pith. machine review for the scientific record. sign in

arxiv: 2604.21355 · v1 · submitted 2026-04-23 · 💻 cs.RO

Recognition: unknown

RPG: Robust Policy Gating for Smooth Multi-Skill Transitions in Humanoid Fighting

Yucheng Xin , Jiacheng Bao , Yubo Dong , Xueqian Wang , Bin Zhao , Xuelong Li , Junbo Tan , Dong Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-09 22:10 UTC · model grok-4.3

classification 💻 cs.RO
keywords humanoid robotmulti-skill policyfighting skillstransition randomizationreinforcement learningrobust controlskill compositionUnitree G1
0
0 comments X

The pith

Randomizing skill transitions and timings during training lets one policy produce stable, smooth multi-skill humanoid fighting.

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

The paper addresses instability when humanoid robots switch between different fighting skills. Separate per-skill policies or direct motion imitation both produce jerky or falling behavior because the ending state of one skill rarely matches the starting state of the next. The authors add deliberate randomization of how one motion connects to another and of when each skill begins, then train a single unified policy on these varied transitions. This produces agile fighting sequences that remain stable and can be interrupted or extended at any moment. The same pipeline also merges basic locomotion with the fighting skills, enabling continuous combat of arbitrary length.

Core claim

By training a hybrid expert policy with motion transition randomization and temporal randomization, a single unified policy generates agile fighting actions that maintain stability and smoothness across skill changes, removing the out-of-domain disturbances caused by state mismatches and supporting a control pipeline that integrates locomotion for humanlike combat of any duration.

What carries the argument

Motion transition randomization and temporal randomization inside a hybrid expert policy framework that trains one unified policy instead of switching among separate skill policies.

If this is right

  • A single policy can chain any number of fighting skills without separate switching logic or recovery steps.
  • Fighting sequences can be started, stopped, or altered at arbitrary times while the robot remains balanced.
  • Locomotion and fighting skills can be combined so the robot can move into position and fight without pausing.
  • The same randomization approach supports long-duration combat that continues until explicitly interrupted.
  • Real-robot experiments on the Unitree G1 confirm that the trained policy transfers without additional fine-tuning for transitions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same randomization of connection points and timing could be applied to other multi-skill robotic tasks such as manipulation or navigation.
  • Training-time randomization may reduce the need for explicit state-alignment modules when composing learned behaviors.
  • If the randomization distribution is widened further, the policy might handle entirely novel skill combinations without retraining.

Load-bearing premise

Randomizing transitions and timings in training will be enough to remove all state-mismatch disturbances that appear when the same skills are executed on the physical robot.

What would settle it

Frequent falls or visible jerks during skill changes when the trained policy is deployed on the Unitree G1 performing a sequence of fighting moves that was never seen exactly in training.

Figures

Figures reproduced from arXiv: 2604.21355 by Bin Zhao, Dong Wang, Jiacheng Bao, Junbo Tan, Xuelong Li, Xueqian Wang, Yubo Dong, Yucheng Xin.

Figure 1
Figure 1. Figure 1: Policy Transition Demonstration. We conducted policy transition tests for robotic combat motions using the proposed RPG. Punching and Sword Swing motions primarily involve the upper body, whereas Jumping and Kicking motions are mainly lower-body actions. We demonstrate 4 distinct policy transition combinations here, highlighting the motion capabilities during transition between upper and lower-body strateg… view at source ↗
Figure 2
Figure 2. Figure 2: Recovery Tests. Due to the short duration of the Jumping motion, it was excluded from consideration. For the other motions, the robot demonstrated the ability to avoid foot sliding and resume a stable, human-like standing posture from any interrupted state during execution. 2. We introduced a novel method incorporating both policy-transition and temporal randomization during expert policies training, and a… view at source ↗
Figure 3
Figure 3. Figure 3: Framework. a) Collection of human demonstration data for combat motions (jumping, punching, sword swing, kicking), with video-based motion conversion and retargeting to adapt to the Unitree G1 humanoid robot; b) Expert networks are trained for each motion type via imitation learning. All experts are loaded concurrently, though only one policy is updated per step. Policy-transition and temporal randomizatio… view at source ↗
Figure 4
Figure 4. Figure 4: Policy Transition Example: Jumping-Punching-Sword Swing [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Humanoid robots have demonstrated impressive motor skills in a wide range of tasks, yet whole-body control for humanlike long-time, dynamic fighting remains particularly challenging due to the stringent requirements on agility and stability. While imitation learning enables robots to execute human-like fighting skills, existing approaches often rely on switching among multiple single-skill policies or employing a general policy to imitate input reference motions. These strategies suffer from instability when transitioning between skills, as the mismatch of initial and terminal states across skills or reference motions introduces out-of-domain disturbances, resulting in unsmooth or unstable behaviors. In this work, we propose RPG, a hybrid expert policy framework, for smooth and stable humanoid multi-skills transition. Our approach incorporates motion transition randomization and temporal randomization to train a unified policy that generates agile fighting actions with stability and smoothness during skill transitions. Furthermore, we design a control pipeline that integrates walking/running locomotion with fighting skills, allowing humanlike long-time combat of arbitrary duration that can be seamlessly interrupted or transit action policies at any time. Extensive experiments in simulation demonstrate the effectiveness of the proposed framework, and real-world deployment on the Unitree G1 humanoid robot further validates its robustness and applicability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes RPG, a hybrid expert policy framework for humanoid robots to achieve smooth and stable transitions between multiple fighting skills. It trains a unified policy via motion transition randomization and temporal randomization to mitigate out-of-domain disturbances from state mismatches across skills. The approach further integrates locomotion (walking/running) with fighting skills to enable long-duration, arbitrarily interruptible combat sequences. Effectiveness is asserted via simulation experiments and real-world deployment on the Unitree G1 humanoid.

Significance. If the empirical results hold with proper quantification, the work would meaningfully advance whole-body control for dynamic, multi-skill humanoid tasks by reducing reliance on explicit policy switching. The real-world hardware validation on the Unitree G1 is a concrete strength that helps bridge sim-to-real gaps, which is valuable in this domain.

major comments (1)
  1. Abstract: The central claim that motion transition randomization plus temporal randomization produces a unified policy with stability and smoothness during skill transitions is asserted without any metrics, baselines, ablation studies, or failure-mode discussion. This is load-bearing for the claim, as it leaves unaddressed whether the randomization distributions are dense enough over posture, velocity, and contact-force mismatches to eliminate out-of-domain disturbances both in simulation and on hardware.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address the single major comment below and will revise the manuscript to strengthen the presentation of our claims.

read point-by-point responses

  1. Referee: Abstract: The central claim that motion transition randomization plus temporal randomization produces a unified policy with stability and smoothness during skill transitions is asserted without any metrics, baselines, ablation studies, or failure-mode discussion. This is load-bearing for the claim, as it leaves unaddressed whether the randomization distributions are dense enough over posture, velocity, and contact-force mismatches to eliminate out-of-domain disturbances both in simulation and on hardware.

    Authors: We agree that the abstract, in its current form, presents the benefits of motion transition randomization and temporal randomization at a high level without embedding quantitative metrics, explicit baseline comparisons, ablation results, or failure-mode analysis. The full manuscript reports extensive simulation experiments (Section 4) that include metrics for transition stability (e.g., center-of-mass deviation and success rates) and smoothness, comparisons against policy-switching and single-policy baselines, ablations isolating each randomization component, and discussion of failure cases under large state mismatches. These results support that the randomization covers relevant posture, velocity, and contact variations sufficiently for robust performance in both simulation and hardware deployment. To make the abstract self-contained and directly responsive to this concern, we will revise it to incorporate concise references to the key quantitative findings, ablations, and failure-mode observations from the experimental sections. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical training procedure is self-contained

full rationale

The manuscript describes a hybrid expert policy trained via motion transition randomization and temporal randomization, with performance evaluated through external simulation rollouts and hardware deployment on the Unitree G1. No equations, fitted parameters, or self-citations are invoked as load-bearing steps that reduce the central claim to its own inputs by construction. The randomization is presented as an input design choice whose effectiveness is measured against independent benchmarks rather than being tautologically redefined as the output. This is the standard non-circular pattern for an empirical robotics policy paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, the paper introduces no explicit free parameters, mathematical axioms, or new physical entities; it relies on standard concepts from imitation learning and reinforcement learning whose details are not supplied.

pith-pipeline@v0.9.0 · 5530 in / 1164 out tokens · 55410 ms · 2026-05-09T22:10:52.668627+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 10 canonical work pages

  1. [1]

    Learning Hu- manoid Locomotion over Challenging Terrain,

    I. Radosavovic, S. Kamat, T. Darrell, and J. Malik, “Learning Hu- manoid Locomotion over Challenging Terrain,” Oct. 2024

    2024

  2. [2]

    Humanoid parkour learning,

    Z. Zhuang, S. Yao, and H. Zhao, “Humanoid parkour learning,”arXiv preprint arXiv:2406.10759, 2024

    work page arXiv 2024

  3. [3]

    Dribble Master: Learning Agile Humanoid Dribbling Through Legged Locomotion,

    Z. Wang, J. Zhou, and Q. Wu, “Dribble Master: Learning Agile Humanoid Dribbling Through Legged Locomotion,” May 2025

    2025

  4. [4]

    A Unified and General Humanoid Whole-Body Controller for Fine-Grained Loco- motion,

    Y . Xue, W. Dong, M. Liu, W. Zhang, and J. Pang, “A Unified and General Humanoid Whole-Body Controller for Fine-Grained Loco- motion,” Feb. 2025

    2025

  5. [5]

    Learning Robust Motion Skills via Critical Adversarial Attacks for Humanoid Robots,

    Y . Zhang, Z. Cao, B. Nie, H. Li, and Y . Gao, “Learning Robust Motion Skills via Critical Adversarial Attacks for Humanoid Robots,” Jul. 2025

    2025

  6. [6]

    Homie: Humanoid loco- manipulation with isomorphic exoskeleton cockpit,

    Q. Ben, F. Jia, J. Zeng, J. Dong, D. Lin, and J. Pang, “Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit,” arXiv preprint arXiv:2502.13013, 2025

    work page arXiv 2025

  7. [7]

    Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning,

    Z. Gu, J. Li, W. Shen, W. Yu, Z. Xie, S. McCrory, X. Cheng, A. Shamsah, R. Griffin, C. K. Liu, A. Kheddar, X. B. Peng, Y . Zhu, G. Shi, Q. Nguyen, G. Cheng, H. Gao, and Y . Zhao, “Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning,” Jan. 2025

    2025

  8. [8]

    Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space,

    Y . Liu, Z. Zhang, H. Wang, and L. Yi, “Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space,” May 2025

    2025

  9. [9]

    Hierarchical Vision-Language Planning for Multi-Step Humanoid Manipulation,

    A. Schakkal, B. Zandonati, Z. Yang, and N. Azizan, “Hierarchical Vision-Language Planning for Multi-Step Humanoid Manipulation,” Jun. 2025

    2025

  10. [10]

    ULC: A Unified and Fine-Grained Controller for Humanoid Loco-Manipulation,

    W. Sun, L. Feng, B. Cao, Y . Liu, Y . Jin, and Z. Xie, “ULC: A Unified and Fine-Grained Controller for Humanoid Loco-Manipulation,” Jul. 2025

    2025

  11. [11]

    MaskedManipulator: Versatile Whole-Body Control for Loco- Manipulation,

    C. Tessler, Y . Jiang, E. Coumans, Z. Luo, G. Chechik, and X. B. Peng, “MaskedManipulator: Versatile Whole-Body Control for Loco- Manipulation,” May 2025

    2025

  12. [12]

    Human-humanoid collaborative carrying,

    D. J. Agravante, A. Cherubini, A. Sherikov, P.-B. Wieber, and A. Kheddar, “Human-humanoid collaborative carrying,”IEEE Trans- actions on Robotics, vol. 35, no. 4, pp. 833–846, 2019

    2019

  13. [13]

    Available: https://arxiv.org/abs/2502.01143

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Panet al., “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,”arXiv preprint arXiv:2502.01143, 2025

    work page arXiv 2025

  14. [14]

    BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,

    T. E. Truong, Q. Liao, X. Huang, G. Tevet, C. K. Liu, and K. Sreenath, “BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,” Aug. 2025

    2025

  15. [15]

    KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills,

    W. Xie, J. Han, J. Zheng, H. Li, X. Liu, J. Shi, W. Zhang, C. Bai, and X. Li, “KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills,” Jun. 2025

    2025

  16. [16]

    Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning,

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi, “Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,”arXiv preprint arXiv:2406.08858, 2024

    work page arXiv 2024

  17. [17]

    HuB: Learning Extreme Humanoid Balance,

    T. Zhang, B. Zheng, R. Nai, Y . Hu, Y .-J. Wang, G. Chen, F. Lin, J. Li, C. Hong, K. Sreenath, and Y . Gao, “HuB: Learning Extreme Humanoid Balance,” May 2025

    2025

  18. [18]

    Learning human-to-humanoid real-time whole-body teleoperation,

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 8944–8951

    2024

  19. [19]

    GMT: General Motion Tracking for Humanoid Whole-Body Control,

    Z. Chen, M. Ji, X. Cheng, X. Peng, X. B. Peng, and X. Wang, “GMT: General Motion Tracking for Humanoid Whole-Body Control,” Jun. 2025

    2025

  20. [20]

    HITTER: A HumanoId Table TEnnis Robot via Hierarchical Planning and Learning,

    Z. Su, B. Zhang, N. Rahmanian, Y . Gao, Q. Liao, C. Regan, K. Sreenath, and S. S. Sastry, “HITTER: A HumanoId Table TEnnis Robot via Hierarchical Planning and Learning,” Aug. 2025

    2025

  21. [21]

    Robust and versatile bipedal jumping control through reinforcement learning,

    Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath, “Robust and versatile bipedal jumping control through reinforcement learning,”arXiv preprint arXiv:2302.09450, 2023

    work page arXiv 2023

  22. [22]

    Expressive whole-body control for humanoid robots,

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Ex- pressive whole-body control for humanoid robots,”arXiv preprint arXiv:2402.16796, 2024

    work page arXiv 2024

  23. [23]

    ExBody2: Advanced expressive humanoid whole-body control,

    M. Ji, X. Peng, F. Liu, J. Li, G. Yang, X. Cheng, and X. Wang, “Exbody2: Advanced expressive humanoid whole-body control,”arXiv preprint arXiv:2412.13196, 2024

    work page arXiv 2024

  24. [24]

    Embrace collisions: Humanoid shad- owing for deployable contact-agnostics motions,

    Z. Zhuang and H. Zhao, “Embrace collisions: Humanoid shad- owing for deployable contact-agnostics motions,”arXiv preprint arXiv:2502.01465, 2025

    work page arXiv 2025

  25. [25]

    Robot parkour learning,

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

    work page arXiv 2023

  26. [26]

    Visual Imitation Enables Contextual Humanoid Control,

    A. Allshire, H. Choi, J. Zhang, D. McAllister, A. Zhang, C. M. Kim, T. Darrell, P. Abbeel, J. Malik, and A. Kanazawa, “Visual Imitation Enables Contextual Humanoid Control,” May 2025

    2025

  27. [27]

    From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots,

    Y . Wang, M. Yang, W. Zeng, Y . Zhang, X. Xu, H. Jiang, Z. Ding, and Z. Lu, “From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots,” Jun. 2025

    2025

  28. [28]

    Okami: Teaching humanoid robots manipulation skills through single video imitation,

    J. Li, Y . Zhu, Y . Xie, Z. Jiang, M. Seo, G. Pavlakos, and Y . Zhu, “Okami: Teaching humanoid robots manipulation skills through single video imitation,” in8th Annual Conference on Robot Learning, 2024

    2024

  29. [29]

    Humanplus: Humanoid shadowing and imitation from humans,

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn, “Humanplus: Humanoid shadowing and imitation from humans,”arXiv preprint arXiv:2406.10454, 2024

    work page arXiv 2024

  30. [30]

    Discovery of skill switching criteria for learning agile quadruped locomotion,

    W. Yu, F. Acero, V . Atanassov, C. Yang, I. Havoutis, D. Kanoulas, and Z. Li, “Discovery of skill switching criteria for learning agile quadruped locomotion,” Feb. 2025

    2025

  31. [31]

    Ase: Large- scale reusable adversarial skill embeddings for physically simulated characters,

    X. B. Peng, Y . Guo, L. Halper, S. Levine, and S. Fidler, “Ase: Large- scale reusable adversarial skill embeddings for physically simulated characters,”ACM Transactions On Graphics (TOG), vol. 41, no. 4, pp. 1–17, 2022

    2022

  32. [32]

    Evaluation-Time Policy Switching for Offline Reinforcement Learning,

    N. S. Neggatu, J. Houssineau, and G. Montana, “Evaluation-Time Policy Switching for Offline Reinforcement Learning,” Mar. 2025

    2025

  33. [33]

    Robust Policy Switching for Antifragile Reinforcement Learning for UA V Deconfliction in Adversarial Envi- ronments,

    D. K. Panda and W. Guo, “Robust Policy Switching for Antifragile Reinforcement Learning for UA V Deconfliction in Adversarial Envi- ronments,” Jun. 2025

    2025

  34. [34]

    Expert Composer Policy: Scalable Skill Repertoire for Quadruped Robots,

    G. Christmann, Y .-S. Luo, and W.-C. Chen, “Expert Composer Policy: Scalable Skill Repertoire for Quadruped Robots,” Mar. 2024

    2024

  35. [35]

    Learning Humanoid Standing-up Control across Diverse Postures,

    T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, and J. Pang, “Learning Humanoid Standing-up Control across Diverse Postures,” Feb. 2025

    2025

  36. [36]

    End-to-End Humanoid Robot Safe and Comfortable Locomotion Policy,

    Z. Wang, X. Yang, J. Zhao, J. Zhou, T. Ma, Z. Gao, A. Ajoudani, and J. Liang, “End-to-End Humanoid Robot Safe and Comfortable Locomotion Policy,” Aug. 2025

    2025

  37. [37]

    Amass: Archive of motion capture as surface shapes,

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black, “Amass: Archive of motion capture as surface shapes,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5442–5451

    2019

  38. [38]

    World-Grounded Human Motion Recovery via Gravity- View Coordinates,

    Z. Shen, H. Pi, Y . Xia, Z. Cen, S. Peng, Z. Hu, H. Bao, R. Hu, and X. Zhou, “World-Grounded Human Motion Recovery via Gravity- View Coordinates,” inSIGGRAPH Asia 2024 Conference Papers, Dec. 2024, pp. 1–11

    2024

  39. [39]

    Perpetual Humanoid Control for Real-time Simulated Avatars,

    Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu, “Perpetual Humanoid Control for Real-time Simulated Avatars,” Sep. 2023

    2023