Pith. sign in

REVIEW 4 major objections 5 minor 41 references

AT-Drone: Benchmarking Adaptive Teaming in Multi-Drone Pursuit

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read AT-Drone is a new benchmark that lets researchers train and evaluate drone teams that must coordinate with partners they have never worked with before, and backs it with a real Crazyflie deployment pipeline.

desk verdict Useful benchmark artifact for adaptive teaming in drone pursuit, but the 'unseen teammate' evaluation is too in-family to back the headline claim. read the letter →

arxiv 2502.09762 v2 pith:ZSSB7K5T submitted 2025-02-13 cs.RO cs.AI

Yang Li , Junfan Chen , Feng Xue , Jiabin Qiu , Wenbin Li , Qingrui Zhang , Ying Wen , Wei Pan This is my paper · ORCID
classification cs.ROcs.AI
keywords adaptiveteamingmulti-dronepursuitzero-shotcoordinationad-hocteamworkbenchmarkmulti-agentreinforcementlearningCrazyfliedronesunseendronezoo
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

AT-Drone is presented as the first benchmark built specifically for adaptive teaming in multi-drone pursuit: training a team of pursuers to work with partners they have never coordinated with before. The paper's claim is that it closes the gap between adaptive teaming studies, which have mostly lived in discrete video games such as Overcooked and Hanabi, and real multi-robot systems that operate with continuous actions. To do this, AT-Drone provides four configurable pursuit environments, an algorithm zoo with seven learning and rule-based methods, and standardized evaluation protocols built around three 'unseen drone zoo' configurations that supply unfamiliar partner behaviors. It also includes a real-world deployment pipeline using Crazyflie drones and edge computing, with a case study in the hardest environment. If the claim holds, the benchmark gives the research community a common testbed for adaptive teaming in continuous, physically grounded multi-agent problems.

What carries the argument

The load-bearing object is the 'unseen drone zoo,' a set of three teammate behaviors—Greedy, VICSEK, and IPPO self-play—packaged into three evaluation configurations. It is what makes the 'adaptive' part testable: learners must generalize to partner policies they never trained with. The benchmark also relies on the AT-Dec-POMDP formulation and the distributed training framework with self-play and PBT, plus the HOLA-Drone V2 max-min preference oracle and NAHT-D teammate-modelling network. Each component is needed: the zoos define the distribution shift, the algorithm zoo supplies baselines, and the metrics quantify coordination quality.

What would settle it

Pair a policy trained on AT-Drone with teammates built independently by another research group (or by human teleoperation) in the same 4p2e3o and 4p3e5o environments; if its success rate no longer beats a policy trained only against the provided zoos, the claim that AT-Drone measures adaptive teaming to unfamiliar partners is falsified.

Watch

Extended reading notes

Core claim

On its own terms, AT-Drone's central claim is that multi-drone pursuit can be turned into a repeatable benchmark for adaptive teaming, and that its environment configurator, algorithm zoo, unseen drone zoos, and real-world deployment pipeline together make this possible. The paper defines adaptive teaming in this setting as an AT-Dec-POMDP where N learners must cooperate with M previously unseen partners to capture K evaders. It then demonstrates the benchmark across four increasingly hard environments (4p2e3o, 4p2e1o, 4p2e5o, and 4p3e5o) and three unseen-partner configurations, reporting success rate, collision rate, average success timesteps, and reward. The experiments show that HOLA-Drone V2 generally outperforms SP and PBT in zero-shot coordination, while the NAHT-D teammate-modelling variant's decoder helps in simpler scenarios but can hurt in the hardest one.

Load-bearing premise

The evaluation presumes that the three predefined unseen drone zoos — a Greedy pursuer, a VICSEK swarm-style follower, and IPPO self-play policies — stand in for the range of unfamiliar teammates a real drone would meet; if real-world partners behave differently, the benchmark's success rates may not predict field coordination.

Editorial extensions

If this is right

  • Researchers get a common ground to compare zero-shot coordination and ad-hoc teamwork algorithms on continuous-action, multi-pursuer multi-evader tasks, rather than only discrete games like Overcooked and Hanabi.
  • The deployment pipeline lets simulation-trained policies be tested on physical Crazyflie drones, so claims about real-world transfer are checkable.
  • The four environments and three zoos provide standardized difficulty levels, so progress can be measured as a single success-rate number.
  • The HOLA-Drone V2 and NAHT-D results give direct baselines for future methods, including which scenario complexities favor simpler teammate modelling.

Reading between the lines

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

  • The benchmark could be extended to include partners trained by independent institutions or human teleoperators, which would strengthen the claim that the zoos measure true adaptivity rather than overfitting to a fixed teammate distribution.
  • The finding that the NAHT-D decoder hurts in the hardest environment suggests a trade-off between model capacity and robustness; future work might test whether an adaptive gating mechanism for teammate modelling would improve performance.
  • Because the same IPPO family is used for both self-play zoo generation and learner training, the benchmark's difficulty could be underestimated relative to truly foreign teammates; this is a testable hypothesis.
  • The configurator could be pointed at new tasks beyond pursuit—such as coverage or search-and-rescue—without changing the core evaluation protocol, which would test whether the benchmark structure generalizes.
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

4 major / 5 minor

Summary. AT-Drone is presented as the first benchmark specifically for adaptive teaming in multi-drone pursuit. The paper contributes a configurable Gymnasium-style simulator with four pursuit scenarios, a distributed training framework with an algorithm zoo (SP, PBT, HOLA-Drone V2 and an ablation, MAPPO, NAHT-D and an ablation), three 'unseen drone zoo' teammate configurations, four evaluation metrics, and a Crazyflie/Jetson deployment pipeline. Experiments report success rate, collision rate, timesteps, and reward across the scenarios and zoos, together with a real-world case study.

Significance. The benchmark fills a genuine gap: existing adaptive-teaming benchmarks are mostly discrete-action games such as Overcooked, Hanabi, and SMAC, whereas multi-drone pursuit is continuous, safety-constrained, and directly relevant to field robotics. The paper's strongest assets are reproducibility (code, weights, and videos are promised), a standardized environment configurator, and a real-world deployment pipeline on Crazyflie drones with edge devices. If the evaluation protocol were validated, AT-Drone would be a useful community testbed. The authors' Section 6 limitations are candid, but the central validity threat is not addressed there: the 'unseen' partners used for evaluation are all constructed within the same in-house training family, so the benchmark has not yet been shown to measure adaptation to genuinely unfamiliar teammates.

major comments (4)
  1. [Section 3.3 and Appendix B] The 'unseen drone zoos' do not provide genuinely unfamiliar teammates, which is the construct the benchmark is built to measure. Zoo 1 is a hand-coded greedy policy, Zoo 2 consists of two IPPO self-play policies, and Zoo 3 is a random mixture of the others; Appendix B states that the learned zoo members are trained with 'an IPPO-based self-play algorithm,' and Section 3.3 states that the SP, PBT, and HOLA-Drone (V2) learners are also IPPO-based. Thus the learned partners the learners meet come from the same training family as the learners, differing only by seeds and skill level. 'Unseen' therefore means held-out seeds within one algorithm family, not previously unseen teammate types. Since Definition 3.1 defines adaptive teaming by collaboration with previously unseen partners, the current protocol does not establish the validity of AT-Drone as an adaptive-teaming benchmark. I recommend adding an out-of-family held-out zoo, for example partners trained with a different MARL algorithm or reward function, or independent heuristic controllers, and reporting a per-partner success matrix; until then, the term 'unseen' should be qualified as 'held-out within family.'
  2. [Section 4, Table 2, and Figure 3] The headline comparisons are not supported by statistical tests. All results are given as mean ± std, but no number of seeds or trials, no paired tests, and no effect sizes are reported, while the text makes categorical claims such as HOLA-Drone (V2) 'consistently demonstrates superior performance' and NAHT-D 'outperforms' MAPPO. In the reported point estimates the ablations are often not worse than the full methods: HOLA-Drone (V2) w/o g is comparable or better on 4p2e1o, and NAHT-D w/o Dec is better on 4p2e5o and 4p3e5o. These claims are therefore not merely unsupported; they appear contradicted by the paper's own tables and figures. The benchmark should report statistical significance for the central comparisons, or the conclusions should be restricted to descriptive observations.
  3. [Figure 2 and Section 4] The red dotted best-response lines are not a valid upper bound for adaptive teaming as defined in Definition 3.1. The text says these policies are 'specifically trained for each unseen teammate zoo,' which means the BR policy is trained with knowledge of the exact test-time partner distribution and measures the performance of a specialized policy, not a zero-shot-coordination ceiling. Moreover, a BR trained on the same in-house zoo cannot serve as evidence that the benchmark's 'unseen' partners are genuinely unfamiliar. I suggest either removing the BR lines from the comparison or renaming them as an 'in-distribution specialization upper bound'; if they are kept, a second reference such as a learner given the partner identity as additional input should also be reported.
  4. [Section 6] The limitations section candidly states that the real-world system is simple, the fleet is small, and perception is not handled; I agree with these qualifications, and they should be carried into the abstract and conclusion. The abstract says comprehensive experiments 'confirm AT-Drone's effectiveness,' but the experiments are mostly simulation-based with a single qualitative real-world case study; the real-world results should be described as a feasibility demonstration rather than validation of adaptive teaming performance.
minor comments (5)
  1. [Appendix A] The term 'AT-MDP framework' is not introduced anywhere else in the paper and should be made consistent with 'AT-Drone.'
  2. [Table 3] The row labeled 'Entropy coefficient(ϵclip)' uses a clip-ratio symbol; this should be split into two distinct hyperparameters, the PPO clip ratio and the entropy coefficient.
  3. [Appendix E] Appendix E says Unseen Zoo 3 'consists entirely of PPO-based self-play policies,' but Section 3.3 defines Zoo 3 as a random mixture of Greedy, VICSEK, and self-play drones; these statements need to be reconciled.
  4. [Figure 5] In the JSON example, `task_name` is assigned the bare value `4p2e1o` without quotes, so the example is not syntactically valid JSON.
  5. [Section 3.1] The definition sentence contains a duplicated 'where: where,' and the symbol C is used both for the set of cooperative agents and for the team size, which is confusing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: AT-Drone's benchmark contributions are empirical and self-contained; the in-house 'unseen drone zoo' limits external validity but does not make any evaluation result true by construction.

full rationale

AT-Drone is a benchmark and experimental paper rather than a formal derivation: the central contributions are an environment configurator, a deployment pipeline, an algorithm zoo, and an evaluation protocol. The reported comparisons (SUC, COL, AST, REW) are measured outcomes of trained policies; no quantity is obtained by solving an equation that already contains the conclusion. The nearest candidate for circularity is the 'unseen drone zoo': Appendix B says the learned zoo members are 'IPPO-based self-play' policies, and Section 3.3 says SP and PBT learners are also IPPO-based, so evaluation partners are in-house and same-family rather than independently produced. This is a legitimate benchmark-design and external-validity weakness, but it is not circular: performance on these zoos is not forced by construction, the best-response lines are explicitly reference points 'specifically trained for each unseen teammate zoo,' and the real-world Crazyflie experiments provide independent evidence. HOLA-Drone V2 builds on the authors' earlier HOLA-Drone [24] preference hypergraph; that self-citation supplies an algorithmic component, but the paper's own ablations and physical experiments carry the evaluation, so the self-citation is not load-bearing. No step of the paper's argument reduces to its inputs by definition.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on multiple hand-chosen environment parameters (capture range, speed limits, horizon) and on domain assumptions about simulation fidelity and the representativeness of the in-house unseen zoos. No new physical entities are introduced; the algorithmic innovations (preference hypergraph oracle, NAHT-D decoder) are framework-level constructs whose benefit is empirically mixed.

free parameters (8)
  • capture_range = 0.2 m
    Sets success threshold for capturing evaders; directly defines SUC metric. Chosen by authors, not justified by external benchmark.
  • safe_radius = 0.1 m
    Defines collision threshold for drone-obstacle contact; affects COL metric.
  • reception_range = 2 (units not specified)
    Observation radius for agents; determines what information policies receive.
  • velocity_p = 0.3
    Pursuer speed in simulator; relative speed vs evaders affects task difficulty.
  • velocity_e = 0.6
    Evader speed; faster than pursuers, making pursuit possible only through coordination.
  • task_horizon = 100 timesteps
    Episode length; longer horizon would change success rates.
  • fps = 10
    Simulation frame rate; affects control frequency and action semantics.
  • self-play drone skill levels = 70% and 54% success rate
    Hand-selected thresholds to create two distinct unseen teammate skill levels in Unseen Zoo 2.
assumptions (5)
  • domain assumption The planar kinematic model with global motion-capture state and angular steering action captures the essential dynamics of real Crazyflie pursuit (Section 3.2 Simulation and Deployment).
    The real-world deployment provides only a single case study, so the fidelity of the simulation model is not systematically validated.
  • ad hoc to paper The three predefined unseen drone zoos (Greedy, VICSEK, Self-Play IPPO) represent the space of unfamiliar teammates an agent will encounter (Section 3.3 Evaluation).
    All zoos are constructed in-house; no external or human evaluation shows they cover realistic teammate diversity.
  • domain assumption Learning policies with IPPO, self-play, and PBT will produce adaptive teaming behavior (Section 3.3 Training).
    The paper assumes these standard RL recipes suffice; no comparison to non-learning or optimization-based pursuit methods is provided.
  • ad hoc to paper The preference hypergraph max-min oracle in HOLA-Drone V2 improves worst-case coordination (Appendix C, Eqs. 6-7).
    The oracle's benefit is shown only as an ablation vs. w/o g; the theoretical justification for preference centrality as a coordination measure is asserted, not derived.
  • ad hoc to paper NAHT-D's autoencoder teammate model reconstructing action distributions captures enough behavior to aid coordination (Appendix D).
    Ablations often show NAHT-D w/o Dec outperforms full NAHT-D, undermining this premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AT-Drone: Benchmarking Adaptive Teaming in Multi-Drone Pursuit." pith.science (2026). https://pith.science/paper/ZSSB7K5T

@misc{pith2026250209762,
  author       = {Pith},
  title        = {Pith review of: AT-Drone: Benchmarking Adaptive Teaming in Multi-Drone Pursuit},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSSB7K5T}},
  note         = {Machine review of arXiv:2502.09762}
}
read the original abstract

Adaptive teaming-the capability of agents to effectively collaborate with unfamiliar teammates without prior coordination-is widely explored in virtual video games but overlooked in real-world multi-robot contexts. Yet, such adaptive collaboration is crucial for real-world applications, including border surveillance, search-and-rescue, and counter-terrorism operations. To address this gap, we introduce AT-Drone, the first dedicated benchmark explicitly designed to facilitate comprehensive training and evaluation of adaptive teaming strategies in multi-drone pursuit scenarios. AT-Drone makes the following key contributions: (1) An adaptable simulation environment configurator that enables intuitive and rapid setup of adaptive teaming multi-drone pursuit tasks, including four predefined pursuit environments. (2) A streamlined real-world deployment pipeline that seamlessly translates simulation insights into practical drone evaluations using edge devices and Crazyflie drones. (3) A novel algorithm zoo integrated with a distributed training framework, featuring diverse algorithms explicitly tailored, for the first time, to multi-pursuer and multi-evader settings. (4) Standardized evaluation protocols with newly designed unseen drone zoos, explicitly designed to rigorously assess the performance of adaptive teaming. Comprehensive experimental evaluations across four progressively challenging multi-drone pursuit scenarios confirm AT-Drone's effectiveness in advancing adaptive teaming research. Real-world drone experiments further validate its practical feasibility and utility for realistic robotic operations. Videos, code and weights are available at \url{https://sites.google.com/view/at-drone}.

Figures

Figures reproduced from arXiv: 2502.09762 by the authors.

Figure 1
Figure 1. Overview of the AT-Drone Benchmark, comprising four key components: (I) a customizable [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Success rate (SUC) across different difficulty levels for adaptive teaming without teammate [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of adaptive teaming with teammate modeling across environments [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of four multi-drone pursuit environments in real world. The environments [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: An example of environment configuration file. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: An example of a hypergraph representation (left) and its corresponding preference hyper [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Overview of our proposed HOLA-Drone (V2) algorithm. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Case Study: This example demonstrates the capture strategy executed by NAHT-D learners and unseen drone partners from unseen zoo 3 in the superhard environment 4p3e5o. The red circles denote pursuers, and the black squares represent evaders. In this scenario, four purs…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [1]

    T. H. Chung, G. A. Hollinger, and V . Isler. Search and pursuit-evasion in mobile robotics: A survey. Autonomous robots, 31:299–316, 2011

    work page 2011

  2. [2]

    Zhang, D

    Z. Zhang, D. Zhang, Q. Zhang, W. Pan, and T. Hu. DACOOP-A: Decentralized adaptive cooperative pursuit via attention. IEEE Robotics and Automation Letters , PP:1–8, 11 2023. doi:10.1109/LRA.2023.3331886

    work page arXiv 2023

  3. [3]

    J. P. Queralta, J. Taipalmaa, B. C. Pullinen, V . K. Sarker, T. N. Gia, H. Tenhunen, M. Gabbouj, J. Raitoharju, and T. Westerlund. Collaborative multi-robot search and rescue: Planning, coordination, perception, and active vision. Ieee Access, 8:191617–191643, 2020

    work page 2020

  4. [4]

    Shah and M

    K. Shah and M. Schwager. Multi-agent Cooperative Pursuit-Evasion Strategies Under Uncer- tainty. In N. Correll, M. Schwager, and M. Otte, editors, Distributed Autonomous Robotic Systems, pages 451–468, Cham, 2019. Springer International Publishing. ISBN 978-3-030- 05816-6. doi:10.1007/978-3-030-05816-6 32

    work page doi:10.1007/978-3-030-05816-6 2019

  5. [5]

    Group chasing tactics: how to catch a faster prey?

    M. Janosov, C. Vir´agh, G. V ´as´arhelyi, and T. Vicsek. Group chasing tactics: how to catch a faster prey? New Journal of Physics , 19(5):053003, May 2017. ISSN 1367-2630. doi: 10.1088/1367-2630/aa69e7. URL http://arxiv.org/abs/1701.00284. arXiv:1701.00284 [physics]

    work page Pith review arXiv 2017

  6. [6]

    Z. Zhou, W. Zhang, J. Ding, H. Huang, D. M. Stipanovi´c, and C. J. Tomlin. Cooperative pursuit with V oronoi partitions. Automatica, 72:64–72, Oct. 2016. ISSN 0005-1098. doi:10.1016/ j.automatica.2016.05.007. URL https://www.sciencedirect.com/science/article/ pii/S0005109816301911

    work page 2016

  7. [7]

    J. Chen, G. Li, C. Yu, X. Yang, B. Xu, H. Yang, and Y . Wang. A dual curriculum learning framework for multi-uav pursuit-evasion in diverse environments, 2024. URLhttps://arxiv. org/abs/2312.12255. 9

    work page Pith review arXiv 2024

  8. [8]

    S. Qi, X. Huang, P. Peng, X. Huang, J. Zhang, and X. Wang. Cascaded Attention: Adap- tive and Gated Graph Attention Network for Multiagent Reinforcement Learning. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3769–3779, Mar. 2024. ISSN 2162-2388. doi:10.1109/TNNLS.2022.3197918. URL https://ieeexplore.ieee.org/ abstract/document/9913678....

    work page arXiv 2024

Show all 41 references
  1. [9]

    de Souza, R

    C. de Souza, R. Newbury, A. Cosgun, P. Castillo, B. Vidolov, and D. Kuli´c. Decentralized Multi- Agent Pursuit Using Deep Reinforcement Learning. IEEE Robotics and Automation Letters, 6 (3):4552–4559, July 2021. ISSN 2377-3766. doi:10.1109/LRA.2021.3068952. URL https:// ieeexp...

    2021

  2. [10]

    S. Li, Y . Wu, X. Cui, H. Dong, F. Fang, and S. Russell. Robust Multi-Agent Reinforcement Learning via Minimax Deep Deterministic Policy Gradient.Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):4213–4220, July 2019. ISSN 2374-3468. doi:10.1609/aaai. v33i0...

    2019 doi

  3. [11]

    Matignon, G

    L. Matignon, G. J. Laurent, and N. Le Fort-Piat. Hysteretic Q-learning : an algorithm for Decentralized Reinforcement Learning in Cooperative Multi-Agent Teams. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 64–69, Oct. 2007. doi:10. 1109/IR...

    2007

  4. [12]

    other-play

    H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “other-play” for zero-shot coordination. In International Conference on Machine Learning, pages 4399–4410. PMLR, 2020

    2020

  5. [13]

    Stone, G

    P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Col- laboration without pre-coordination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1504–1509, 2010

    2010

  6. [14]

    Y . Li, S. Zhang, J. Sun, Y . Du, Y . Wen, X. Wang, and W. Pan. Cooperative open-ended learning framework for zero-shot coordination. In International Conference on Machine Learning, pages 20470–20484. PMLR, 2023

    2023

  7. [15]

    X. Wang, S. Zhang, W. Zhang, W. Dong, J. Chen, Y . Wen, and W. Zhang. Zsc-eval: An evaluation toolkit and benchmark for multi-agent zero-shot coordination. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024

    2024

  8. [16]

    Lucas and R

    K. Lucas and R. E. Allen. Any-play: An intrinsic augmentation for zero-shot coordination. In P. Faliszewski, V . Mascardi, C. Pelachaud, and M. E. Taylor, editors, 21st International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2022, Auckland, New Zealand, May...

    2022

  9. [17]

    Canaan, X

    R. Canaan, X. Gao, J. Togelius, A. Nealen, and S. Menzel. Generating and adapting to diverse ad hoc partners in hanabi. IEEE Transactions on Games, 15(2):228–241, 2022

    2022

  10. [18]

    N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V . Dumoulin, S. Moitra, E. Hughes, et al. The hanabi challenge: A new frontier for ai research. Artificial Intelligence, 280:103216, 2020

    2020

  11. [19]

    Barrett, P

    S. Barrett, P. Stone, and S. Kraus. Empirical evaluation of ad hoc teamwork in the pursuit domain. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 2, pages 567–574, 2011. 10

    2011

  12. [20]

    Papoudakis, F

    G. Papoudakis, F. Christianos, and S. Albrecht. Agent modelling under partial observability for deep reinforcement learning. Advances in Neural Information Processing Systems , 34: 19210–19222, 2021

    2021

  13. [21]

    C. Wang, M. A. Rahman, I. Durugkar, E. Liebman, and P. Stone. N-agent ad hoc teamwork. Advances in Neural Information Processing Systems, 37:111832–111862, 2024

    2024

  14. [22]

    Samvelyan, T

    M. Samvelyan, T. Rashid, C. S. De Witt, G. Farquhar, N. Nardelli, T. G. Rudner, C.-M. Hung, P. H. Torr, J. Foerster, and S. Whiteson. The starcraft multi-agent challenge. arXiv preprint arXiv:1902.04043, 2019

    1902 arXiv

  15. [23]

    Zhang, M

    Y . Zhang, M. Ding, Y . Yuan, J. Zhang, Q. Yang, G. Shi, F. Jiang, and M. Lu. Multi-uav cooperative pursuit of a fast-moving target uav based on the gm-td3 algorithm. Drones, 8(10): 557, 2024

    2024

  16. [24]

    Y . Li, D. Zhang, J. Chen, Y . Wen, Q. Zhang, S. Mou, and W. Pan. Hola-drone: Hypergraphic open-ended learning for zero-shot multi-drone cooperative pursuit. CoRR, abs/2409.08767,

    arXiv

  17. [25]

    Carroll, R

    M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the utility of learning about humans for human-ai coordination. Advances in neural information processing systems, 32, 2019

    2019

  18. [26]

    A. Lupu, B. Cui, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International Conference on Machine Learning (ICML), pages 7204–7213. PMLR, 2021

    2021

  19. [27]

    R. Zhao, J. Song, Y . Yuan, H. Hu, Y . Gao, Y . Wu, Z. Sun, and W. Yang. Maximum entropy population-based training for zero-shot human-ai coordination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 6145–6153, 2023

    2023

  20. [28]

    Charakorn, P

    R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Generating diverse cooperative agents by learning incompatible policies. In The Eleventh International Conference on Learning Representations, 2023

    2023

  21. [29]

    C. Yu, J. Gao, W. Liu, B. Xu, H. Tang, J. Yang, Y . Wang, and Y . Wu. Learning zero-shot cooperation with humans, assuming humans are biased. arXiv preprint arXiv:2302.01605, 2023

    2023 arXiv

  22. [30]

    Y . Li, S. Zhang, J. Sun, W. Zhang, Y . Du, Y . Wen, X. Wang, and W. Pan. Tackling cooperative incompatibility for zero-shot human-ai coordination. Journal of Artificial Intelligence Research, 80:1139–1185, 2024

    2024

  23. [31]

    Barrett, A

    S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with new teammates. Artificial Intelligence, 242:132–171, 2017

    2017

  24. [32]

    S. Chen, E. Andrejczuk, Z. Cao, and J. Zhang. Aateam: Achieving the ad hoc teamwork by employing the attention mechanism. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7095–7102, 2020

    2020

  25. [33]

    M. A. Rahman, N. Hopner, F. Christianos, and S. V . Albrecht. Towards open ad hoc teamwork using graph-based policy learning. In International conference on machine learning, pages 8776–8786. PMLR, 2021

    2021

  26. [34]

    J. Wang, Y . Li, Y . Zhang, W. Pan, and S. Kaski. Open ad hoc teamwork with cooperative game theory. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  27. [35]

    Fujimoto, H

    S. Fujimoto, H. Hoof, and D. Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pages 1587–1596. PMLR, 2018. 11

    2018

  28. [36]

    Lillicrap

    T. Lillicrap. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015

    2015 arXiv

  29. [37]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms, 2017

    2017

  30. [38]

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu. The surprising effectiveness of ppo in cooperative multi-agent games. Advances in Neural Information Processing Systems, 35:24611–24624, 2022

    2022

  31. [39]

    Janosov, C

    M. Janosov, C. Vir´agh, G. V´as´arhelyi, and T. Vicsek. Group chasing tactics: How to catch a faster prey. New Journal of Physics, 19, 05 2017. doi:10.1088/1367-2630/aa69e7

    2017 doi

  32. [40]

    players

    D. Balduzzi, M. Garnelo, Y . Bachrach, W. Czarnecki, J. Perolat, M. Jaderberg, and T. Graepel. Open-ended learning in symmetric zero-sum games. In International Conference on Machine Learning, pages 434–443. PMLR, 2019. 12 pursuers evaders obstacles Env: 4p2e3o Env: 4p2e5oEnv:...

    2019

  33. [2024]

    URLhttps://doi.org/10.48550/arXiv.2409

    doi:10.48550/ARXIV .2409.08767. URLhttps://doi.org/10.48550/arXiv.2409. 08767

    doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.