Pith. sign in

REVIEW 3 major objections 4 minor 54 references

Advancing Learnable Multi-Agent Pathfinding Solvers with Active Fine-Tuning

T0 review · 3 major / 4 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 Fine-tuning a 2M-parameter MAPF policy on its own failure states matches a model 50 times larger.

desk verdict The abstract's 'surpasses all' claim is contradicted by the paper's own figures, but the delta-selection idea is new and the 1M-agent demo is real; it needs a matched ablation and honest claims before acceptance. read the letter →

arxiv 2506.23793 v1 pith:J6HXY6Q3 submitted 2025-06-30 cs.AI cs.LGcs.MA

Anton Andreychuk , Konstantin Yakovlev , Aleksandr Panov , Alexey Skrynnik This is my paper
classification cs.AIcs.LGcs.MA
keywords multi-agentpathfindingimitationlearningactivefine-tuningdatasetaggregationdistributionalshifttransformerpolicyscalabilitydeltadatageneration
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

MAPF-GPT-DDG is a fine-tuned version of the MAPF-GPT imitation-learning solver. The paper introduces $\Delta$ Data Generation (DDG), a reward-free active fine-tuning method that detects the states where the learned policy performs worst and corrects them using a centralized high-quality solver. The authors claim that this lets a 2-million-parameter model match or exceed the original 85-million-parameter MAPF-GPT on several benchmarks and that the resulting solver handles instances with over a million agents. The central bet is that selectively correcting the most problematic states is enough to close most of the gap to a much larger model.

What carries the argument

$\Delta$ Data Generation (DDG), formalized in Algorithm 1, is the mechanism that carries the argument. It rolls out the current policy, samples every h-th state (h=16), runs a 2-second LaCAM* solver from each sampled state, and computes the cost difference δ_i = cost(sol_{i+1}) - cost(sol_i). The state with the largest δ is treated as the point where the policy underperforms; if δ exceeds a threshold δ_min=3, a 10-second LaCAM* run produces a high-quality solution, and the first k=32 observation-action pairs of that solution are added to the generated dataset. Fine-tuning interleaves minibatches from this generated data and the original expert dataset in a 1:3 ratio, so the correction data is injected without drifting away from the pretrained behavior.

What would settle it

Compare DDG against a control that selects the same number of correction states uniformly at random from the same trajectories, keeping store length, solver time, and fine-tuning schedule fixed. If the control matches DDG's success rate and loss curves, then the delta heuristic, not the volume of corrections, is doing the work.

Watch

Extended reading notes

Core claim

The paper claims that distributional shift in decentralized MAPF policies can be corrected by identifying, for each generated trajectory, the state at which the cost of a fast approximate solver jumps the most between successive sampled states, and then relabeling the next few actions after that state with a more accurate solver's solution. The fine-tuning mixes this generated data with the original expert dataset to avoid catastrophic forgetting. On the POGEMA benchmark, the resulting MAPF-GPT-DDG (2M parameters) matches or outperforms MAPF-GPT-85M on Random, Mazes, Warehouse, and Cities Tiles maps, while on Warehouse maps, SCRIMP retains the best success rate. The paper also reports solving an instance with 1,048,576 agents on a 2048x2048 map, with per-agent decision time around 160 microseconds.

Load-bearing premise

The load-bearing premise is that a large jump in the cost of a fast approximate solver between consecutive sampled states marks where the learned policy actually fails, so correcting that exact state is the highest-value use of expert solver time.

Editorial extensions

If this is right

  • A 2M-parameter model can reach the performance of an 85M-parameter model on MAPF benchmarks, cutting training GPU-hours roughly threefold.
  • DDG-style active fine-tuning can be applied on top of existing imitation-learned policies without a reward function, making it a general tool for distribution-shift repair.
  • Learnable MAPF solvers can scale to over a million agents on large open maps, with per-agent decision time around 160 microseconds.
  • The method's gains transfer from the training map types (Random, Mazes) to out-of-distribution types (Warehouse, Cities Tiles) for success rate and solution cost.
  • The ablation suggests that recording a short sequence of corrective actions (store length 32) may be essential, since DAgger, which records only one action, performs worse than default training.

Reading between the lines

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

  • If the delta heuristic truly identifies policy failure, then any cheap proxy for local solution-cost degradation, such as the policy's own uncertainty estimates, could replace the approximate solver and make the method cheaper.
  • The million-agent demonstration uses an empty 2048x2048 map with start-goal distances capped at 64; the scalability claim may not transfer to dense obstacle maps or long-horizon instances without further evidence.
  • Because the ablation against DAgger changes several factors at once (store length, solver time, and selection rule), a follow-up isolating the delta selection rule from the store length would settle what DDG actually contributes beyond collecting more data per query.
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

3 major / 4 minor

Summary. The paper proposes an active fine-tuning method, Delta Data Generation (DDG), for the pretrained MAPF-GPT imitation-learning solver. During data generation, the current policy rolls out in randomly generated instances; every h-th state is solved by a fast approximate LaCAM* solver, and the state with the largest cost jump delta is then solved by a slower accurate LaCAM* solver, whose first k observation-action pairs are added to a generated dataset Dg. The pretrained 2M-parameter MAPF-GPT is fine-tuned on minibatches that mix the original expert dataset De with Dg. The paper reports that the resulting MAPF-GPT-DDG-2M improves success rate and solution cost relative to MAPF-GPT-85M on some map types, is competitive with SCRIMP, DCC, and EPH on the POGEMA benchmark, and can be run on a 2048x2048 empty map with up to 1,048,576 agents.

Significance. If the results are taken at face value, the contribution is useful: it shows that a 2M-parameter imitation-learning MAPF policy can be improved to match a 50x larger model by targeted expert relabeling, and it introduces a cheap heuristic for focusing expert queries on states where the policy appears to fail. The paper uses a well-established benchmark (POGEMA), builds on publicly available pretrained weights, and reports confidence intervals for success rates, which supports reproducibility. However, the headline claim that MAPF-GPT-DDG 'surpasses all existing learning-based MAPF solvers' is not supported by the paper's own data, and the ablation does not isolate the proposed delta-selection rule. The method-specific contribution is therefore not yet established with the evidence presented.

major comments (3)
  1. [Abstract and Section VI.B] The abstract states that 'MAPF-GPT-DDG surpasses all existing learning-based MAPF solvers, including the original MAPF-GPT, regarding solution quality,' but the paper's own Fig. 2 shows SCRIMP with the best success rate on Warehouse maps, and on Random and Cities Tiles maps MAPF-GPT-DDG is on par with MAPF-GPT-85M rather than strictly better. The conclusion already retreats to 'in certain scenarios.' The central comparative claim should be reworded to match the evidence, e.g., per-map-type improvements and particular density regimes, or the missing comparisons must be supplied.
  2. [Section V.C and Algorithm 1] The ablation that is intended to demonstrate the advantage of DDG is confounded. In the DAgger branch, the store length is 1 and the expert solver is limited to 2 seconds, while in the DDG branch the store length is k=32 and the accurate solver is run for 10 seconds (Section V.D). Thus the comparison changes at least three factors at once: the selection rule (max-delta vs. no selection), the number of expert actions stored per query, and the quality of the expert solver. The observed gain could come entirely from longer expert rollouts or more solver time, without any contribution from delta-based selection. Please add an ablation that holds k and solver time fixed and compares max-delta selection against random selection; this is the load-bearing premise of DDG.
  3. [Section VI.D and Table I] The scalability claim needs calibration. Table I reports a 99.9% independent success rate at 1,048,576 agents, not a 100% solve rate, and all runs are on a single 2048x2048 empty map with start-goal distance limited to 64. The statement that MAPF-GPT-DDG 'can work with MAPF instances involving up to 1 million agents' should therefore be restricted to this narrow, easy distribution, and ideally accompanied by a baseline at the same scale; otherwise the claimed 'milestone' is not comparable to any existing solver.
minor comments (4)
  1. [Algorithm 1] The input parameter G is never used in the loop body, which runs over L instead; either remove G or use it in the termination condition.
  2. [Section V.D] Please specify the number of outer-loop phases L and the number of fine-tuning iterations per phase J explicitly, since the text reports 'one phase of fine-tuning run for 1,000 iterations' and 'an additional 340,000 iterations' without reconciling the two.
  3. [Related Work] The text mentions 'GT4-V and Gemini' in reference to multimodal models; this appears to be a typo for 'GPT-4V' and should be corrected.
  4. [Section VI.A] The SoC ratio plots in Fig. 3 do not include confidence intervals or significance tests; please state the number of instances per cell and indicate which observed differences are statistically reliable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DDG is an empirical fine-tuning procedure evaluated on external benchmarks, and the claimed improvements are not forced by construction.

full rationale

MAPF-GPT-DDG is a training procedure, not a derived theoretical result, so there is no equation chain in which a predicted quantity is identical to an input by construction. The fine-tuning target is the LaCAM* expert (an external anytime MAPF solver, reference [52]), and the model is evaluated on held-out instances from the POGEMA benchmark against independent baselines (SCRIMP, DCC, EPH) as well as the base MAPF-GPT models. Gains in success rate and solution-cost ratio are therefore measured against external data, not forced by the training objective. The main self-citations, MAPF-GPT [24] as the base model and POGEMA [54] as the benchmark and evaluation protocol, are appropriate reuse of the authors' own publicly available artifacts; they do not supply the load-bearing justification for the claimed improvement, because the improvement is an empirical result displayed in Figs. 2 and 3. The delta selection rule (Eq. 4) is asserted to identify states where the policy underperforms; this is an unvalidated heuristic, and the ablation in Section V.C confounds selection rule, store length, and expert time. However, an unvalidated or confounded heuristic is a correctness concern, not circularity, since the final evaluation is independent of the heuristic's truth. The abstract's 'surpasses all' wording is also contradicted by the paper's own Fig. 2 (SCRIMP is best on Warehouse, and MAPF-GPT-DDG is only on par with MAPF-GPT-85M on Random and Cities Tiles), but overclaiming is not a circular-derivation issue. No step reduces to its own input by definition or by a self-citation chain, so no circular steps are reported.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method is empirical and introduces no new physical or mathematical entities. The main free parameters are hyperparameters of the fine-tuning pipeline, chosen by hand and not tuned on a held-out set. The key domain assumption is that the delta of approximate solver costs identifies policy weakness.

free parameters (7)
  • delta threshold δmin = 3
    States with delta above this threshold are selected for accurate relabeling. Chosen in Section V-D to balance data collection time and training cost.
  • split length h = 16
    Trajectories are sampled every h-th state to form candidate states. Set in Section V-D.
  • store length k = 32
    Number of observation-action pairs taken from the accurate solver's solution for each selected state. Section V-D.
  • Bg:Be ratio = 1:3
    Proportion of generated to expert data in each fine-tuning mini-batch. Section V-D.
  • fine-tuning iterations per phase J = 1000
    Number of gradient steps per data generation cycle. Section V-D.
  • approximate solver time limit = 2 seconds
    LaCAM* time limit for cost estimation in delta computation. Section V-D.
  • accurate solver time limit = 10 seconds
    LaCAM* time limit for generating corrective data. Section V-D.
assumptions (4)
  • domain assumption A decentralized policy with local observations can effectively solve MAPF.
    The entire approach treats MAPF as a decentralized sequential decision-making problem, relying on local observations. Invoked throughout Section III and used to justify the model input representation.
  • domain assumption LaCAM* with 2s and 10s time limits provides adequate approximate and accurate expert behavior.
    The approximate solver's cost estimates drive state selection, and the accurate solver produces corrective data. Section V-D specifies these time limits without validating their quality relative to other choices.
  • domain assumption The delta between approximate solver costs at successive states is a proxy for policy weakness.
    Algorithm 1 selects the state with maximal δi as the policy's most problematic state. This heuristic is the core of DDG and is not independently tested.
  • domain assumption Fine-tuning on a mix of old expert data and new generated data prevents catastrophic forgetting.
    Section V-B mixes mini-batches from De and Dg to keep the policy near its initialization. This is a standard assumption in fine-tuning and is not verified for this specific setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Learnable Multi-Agent Pathfinding Solvers with Active Fine-Tuning." pith.science (2026). https://pith.science/paper/J6HXY6Q3

@misc{pith2026250623793,
  author       = {Pith},
  title        = {Pith review of: Advancing Learnable Multi-Agent Pathfinding Solvers with Active Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J6HXY6Q3}},
  note         = {Machine review of arXiv:2506.23793}
}
read the original abstract

Multi-agent pathfinding (MAPF) is a common abstraction of multi-robot trajectory planning problems, where multiple homogeneous robots simultaneously move in the shared environment. While solving MAPF optimally has been proven to be NP-hard, scalable, and efficient, solvers are vital for real-world applications like logistics, search-and-rescue, etc. To this end, decentralized suboptimal MAPF solvers that leverage machine learning have come on stage. Building on the success of the recently introduced MAPF-GPT, a pure imitation learning solver, we introduce MAPF-GPT-DDG. This novel approach effectively fine-tunes the pre-trained MAPF model using centralized expert data. Leveraging a novel delta-data generation mechanism, MAPF-GPT-DDG accelerates training while significantly improving performance at test time. Our experiments demonstrate that MAPF-GPT-DDG surpasses all existing learning-based MAPF solvers, including the original MAPF-GPT, regarding solution quality across many testing scenarios. Remarkably, it can work with MAPF instances involving up to 1 million agents in a single environment, setting a new milestone for scalability in MAPF domains.

Figures

Figures reproduced from arXiv: 2506.23793 by the authors.

Figure 1
Figure 1. The general pipeline of the proposed active fine-tuning approach — Delta-Driven Data Generation (DDG). The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Success rates of the approaches on different map types depending on the number of agents in the instances (higher [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. SoC ratio relative to solutions found by the LaCAM* approach (lower is better). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Validation loss and success rate on subset of instances [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 45 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,” in Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI 2021), 2021, pp. 11 272–11 281

    work page 2021

  2. [2]

    Intersection coordination with priority-based search for autonomous vehicles,

    J. Li, E. Lin, H. L. Vu, S. Koenig et al. , “Intersection coordination with priority-based search for autonomous vehicles,” in Proceedings of the 37th AAAI Conference on Artificial Intelligence (AAAI 2023) , 2023, pp. 11 578–11 585

    work page 2023

  3. [3]

    Multi-agent pathfinding: Definitions, variants, and benchmarks,

    R. Stern, N. R. Sturtevant, A. Felner, S. Koenig, H. Ma, T. T. Walker, J. Li, D. Atzmon, L. Cohen, T. S. Kumar et al., “Multi-agent pathfinding: Definitions, variants, and benchmarks,” in Proceedings of the 12th Annual Symposium on Combinatorial Search (SoCS 2019) , 2019, pp. 151–158

    work page 2019

  4. [4]

    Multi-agent path finding with kinematic constraints

    W. H ¨onig, T. S. Kumar, L. Cohen, H. Ma, H. Xu, N. Ayanian, and S. Koenig, “Multi-agent path finding with kinematic constraints.” in Proceedings of The 26th International Conference on Automated Planning and Scheduling (ICAPS 2016) , 2016, pp. 477–485

    work page 2016

  5. [5]

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

    H. Ma, W. H ¨onig, T. K. S. Kumar, N. Ayanian, and S. Koenig, “Lifelong path planning with kinematic constraints for multi-agent pickup and delivery,” in Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI 2019) , 2019, pp. 7651–7658

    work page 2019

  6. [6]

    Prioritized multi- agent path finding for differential drive robots,

    K. Yakovlev, A. Andreychuk, and V . V orobyev, “Prioritized multi- agent path finding for differential drive robots,” in Proceedings of the 2019 European Conference on Mobile Robots (ECMR 2019) , 2019, pp. 1–6

    work page 2019

  7. [7]

    Finding optimal solutions to cooperative pathfinding problems,

    T. S. Standley, “Finding optimal solutions to cooperative pathfinding problems,” in Proceedings of The 24th AAAI Conference on Artificial Intelligence (AAAI 2010) , 2010, pp. 173–178

    work page 2010

  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 54 references
  1. [9]

    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,” in Proceedings of The 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2011), 2011, pp. 3260–3267

    2011

  2. [10]

    Efficient sat approach to multi-agent path finding under the sum of costs objec- tive,

    P. Surynek, A. Felner, R. Stern, and E. Boyarski, “Efficient sat approach to multi-agent path finding under the sum of costs objec- tive,” in Proceedings of the 22nd European Conference on Artificial Intelligence (ECAI 2016) . IOS Press, 2016, pp. 810–818

    2016

  3. [11]

    Priority inheritance with backtracking for iterative multi-agent path finding,

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

    2022

  4. [12]

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

    K. Okumura, “Lacam: Search-based algorithm for quick multi-agent pathfinding,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023, pp. 11 655–11 662

    2023

  5. [13]

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

    J. Li, Z. Chen, D. Harabor, P. J. Stuckey, and S. Koenig, “Mapf- lns2: Fast repairing for multi-agent path finding via large neighbor- hood search,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 2022, pp. 10 256–10 265

    2022

  6. [14]

    An optimization variant of multi-robot path planning is intractable,

    P. Surynek, “An optimization variant of multi-robot path planning is intractable,” in Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI 2010) , 2010, pp. 1261–1263

    2010

  7. [15]

    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,” in Proceedings of the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2020)....

    2020

  8. [16]

    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

  9. [17]

    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,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 9301–9308

    2023

  10. [18]

    Distributed heuristic multi-agent path finding with communication,

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

    2021

  11. [19]

    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

  12. [20]

    Ensembling prioritized hybrid policies for multi-agent pathfinding,

    H. Tang, F. Berto, and J. Park, “Ensembling prioritized hybrid policies for multi-agent pathfinding,” in 2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 8047–8054

    2024

  13. [21]

    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,” in Proceedings of the 38th AAAI Conference on Artificial Intelligence (AAAI 2024) , 2024

    2024

  14. [22]

    When to switch: planning and learning for partially observable multi-agent pathfinding,

    A. Skrynnik, A. Andreychuk, K. Yakovlev, and A. I. Panov, “When to switch: planning and learning for partially observable multi-agent pathfinding,” IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  15. [23]

    A comprehensive review on lever- aging machine learning for multi-agent path finding,

    J.-M. Alkazzi and K. Okumura, “A comprehensive review on lever- aging machine learning for multi-agent path finding,” IEEE Access , 2024

    2024

  16. [24]

    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,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 23 126–23 134

    2025

  17. [25]

    Multi-agent imitation learning: Value is easy, regret is hard,

    J. Tang, G. Swamy, F. Fang, and S. Wu, “Multi-agent imitation learning: Value is easy, regret is hard,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

    2024

  18. [26]

    Learning multi-agent behaviors from distributed and streaming demonstrations,

    S. Liu and M. Zhu, “Learning multi-agent behaviors from distributed and streaming demonstrations,” Advances in Neural Information Pro- cessing Systems, vol. 36, 2024

    2024

  19. [27]

    Bayesian multi- type mean field multi-agent imitation learning,

    F. Yang, A. Vereshchaka, C. Chen, and W. Dong, “Bayesian multi- type mean field multi-agent imitation learning,” Advances in Neural Information Processing Systems , vol. 33, pp. 2469–2478, 2020

    2020

  20. [28]

    Multi-agent generative adversarial imitation learning,

    J. Song, H. Ren, D. Sadigh, and S. Ermon, “Multi-agent generative adversarial imitation learning,” Advances in neural information pro- cessing systems, vol. 31, 2018

    2018

  21. [29]

    Gailpg: Multi-agent policy gradient with generative adversarial imitation learning,

    W. Li, S. Huang, Z. Qiu, and A. Song, “Gailpg: Multi-agent policy gradient with generative adversarial imitation learning,” IEEE Trans- actions on Games , 2024

    2024

  22. [30]

    Multi-agent imitation learning with copulas,

    H. Wang, L. Yu, Z. Cao, and S. Ermon, “Multi-agent imitation learning with copulas,” in Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13–17, 2021, Proceedings, Part I 21 . Springer, 2021, pp...

    2021

  23. [31]

    Conditional imitation learning for multi-agent games,

    A. Shih, S. Ermon, and D. Sadigh, “Conditional imitation learning for multi-agent games,” in 2022 17th ACM/IEEE International Conference on Human-Robot Interaction (HRI) . IEEE, 2022, pp. 166–175

    2022

  24. [32]

    Coordinated multi-agent imitation learning,

    H. M. Le, Y . Yue, P. Carr, and P. Lucey, “Coordinated multi-agent imitation learning,” in International Conference on Machine Learning. PMLR, 2017, pp. 1995–2003

    2017

  25. [33]

    Offline pre-trained multi-agent decision transformer,

    L. Meng, M. Wen, C. Le, X. Li, D. Xing, W. Zhang, Y . Wen, H. Zhang, J. Wang, Y . Yang et al. , “Offline pre-trained multi-agent decision transformer,” Machine Intelligence Research, vol. 20, no. 2, pp. 233– 248, 2023

    2023

  26. [34]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V . Panneershelvam, M. Lanctot et al. , “Mastering the game of go with deep neural networks and tree search,” nature, vol. 529, no. 7587, pp. 484–489, 2016

    2016

  27. [35]

    Amortized planning with large-scale transformers: A case study on chess,

    A. Ruoss, G. Deletang, S. Medapati, J. Grau-Moya, L. K. Wenliang, E. Catt, J. Reid, C. A. Lewis, J. Veness, and T. Genewein, “Amortized planning with large-scale transformers: A case study on chess,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  28. [36]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill et al. , “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  29. [37]

    Foundation models for decision making: Problems, methods, and opportunities,

    S. Yang, O. Nachum, Y . Du, J. Wei, P. Abbeel, and D. Schuurmans, “Foundation models for decision making: Problems, methods, and opportunities,” arXiv preprint arXiv:2303.04129 , 2023

    2023 arXiv

  30. [38]

    Foundation models in robotics: Applications, challenges, and the future,

    R. Firoozi, J. Tucker, S. Tian, A. Majumdar, J. Sun, W. Liu, Y . Zhu, S. Song, A. Kapoor, K. Hausman et al. , “Foundation models in robotics: Applications, challenges, and the future,” The International Journal of Robotics Research , 2023

    2023

  31. [39]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu et al., “Octo: An open-source generalist robot policy,” arXiv preprint arXiv:2405.12213 , 2024

    2024 arXiv

  32. [40]

    OpenVLA: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “OpenVLA: An open-source vision-language-action model,” in 8th Annual C...

    2024

  33. [41]

    Magentic- one: A generalist multi-agent system for solving complex tasks,

    A. Fourney, G. Bansal, H. Mozannar, C. Tan, E. Salinas, F. Niedtner, G. Proebsting, G. Bassman, J. Gerrits, J. Alber et al. , “Magentic- one: A generalist multi-agent system for solving complex tasks,” arXiv preprint arXiv:2411.04468, 2024

    2024 arXiv

  34. [42]

    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,”arXiv preprint arXiv:2409.14491, 2024

    2024 arXiv

  35. [43]

    Multi-agent systems and foundation models enable autonomous supply chains: Opportunities and challenges,

    L. Xu, S. Almahri, S. Mak, and A. Brintrup, “Multi-agent systems and foundation models enable autonomous supply chains: Opportunities and challenges,” IFAC-PapersOnLine, vol. 58, no. 19, pp. 795–800, 2024

    2024

  36. [44]

    Fine-tuning multi- modal transformer models for generating actions in virtual and real environments,

    A. Staroverov, A. S. Gorodetsky, A. S. Krishtopik, U. A. Izmesteva, D. A. Yudin, A. K. Kovalev, and A. I. Panov, “Fine-tuning multi- modal transformer models for generating actions in virtual and real environments,” Ieee Access, vol. 11, pp. 130 548–130 559, 2023

    2023

  37. [45]

    Archer: Training language model agents via hierarchical multi-turn rl,

    Y . Zhou, A. Zanette, J. Pan, S. Levine, and A. Kumar, “Archer: Training language model agents via hierarchical multi-turn rl,” in International Conference on Machine Learning . PMLR, 2024, pp. 62 178–62 209

    2024

  38. [46]

    Fine-tuning large vision-language models as decision-making agents via reinforcement learning,

    S. Zhai, H. Bai, Z. Lin, J. Pan, P. Tong, Y . Zhou, A. Suhr, S. Xie, Y . LeCun, Y . Ma et al. , “Fine-tuning large vision-language models as decision-making agents via reinforcement learning,” Advances in Neural Information Processing Systems, vol. 37, pp. 110 935–110 971, 2025

    2025

  39. [47]

    Robot utility models: General policies for zero-shot deployment in new environments,

    H. Etukuru, N. Naka, Z. Hu, S. Lee, J. Mehu, A. Edsinger, C. Paxton, S. Chintala, L. Pinto, and N. M. M. Shafiullah, “Robot utility models: General policies for zero-shot deployment in new environments,”arXiv preprint arXiv:2409.05865, 2024

    2024 arXiv

  40. [48]

    The in- creasing cost tree search for optimal multi-agent pathfinding,

    G. Sharon, R. Stern, and A. Goldenberg, Meir aand Felner, “The in- creasing cost tree search for optimal multi-agent pathfinding,”Artificial intelligence, vol. 195, pp. 470–495, 2013

    2013

  41. [49]

    Searching with consistent prioritization for multi-agent path finding,

    H. Ma, D. Harabor, P. J. Stuckey, J. Li, and S. Koenig, “Searching with consistent prioritization for multi-agent path finding,” in Proceedings of the AAAI conference on artificial intelligence , vol. 33, 2019, pp. 7643–7650

    2019

  42. [50]

    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

  43. [51]

    Amortized planning with large-scale transformers: A case study on chess,

    A. Ruoss, G. Del ´etang, S. Medapati, J. Grau-Moya, K. Li, E. Catt, J. Reid, C. Lewis, J. Veness, and T. Genewein, “Amortized planning with large-scale transformers: A case study on chess,” Advances in Neural Information Processing Systems , vol. 37, pp. 65 765–65 790, 2025

    2025

  44. [52]

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

    K. Okumura, “Engineering lacam*: Towards real-time, large-scale, and near-optimal multi-agent pathfinding,” in Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems, 2024, pp. 1501–1509

    2024

  45. [53]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proceedings of the fourteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2011, ...

    2011

  46. [54]

    Pogema: A benchmark platform for cooperative multi-agent pathfinding,

    A. Skrynnik, A. Andreychuk, A. Borzilov, A. Chernyavskiy, K. Yakovlev, and A. Panov, “Pogema: A benchmark platform for cooperative multi-agent pathfinding,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

Pith tools

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