Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Toward Real-World Cooperative and Competitive Soccer with Quadrupedal Robot Teams

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Quadruped teams play real soccer without a central planner

desk verdict A credible simulation-side systems result with a genuine first demonstration of decentralized multi-robot legged soccer, but the real-world evidence is anecdotal and the zero-shot transfer claim is stronger than the data. read the letter →

arxiv 2505.13834 v2 pith:GUJK7UUE submitted 2025-05-20 cs.RO cs.AI

Zhi Su , Yuman Gao , Emily Lukas , Yunfei Li , Jiaze Cai , Faris Tulbah , Fei Gao , Chao Yu
show 3 more authors
Zhongyu Li Yi Wu Koushil Sreenath
This is my paper · ORCID
classification cs.ROcs.AI
keywords quadrupedrobotsrobotsoccermulti-agentreinforcementlearningfictitiousself-playhierarchicalsim-to-realtransferdecentralizedcontrolLiDARperception
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

This paper tries to establish that fully autonomous, cooperative and competitive soccer is possible for teams of quadruped robots operating in the real world with no external infrastructure. The proposed recipe is a hierarchical multi-agent reinforcement learning framework: separately trained low-level skills for walking, dribbling, and kicking are composed by a high-level strategy policy trained with Multi-Agent Proximal Policy Optimization under Fictitious Self-Play. A sympathetic reader would care because the authors show the resulting policies transfer zero-shot from simulation to physical robots for multi-agent games, producing coordinated passing, interception, counterattacks, and even robot-human play using only onboard LiDAR, onboard computation, and broadcast positions. The paper positions this as evidence that long-horizon team coordination in adversarial, dynamic settings can emerge purely through learning rather than handcrafted rules.

What carries the argument

The load-bearing object is the two-level policy stack. At the bottom, a reusable skill library (Walk, Dribble, Kick, and Stop) maps velocity commands to joint targets at 50 Hz under fixed PD gains, with each skill trained separately by reinforcement learning. At the top, a shared recurrent policy observes only egocentric relative positions (ball, teammates, opponents, goals) and outputs a skill type plus one of eight discretized direction commands at 5 Hz; this policy is trained with multi-agent proximal policy optimization under Fictitious Self-Play, which freezes one side and trains the other against a growing population of past opponent snapshots. Cooperation is not rewarded directly; it emerges from the shared scoring objective and the diversity of opponents in the population. The real-world transfer is carried by onboard perception: a LiDAR detects the high-reflectivity ball, LiDAR-inertial odometry provides localization against a pre-mapped environment, and a wireless broadcast shares poses, so no external infrastructure is needed.

What would settle it

Run a controlled 2v1 match where a second defender stands between the attacker and the ball so that the ball leaves the attacker's LiDAR field of view; if the attacker does not consistently recover or re-acquire the ball, the onboard-perception premise behind the real-world claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a fully decentralized multi-quadruped soccer system can be trained entirely in simulation and then run in the real world on quadruped robots without motion capture, centralized planning, or external tracking. The framework stacks a library of low-level motor skills (walk, dribble, kick) under a recurrent high-level policy that selects a skill and a command direction every 0.2 seconds; the high-level policy is trained with MAPPO under Fictitious Self-Play, so attackers and defenders co-evolve against populations of past opponents. In simulation the full hierarchical method beats a flat end-to-end baseline (98.3% versus 37.5% win rate against a static opponent) and converges faster than variants missing the dribble or kick skills. On physical robots the same policies produce coordinated passing, interception, counterattacks, and one robot-human scoring play, with each robot perceiving via laser-based ball detection and LiDAR-inertial localization and sharing only minimal position messages.

Load-bearing premise

The load-bearing premise is that the simulation-trained policies behave on the physical robots just as they do in the simulator, which in turn requires the onboard laser-based ball detection and self-localization to be accurate enough during real play; the paper reports real demonstrations but no quantitative measurement of this accuracy.

Editorial extensions

If this is right

  • The reusable skill library means a new team configuration, such as 2v2 or 3v3, would require retraining only the high-level policy, not the low-level motor skills.
  • Because execution is decentralized and relies only on onboard sensing, the system can play in outdoor spaces where motion-capture systems and centralized planners are impractical.
  • The same architecture supports robot-human teams: a human teammate is treated as another agent in the broadcast observation, and the learned attacker still coordinates passes and scores.
  • FSP's population-based opponent pool prevents strategy collapse, so the final policy exhibits multi-modal behaviors rather than converging to a single local optimum.
  • The paper's own limitation section indicates that scaling to more robots is gated by LiDAR occlusion and sample complexity, not by the hierarchical framework itself.

Reading between the lines

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

  • One extension the paper points to but does not test is training with simulated ball occlusion, which could address the LiDAR-occlusion bottleneck the authors list as the main obstacle to larger teams.
  • The critic value maps could double as an interpretability tool for debugging physical multi-robot coordination, a use the paper only hints at.
  • If the same hierarchy were applied to other adversarial ball games, only the low-level skills and the high-level observation encoding would need to change; this is a conjecture, not a result reported in the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents a hierarchical MARL approach to quadruped soccer. Low-level skills (Walk, Dribble, Kick) are trained separately in IsaacGym, and a shared high-level GRU policy trained with MAPPO and fictitious self-play selects a skill type and an eight-direction command at 5 Hz using only egocentric relative positions. Experiments cover 1v1, 2v1, and 2v2 simulation matches, ablations against a flat end-to-end policy and against reduced skill sets, out-of-domain initialization tests, and qualitative real-world deployments on Unitree Go1 robots using onboard LiDAR, FAST-LIO localization, and broadcast positions, including robot-robot and robot-human games. The central claim is that this is the first fully decentralized multi-quadruped soccer system in the real world and that the policies transfer zero-shot from simulation.

Significance. If the real-world claims were backed by quantitative trials, this would be a notable advance for sim-to-real multi-agent legged robotics: the hierarchical decomposition is sensible, the FSP curriculum is a principled way to avoid opponent overfitting, the ablations support the contribution of skill diversity, and the out-of-domain initialization numbers in Appendix E (89.2%, 81.9%, 68.9%) are genuine holdout evaluations. The value-map analysis is a useful interpretability device. The main gap is that the paper's headline contribution is the real-world system, and that part of the evidence is currently anecdotal and leaves the perception sim-to-real gap unquantified.

major comments (3)
  1. [Section 4.3 and Appendix G] The real-world evidence consists of selected episodes and value-map stills (Fig. 7) plus representative 1v1 cases in Appendix B, but no aggregate numbers are reported: there is no count of trials, goal totals, possession time, pass-completion rate, or comparison against a scripted or teleoperated baseline. Because the paper's headline contribution is a real-world multi-robot soccer system, this evidence base is anecdotal relative to the claim. Please add quantitative field-trial statistics, or, if such statistics are not available, explicitly label the real-world results as demonstrations and soften the 'first fully decentralized system' claim accordingly.
  2. [Section 3.3, Table 4, Appendix G, Section 6] The zero-shot sim-to-real transfer claim rests on an unmodeled perception channel. The high-level policy is trained on ground-truth relative positions of the ball, teammates, opponents, and goals, and the low-level skills receive exact ball state; deployment obtains these quantities from FAST-LIO localization and thresholded LiDAR reflective-ball detection, with no sensor-noise model, detection-latency model, or occlusion process in training. Section 6 explicitly concedes that LiDAR-based ball detection may suffer from occlusions and proposes simulating occlusion as future work. Please quantify real-world perception quality (detection rate, localization error, latency) and its effect on policy performance, add corresponding robustness training and evaluation, or restrict the zero-shot transfer claim to the conditions under which the perception is accurate.
  3. [Appendix E] The out-of-domain tests vary only initialization positions in simulation with clean observations; they do not evaluate generalization over perception noise, physical parameter variations, or opponent distributions beyond the FSP population. The reported degradation from 89.2% in-domain to 68.9% when attackers start out-of-domain is useful, but it is not evidence about the real-world deployment gap identified above. Please either add domain-randomization tests over sensor and dynamics parameters or state explicitly that the out-of-domain robustness is limited to the simulated observation model.
minor comments (6)
  1. [Table 1] The action-space table is incomplete in the displayed text: the 'Direction Options' column for Walk, Dribble, and Kick is blank, while the Stop row reads '– None'. Please fill in the eight direction options and the per-skill command magnitudes, and repair the table formatting.
  2. [Appendix A] Fig. 8 is captioned as showing four low-level motor skills and includes Receive, but the final skill library in Sec. 3.2 contains only Walk, Dribble, and Kick. Please clarify that Receive was ablated during training and is not part of the deployed policy.
  3. [Section 3.3 and Table 5] The text says 'no explicit reward is given for coordinated behaviors,' but Table 5 includes 'base2ball' (reward for approaching the ball when the agent is the closest teammate) and 'opponent near ball' penalties, which are coordination-relevant shaping terms. Please qualify the statement to distinguish the absence of a direct team-reward term from the presence of teammate-relative shaping rewards.
  4. [Figure 4 caption and Section 4.1] The caption says '20 trails' and should say '20 trials.' In addition, the End2End comparison reports win rates against a static opponent without stating the number of seeds or the variance across seeds; please add this information.
  5. [Section 4.3, Fig. 7(g)] The robot-human collaboration experiment is under-specified: state whether the human teammate is treated as a policy-controlled teammate whose position is broadcast to the robot, how the human's actions are synchronized with the robot's decision loop, and whether the robot policy was ever trained with a human-like teammate. This matters because the claim of robot-human collaboration depends on these protocol details.
  6. [Introduction, Contribution (3)] The phrase 'for the first time' should be scoped precisely. The abstract and introduction should state whether the claimed novelty is the first fully autonomous, learning-based, decentralized multi-quadruped soccer system, and should explicitly compare with prior real-world systems such as Labiosa et al. [9] to make the novelty claim verifiable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the simulation-trained policies are evaluated against held-out opponents and out-of-domain initializations, and the real-world deployment claim rests on empirical transfer assumptions rather than on a reduction to the training objectives.

full rationale

The paper makes no formal derivation that equates any prediction with its inputs. The low-level skills and high-level policy are trained with reward functions and FSP opponent populations, but these are training objectives, not hidden targets. Win rates in Sec. 4.1 (98.3% vs. 37.5%) and Appendix E (in-domain 89.2%, out-of-domain defender 81.9%, out-of-domain attacker 68.9%) are genuine holdout evaluations against fixed opponent policies and initialization distributions, so they are not fitted values renamed as predictions. The FSP population is a curriculum that shapes the policy, but the reported generalization numbers are computed after training against snapshots that do not include the evaluation opponents. The zero-shot sim-to-real claim in Sec. 4.3 is an empirical transfer assertion: the policy is trained on clean observations while deployment uses FAST-LIO localization and reflective-LiDAR ball detection. This is a perceptual gap explicitly acknowledged in Sec. 6 (LiDAR ball detection may suffer from occlusions), but it is not circularity because no equation defines the real system's observations as the training observations, nor is the deployment outcome derived from the training reward. Self-citations appear in the related-work and motivation sections, but the central framework, ablations, out-of-domain tests, and real-world system do not depend on those citations for their validity. The strongest claim is an empirical demonstration; its weakness is the lack of quantitative real-world evaluation, which is a correctness risk rather than a circularity. Overall, the derivation chain from training to evaluation is self-contained and the reported results are not forced by construction.

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

The central capability rests on tuned reward weights, hand-selected FSP thresholds, skill command magnitudes, and the empirical adequacy of IsaacGym and the LiDAR perception stack. No new physical entity or mathematical object is introduced. These are the costs the reader must accept before the claimed real-world coordination follows.

free parameters (4)
  • FSP promotion thresholds (sthres_att, sthres_def) = 0.88 attacker, 0.8 defender (0.8 in 1v1)
    These thresholds decide when a team's snapshot joins the opponent population and when training switches sides; they shape the curriculum, and no sensitivity analysis is reported.
  • High-level reward weights = Table 5: scoring 1000, conceding -1000, out of bounds -500, ball forward 1.0, base2ball 0.3, interference -3, fall…
    Hand-chosen dense and sparse weights guide all emergent behaviors; behavior claims are sensitive to them, and they are not treated as unknowns.
  • Low-level skill command magnitudes and kick threshold = Not reported (vwalk, dribble target speeds, rthres)
    The high-level command magnitude per skill is pre-specified and the kick reward threshold rthres is fixed in Appendix D.1; omitted values are tuned per skill and affect deployment.
  • Direction discretization into eight unit vectors = 8 directions
    Continuous velocity commands are discretized to eight headings in Sec. 3.3, a design choice that constrains the expressiveness of the learned strategy.
assumptions (4)
  • domain assumption IsaacGym simulation with PD gains Kp=35 and Kd=0.5 models Go1 dynamics and ball contact closely enough for zero-shot transfer.
    Sec. 3.2 and Sec. 4.3; no real-world adaptation is performed, so simulation fidelity is load-bearing.
  • domain assumption The pre-mapped environment plus FAST-LIO and reflective-ball LiDAR filtering yields accurate enough positions for all policy observations.
    Appendix G; ball detection, human detection, and mutual localization all rely on this pipeline, and Sec. 6 concedes occlusion as an open failure mode.
  • domain assumption MAPPO with fictitious self-play converges to robust policies under the chosen rewards and population protocol.
    The paper gives training curves and OOD win rates, but convergence is empirical and depends on random seeds, as shown by the failed Walk-only runs in Fig. 4(b).
  • standard math Reward maximization in a Dec-POMG is an appropriate model for cooperative and competitive soccer.
    Appendix D.2; the formal MDP and POMDP machinery is standard and not a source of error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Real-World Cooperative and Competitive Soccer with Quadrupedal Robot Teams." pith.science (2026). https://pith.science/paper/GUJK7UUE

@misc{pith2026250513834,
  author       = {Pith},
  title        = {Pith review of: Toward Real-World Cooperative and Competitive Soccer with Quadrupedal Robot Teams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUJK7UUE}},
  note         = {Machine review of arXiv:2505.13834}
}
read the original abstract

Achieving coordinated teamwork among legged robots requires both fine-grained locomotion control and long-horizon strategic decision-making. Robot soccer offers a compelling testbed for this challenge, combining dynamic, competitive, and multi-agent interactions. In this work, we present a hierarchical multi-agent reinforcement learning (MARL) framework that enables fully autonomous and decentralized quadruped robot soccer. First, a set of highly dynamic low-level skills is trained for legged locomotion and ball manipulation, such as walking, dribbling, and kicking. On top of these, a high-level strategic planning policy is trained with Multi-Agent Proximal Policy Optimization (MAPPO) via Fictitious Self-Play (FSP). This learning framework allows agents to adapt to diverse opponent strategies and gives rise to sophisticated team behaviors, including coordinated passing, interception, and dynamic role allocation. With an extensive ablation study, the proposed learning method shows significant advantages in the cooperative and competitive multi-agent soccer game. We deploy the learned policies to real quadruped robots relying solely on onboard proprioception and decentralized localization, with the resulting system supporting autonomous robot-robot and robot-human soccer matches on indoor and outdoor soccer courts.

Figures

Figures reproduced from arXiv: 2505.13834 by the authors.

Figure 1
Figure 1. The proposed hierarchical framework consists of a high-level strategy policy ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Hierarchical Architecture. (a) Low-level skills architecture. The low-level observa￾tion is formed by concatenating the low-level robot observations o l r,t, ball position pball,t, velocity command ct, and previous joint action at−1. An auxiliary estimator network predicts privileged information unavailable to the policy, including the robot state sˆr,t (e.g., motor stiffness), ball state sˆball,t (e.g., ball veloci… view at source ↗
Figure 3
Figure 3. The FSP training procedure, where each side is [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Ablation Study. (a) Ball trajectories under [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Policy evolution with FSP in 1v1 setting. (a) Co-evolutionary dynamics between attacker [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Rollouts of policy trained with FSP in 2v1 setting. (b) For attackers, training against all [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Real-world behavior analysis. (a)(d) The top-down view of two attackers coordinating to [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Real-world demonstration of low-level skills: (a) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Demonstration of 1v1 real world deployment. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Demonstration of more emergent high-level strategies in 2v1 game. (a)-(b): The defender [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Visualization of the out-of-domain test setup. The highlighted rectangular regions illus [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Demonstration of emergent strategies in 2v2 game. (a): The attackers executed coordi [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Real world deployment overview: (a) Quadrupedal Soccer Hardware: Each robot is equipped with a LiDAR and onboard computer for perception and control. (b) Software Pipeline: The system performs ball detection, human detection, and self-localization from LiDAR and IMU d…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CRAFT: Coaching Reinforcement Learning Autonomously using Foundation Models for Multi-Robot Coordination Tasks

    cs.RO 2025-09 conditional novelty 6.0 of 10

    An LLM/VLM coaching loop that generates curricula and reward functions enabled MARL agents to learn coordinated gate passing, seesaw balancing, and bimanual pot lifting, with one policy transferred to real quadrupeds.

Reference graph

Works this paper leans on

42 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    T. He, C. Zhang, W. Xiao, G. He, C. Liu, and G. Shi. Agile but safe: Learning collision-free high-speed legged locomotion. arXiv preprint arXiv:2401.17583, 2024

    arXiv 2024

  2. [2]

    S. Zhu, R. Huang, L. Mou, and H. Zhao. Robust robot walker: Learning agile locomotion over tiny traps. arXiv preprint arXiv:2409.07409, 2024

    arXiv 2024

  3. [3]

    Zhuang, Z

    Z. Zhuang, Z. Fu, J. Wang, C. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao. Robot parkour learning. arXiv preprint arXiv:2309.05665, 2023

    arXiv 2023

  4. [4]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, and D. Pathak. Extreme parkour with legged robots. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 11443–11450. IEEE, 2024

    2024

  5. [5]

    Z. He, K. Lei, Y . Ze, K. Sreenath, Z. Li, and H. Xu. Learning visual quadrupedal loco- manipulation from demonstrations. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9102–9109. IEEE, 2024

    work page 2024

  6. [6]

    Cheng, D

    J. Cheng, D. Kang, G. Fadini, G. Shi, and S. Coros. Rambo: Rl-augmented model-based optimal control for whole-body loco-manipulation. arXiv preprint arXiv:2504.06662, 2025

    arXiv 2025

  7. [7]

    J. Lee, M. Bjelonic, A. Reske, L. Wellhausen, T. Miki, and M. Hutter. Learning robust au- tonomous navigation and locomotion for wheeled-legged robots. Science Robotics , 9(89): eadi9641, 2024

    2024

  8. [8]

    X. Chen, A. Ghadirzadeh, J. Folkesson, M. Bj ¨orkman, and P. Jensfelt. Deep reinforcement learning to acquire navigation skills for wheel-legged robots in complex environments. In2018 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages 3110–

    work page

Show all 42 references
  1. [9]

    Labiosa, Z

    A. Labiosa, Z. Wang, S. Agarwal, W. Cong, G. Hemkumar, A. N. Harish, B. Hong, J. Kelle, C. Li, Y . Li, et al. Reinforcement learning within the classical robotics stack: A case study in robot soccer. arXiv preprint arXiv:2412.09417, 2024

    2024 arXiv

  2. [10]

    Haarnoja, B

    T. Haarnoja, B. Moran, G. Lever, S. H. Huang, D. Tirumala, J. Humplik, M. Wulfmeier, S. Tun- yasuvunakool, N. Y . Siegel, R. Hafner, et al. Learning agile soccer skills for a bipedal robot with deep reinforcement learning. Science Robotics, 9(89):eadi8022, 2024

    2024

  3. [11]

    Tirumala, M

    D. Tirumala, M. Wulfmeier, B. Moran, S. Huang, J. Humplik, G. Lever, T. Haarnoja, L. Hasen- clever, A. Byravan, N. Batchelor, et al. Learning robot soccer from egocentric vision with deep reinforcement learning. arXiv preprint arXiv:2405.02425, 2024

    2024 arXiv

  4. [12]

    S. Liu, G. Lever, Z. Wang, J. Merel, S. A. Eslami, D. Hennes, W. M. Czarnecki, Y . Tassa, S. Omidshafiei, A. Abdolmaleki, et al. From motor control to team play in simulated humanoid football. Science Robotics, 7(69):eabo0235, 2022

    2022

  5. [13]

    Z. Li, F. Bjelonic, V . Klemm, and M. Hutter. Marladona-towards cooperative team play using multi-agent reinforcement learning. arXiv preprint arXiv:2409.20326, 2024. 10

    2024 arXiv

  6. [14]

    Heinrich, M

    J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In Inter- national conference on machine learning , pages 805–813. PMLR, 2015

    2015

  7. [15]

    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. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

    2022

  8. [16]

    H. Kitano. RoboCup-97: robot soccer world cup I, volume 1395. Springer Science & Business Media, 1998

    1998

  9. [17]

    Browning, J

    B. Browning, J. Bruce, M. Bowling, and M. Veloso. Stp: Skills, tactics, and plays for multi- robot control in adversarial environments. Proceedings of the Institution of Mechanical Engi- neers, Part I: Journal of Systems and Control Engineering , 219(1):33–52, 2005

    2005

  10. [18]

    Behnke, J

    S. Behnke, J. Stuckler, M. Schreiber, H. Schulz, M. Bohnert, and K. Meier. Hierarchical reactive control for a team of humanoid soccer robots. In 2007 7th IEEE-RAS International Conference on Humanoid Robots, pages 622–629. IEEE, 2007

    2007

  11. [19]

    S.-J. Yi, S. McGill, D. Hong, and D. Lee. Hierarchical motion control for a team of humanoid soccer robots. International Journal of Advanced Robotic Systems , 13(1):32, 2016

    2016

  12. [20]

    Kohl and P

    N. Kohl and P. Stone. Policy gradient reinforcement learning for fast quadrupedal locomotion. In IEEE International Conference on Robotics and Automation, 2004. Proceedings. ICRA ’04. 2004, volume 3, pages 2619–2624 V ol.3, 2004

    2004

  13. [21]

    Schwab, Y

    D. Schwab, Y . Zhu, and M. Veloso. Learning skills for small size league robocup. InRoboCup 2018: Robot World Cup XXII , page 83–95, Berlin, Heidelberg, 2018. Springer-Verlag. ISBN 978-3-030-27543-3. doi:10.1007/978-3-030-27544-0 7. URL https://doi.org/10.1007/ 978-3-030-27544-0_7

    2018 doi

  14. [22]

    Y . Ji, G. B. Margolis, and P. Agrawal. Dribblebot: Dynamic legged manipulation in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 5155–5162. IEEE, 2023

    2023

  15. [23]

    Y . Hu, K. Wen, and F. Yu. Dexdribbler: Learning dexterous soccer manipulation via dynamic supervision. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024

    2024

  16. [24]

    D. Zhu, Z. Yang, T. Wu, L. Ge, X. Li, Q. Liu, and X. Li. Dynamic legged ball manipulation on rugged terrains with hierarchical reinforcement learning. arXiv preprint arXiv:2504.14989v1 [cs.RO], Apr. 2025

    2025 arXiv

  17. [25]

    Y . Ji, Z. Li, Y . Sun, X. B. Peng, S. Levine, G. Berseth, and K. Sreenath. Hierarchical reinforce- ment learning for precise soccer shooting skills using a quadrupedal robot. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 1479–1486....

    2022

  18. [26]

    Huang, Z

    X. Huang, Z. Li, Y . Xiang, Y . Ni, Y . Chi, Y . Li, L. Yang, X. B. Peng, and K. Sreenath. Creating a dynamic quadrupedal robotic goalkeeper with reinforcement learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 2715–2722. IEEE, 2023

    2023

  19. [27]

    T. Kim, L. F. Vecchietti, K. Choi, S. Sariel, and D. Har. Two-stage training algorithm for ai robot soccer. PeerJ Computer Science , 7:e718, Sep 2021. doi:10.7717/peerj-cs.718. URL https://doi.org/10.7717/peerj-cs.718

    2021 doi

  20. [28]

    Abreu, L

    M. Abreu, L. P. Reis, and N. Lau. Designing a skilled soccer team for RoboCup: ex- ploring skill-set-primitives through reinforcement learning. Neural Computing and Ap- plications, 2025. doi:10.1007/s00521-025-11151-3. URL https://doi.org/10.1007/ s00521-025-11151-3 . 11

    2025 doi

  21. [29]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021

    2021 arXiv

  22. [30]

    Unitree Robotics, Go1, 2025, https://www.unitree.com/go1, [Online; accessed Apr. 2025]

    2025

  23. [31]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  24. [32]

    G. B. Margolis and P. Agrawal. Walk these ways: Tuning robot control for generalization with multiplicity of behavior. Conference on Robot Learning, 2022

    2022

  25. [33]

    K. Cho, B. van Merrienboer, C. Gulcehre, F. Bougares, H. Schwenk, and Y . Bengio. Learn- ing phrase representations using rnn encoder-decoder for statistical machine translation. In Conference on Empirical Methods in Natural Language Processing (EMNLP 2014) , 2014

    2014

  26. [34]

    Livox, MID-360, 2025, https://www.livoxtech.com/mid-360, [Online; accessed Apr. 2025]

    2025

  27. [35]

    Xu and F

    W. Xu and F. Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters , 6(2):3317–3324, 2021. 12 A Low-level Skills Demonstration Figure 8: Real-world demonstration of low-level skills: (a) ...

    2021

  28. [36]

    Dribbling is only activated when the robot is sufficiently close to the ball; otherwise, it is mapped to a walking command directed toward the ball

  29. [37]

    Kicking is only activated when the robot is sufficiently close to the ball; otherwise, it is mapped to a stationary stepping

  30. [38]

    Once kicking is initiated at a high-level decision step, the robot continues executing the kick until the ball is determined to have moved far away, indicating a successful kick. 17

  31. [39]

    These transition rules are consistently applied during both training and real-world deployment

    Walking toward other robots is prevented by predicting future positions using the current location and velocity command, and checking for potential collisions. These transition rules are consistently applied during both training and real-world deployment. E Out of Domain Test ...

  32. [40]

    In-domain: both the defender and attackers are initialized within the same small region used during training

  33. [41]

    Out-of-domain (defender position): the defender is initialized within a larger region on its side, while the attackers’ initialization remains unchanged

  34. [42]

    Out-of-domain (attacker position): the attackers are initialized within a larger region on their side, while the defender’s initialization remains unchanged (Fig. 11). The win rate in the in-domain setting is 89.2%. In the out-of-domain setting with a randomized defender posit...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.