. *, NO. *, JULY 2026 19

    2017 arXiv

  52. [60]

    Relational inductive biases, deep learning, and graph networks,

    P. W. Battaglia, J. B. Hamrick, V . Bapst, A. Sanchez-Gonzalez, V . Zam- baldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner et al., “Relational inductive biases, deep learning, and graph networks,” arXiv preprint arXiv:1806.01261, 2018

    2018 arXiv

  53. [61]

    Nagphormer: A tokenized graph transformer for node classification in large graphs,

    J. Chen, K. Gao, G. Li, and K. He, “Nagphormer: A tokenized graph transformer for node classification in large graphs,”arXiv preprint arXiv:2206.04910, 2022

    2022 arXiv

  54. [62]

    Theory and applications of hazard plotting for censored failure data,

    W. Nelson, “Theory and applications of hazard plotting for censored failure data,”Technometrics, vol. 14, no. 4, pp. 945–966, 1972

    1972

  55. [63]

    J. D. Kalbfleisch and R. L. Prentice,The statistical analysis of failure time data. John Wiley & Sons, 2002

    2002

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.