Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

The paper shows that residual-action reinforcement learning—correcting a reference motion rather than re-synthesizing it—enables robust long-horizon, high-dynamic humanoid motion tracking that transfers zero-shot to real robots.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Residual-action reinforcement learning, with selective corrections on hip and knee pitch joints, enables zero-shot long-horizon dance tracking on real humanoid robots.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Plausible recipe and honest hardware demos, but the residual-action mechanism is confounded with a reduced DOF set and thin statistics; needs a sharper ablation before the headline holds. the 4 major comments →

arxiv 2509.20717 v2 pith:NPKCWRO7 submitted 2025-09-25 cs.RO cs.AI

RobotDancing: Residual-Action Reinforcement Learning Enables Robust Long-Horizon Humanoid Motion Tracking

Zhenguo Sun , Yibo Peng , Yuan Meng , Xukun Li , Bo-Sheng Huang , Zhenshan Bing , Xinlong Wang , Alois Knoll This is my paper
classification cs.RO cs.AI
keywords residual reinforcement learninghumanoid motion trackinglong-horizon controlsim-to-real transferdance motionsmotion resamplingaction parameterizationwhole-body control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

RobotDancing claims that the brittleness of long-horizon, high-dynamic humanoid motion tracking comes from a design choice: policies that command absolute joint positions try to re-synthesize the motion and let small errors accumulate, while policies that emit residual corrections on top of the reference motion learn to compensate the physics mismatch (actuation limits, latency, friction, inertia) instead. Training one policy per reference sequence with a single observation, reward, and hyperparameter recipe, the paper shows on three humanoid platforms that residual actions reduce error accumulation, improve sample efficiency, and transfer zero-shot to real hardware for multi-minute dance sequences with jumps, spins, and cartwheels. The method also couples this with a two-stage sampling scheme—offline pose-distribution balancing plus online failure-prioritized resampling—that improves coverage of rare poses and speeds learning. A sympathetic reader would take this as evidence that residualization is a practical, scalable alternative to multi-stage distillation or heavy system identification for reference tracking.

Core claim

The paper's central claim is that parameterizing the RL action as a residual joint target—the commanded position equals the reference position plus a learned correction, executed by a low-level PD controller—lets the policy learn the dynamics discrepancy between reference and robot rather than regenerate the whole motion. Because the reference already carries kinematic plausibility, policy capacity is spent on corrections that absorb model–plant mismatch. To implement this, the paper selectively residualizes only high-variance, error-sensitive joints (hip and knee pitch), drives stable joints directly from the reference, and bounds residual magnitudes during training. It also combines offlin

What carries the argument

The load-bearing object is the residual-action parametrization: the commanded joint target is the reference joint position plus a learned residual, which the PD controller tracks. The argument hinges on dividing labor—the reference supplies 'what to do,' the residual supplies 'how to compensate physics'—supported by a selective residualization mask that restricts residuals to hip/knee pitch joints, and by a two-stage resampler (offline distribution-aware balancing + online failure-aware priority) that keeps training concentrated on rare and difficult poses. These pieces together reduce action variance and error accumulation, and are claimed to enable zero-shot sim-to-real and cross-platform

Load-bearing premise

The method's advantage depends on the hand-picked decision to apply residual corrections only to the hip and knee pitch joints; if that choice is wrong for another robot body or motion family, the gains over residualizing all joints or none could shrink or even reverse.

What would settle it

Train the same pipeline on a motion library whose high-variance joints are not hip/knee pitch—for instance an upper-body gesture-heavy sequence—keeping the same selective mask; if the selective variant no longer beats both the no-residual and all-joint baselines on tracking error, then the hand-crafted joint selection, not residual-action learning itself, is carrying the result.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Residual-action policies can track multi-minute, high-energy dance sequences on real humanoid hardware with no test-time modifications, including jumps, spins, and cartwheels.
  • Selective residualization on hip/knee pitch joints reduces global mean per-body position error by roughly 15–20% and mean per-joint position error by about 20% relative to an absolute-command baseline, and outperforms residualizing all joints.
  • The same training recipe (observations, rewards, hyperparameters) transfers across three humanoid platforms and across motions, including zero-shot transfer to simpler, unseen references such as walking.
  • Combining offline distribution-aware balancing with online failure-aware priority adaptation improves sample efficiency and asymptotic reward compared to failure-aware sampling alone.
  • All policies are trained in simulation and transfer to reality without per-sequence tuning, suggesting the dynamics compensation learned is general rather than motion-memorized.

Where Pith is reading between the lines

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

  • Beyond the paper: the residual-action principle should transfer to other reference-tracking domains—manipulation, teleoperation, exoskeletons—where a nominal plan exists but dynamics mismatch is the bottleneck; the same 'correct, don't regenerate' logic should reduce drift.
  • Beyond the paper: the hand-crafted selective mask suggests a learned, morphology-aware gating mechanism (the authors' stated future work) could automatically discover which joints need residuals; if so, the true generalization bottleneck across robot bodies may be the mask, not the residual formulation.
  • Beyond the paper: the proprioception-only controller's success inside a known motion library leaves open whether residual corrections retain their advantage when observations include noisy, delayed perception; combining residual actions with learned disturbance estimation is a plausible extension.
  • Beyond the paper: the observed zero-shot transfer from dance to walking/standing implies a single residual policy might serve as a reusable movement primitive for higher-level planners, potentially reducing per-skill training across a behavior library.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes RobotDancing, a single-stage residual-action RL framework for long-horizon, high-dynamic humanoid motion tracking. Instead of predicting absolute joint commands, the policy outputs residual corrections on top of retargeted reference joint targets, with selective residualization applied only to hip/knee pitch DOFs. Training combines an asymmetric actor-critic PPO setup, vectorized rewards, a two-stage motion sampling strategy (offline distribution-aware balancing plus online failure-aware priority adaptation), curriculum learning, and domain randomization. The authors evaluate on eight LAFAN1 dance sequences in simulation and on a Unitree G1 humanoid, with qualitative cross-platform validation on H1 and H1-2. The central claims are that residual actions reduce error accumulation and improve sample efficiency relative to absolute-action baselines, that selective residualization outperforms both no residualization and full residualization, and that the resulting policies transfer zero-shot to real hardware without test-time modification.

Significance. If substantiated, the framework would be a practical and welcome contribution to humanoid motion tracking: it is simple, uses a unified observation/reward/hyperparameter recipe across motions and platforms, and demonstrates real-robot deployment of multi-minute, high-energy dance behaviors. The strengths of the paper include the explicit residual-action formulation, the modular reward design, the combination of offline and online sampling, and the reported zero-shot sim-to-real transfer on G1 plus qualitative transfer to H1/H1-2. The main reservations concern the strength of the quantitative evidence for the central residualization claim: the current statistics and ablations are underpowered and confounded, and the hardware evaluation is predominantly binary. The core idea is plausible and the real-robot executions are valuable, but the load-bearing comparisons need additional experimental support before the headline claims can be accepted.

major comments (4)
  1. [Section IV-B, Table I] The reported mean±std does not state what the spread is over (seeds, episodes, or timesteps), and no seed counts are given. For the headline comparison, SELECTIVE vs ALL on E_g-mpbpe is 484.72±147.37 vs 548.76±132.64, less than one pooled standard deviation; the gap over NONE is also within overlapping error bars. Without per-seed results, confidence intervals, or paired significance tests, the claimed improvements from residualization, and from selective residualization in particular, are not statistically established. Please report the number of seeds, per-seed means, and appropriate error bars or tests for all entries in Table I.
  2. [Section III-D and Section IV-B] The SELECTIVE condition is confounded: compared to NONE it changes both the action parameterization (residual vs absolute) and the active DOF set (only hip/knee pitch receive actions), while compared to ALL it changes only the DOF mask but with a different action space dimension. The paper never compares residual vs absolute actions under the same action mask. Thus the observed benefit could come from the hand-selected low-dimensional action space, from reduced over-correction on stable joints, or from the residual parameterization itself, rather than from residualization per se. Since residualization is the central mechanism, this ambiguity is load-bearing. Please add ablations such as absolute commands on the same hip/knee-pitch-only DOFs, or residual actions on the same selected DOFs, to isolate the effect. The authors' own statement in Section V that the selective residualization is
  3. [Section IV-C, Table II] The real-robot evidence is reported as binary success/untested flags, with no tracking error, time-to-failure, or number of trials; three of the eight G1 sequences are untested. The claim of "robust long-horizon motion tracking" on hardware therefore rests on qualitative stills and videos rather than quantitative measurements. Please report at least one hardware tracking metric (e.g., E_g-mpbpe, E_mpjpe, or episode duration before failure) for the successfully deployed sequences, along with the number of hardware runs. Without such data, the sim-to-real robustness claim is not quantitatively supported.
  4. [Figures 5-6 and Section IV-B] The learning curves appear to be single runs with no confidence intervals or seed counts. Claims that SELECTIVE shows faster improvement and a higher asymptote than ALL/NONE, and that Failure+Balanced converges faster and to higher reward than Failure alone, are therefore not supported by the presented evidence. Please provide multi-seed learning curves with shaded error bars, and preferably a quantitative summary table for the sampling ablation, mirroring the format of Table I.
minor comments (6)
  1. [Section IV-B] Typo: "I-CTRL fasion" should be "I-CTRL fashion".
  2. [Section III-A] The observation definition in Eq. (1) uses O_t = [P_t; G_{t+1}], but the policy is described as using stacked observations O_{t-k:t} and goals G_{t-k+1:t+1}. Please clarify the notation so single-step and stacked forms are not conflated.
  3. [Table II] The footnote for "walk1 13 [0,50]s" is unclear. State explicitly what cross-motion means here: is the policy trained on dance1 1 evaluated only on the first 50 seconds of walk1 13, and what metric is used? The current formatting obscures an interesting result.
  4. [Section III-C and Section III-E] The symbol s is used both for the segment index in Eq. (9) and for the curriculum penalty scale s_pen(t) in Eq. (5). This is a minor but avoidable collision of notation.
  5. [Section IV-A] The action dimensions are given as d_act in {23,21,19} for G1, H1-2, and H1, respectively. Please clarify, for each platform, which joints are residualized under the selective residualization scheme; the paper only states hip/knee pitch for G1.
  6. [Section IV-C] The sentence "15K iterations training for every sequence is deployable" is vague. Specify the deployability criterion (e.g., success in MuJoCo sim-to-sim validation) and how it was determined.

Circularity Check

0 steps flagged

No circularity: residual-action gains are empirical ablations, not definitional consequences.

full rationale

The paper's central claim—that residual actions improve long-horizon tracking—is supported by direct RL comparisons (NONE vs ALL vs SELECTIVE, Table I; learning curves, Figs. 5–6) under a fixed observation, reward, and hyperparameter recipe. The residual parameterization q_t^tar = q_{t+1}^ref + a_t (Eq. 3) does not by construction force lower tracking error; a_t is learned by PPO and could be zero, harmful, or beneficial. The sampling weights in Eqs. (8)–(9) are computed from motion statistics and online failure signals, not from the evaluation metrics used to support the headline. The hand-crafted choice of hip/knee-pitch residualization is an admitted limitation (Sec. V: 'Our current selective residualization is hand-crafted. While effective, this heuristic may be suboptimal for other morphologies or motion families.') and is a legitimate generalization/confound concern—the SELECTIVE ablation changes both action parameterization and active-DOF set—but it is not circular: the SELECTIVE advantage is empirical and could in principle have gone the other way. Self-citations to co-authored prior work (e.g., [6], [9]) appear only as related work or as a stated source for one sampling heuristic, not as load-bearing proof of the paper's core claims. No equation reduces to itself, no fitted constant is renamed as a prediction, and no uniqueness or ansatz is imported via self-citation. The derivation chain is therefore self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The central claim rests on hand-tuned hyperparameters whose values are not disclosed, and on untested assumptions about the sufficiency of the reference and the simulation gap.

free parameters (5)
  • reward weight vector {lambda_tau, lambda_delta_a, lambda_lim, lambda_cnt, lambda_term} = not reported
    Weights in Eq. (7) are 'tuned once' (Sec. III-B) and reused across tasks; they influence the tradeoff between tracking and regularization and are not listed.
  • Gaussian tolerance sigma_chi = not reported
    Tolerances in Eq. (6) are fixed or slowly adapted (Sec. III-B) and affect the reward shaping.
  • failure-prior temperature beta and EMA factor alpha = not reported
    Parameters in Eq. (9) control how aggressively the sampler targets difficult segments.
  • domain randomization ranges = not reported
    Ranges over friction, masses/inertias, PD gains, delays are described qualitatively in Sec. III-F but not quantified.
  • curriculum schedule and penalty scale s_pen(t) = not reported
    Progressive curriculum described in Sec. III-E; exact schedule omitted.
axioms (3)
  • domain assumption Reference motions from retargeted human data are kinematically plausible but dynamically inconsistent with the robot, so residual correction is well-posed.
    Assumed throughout Sec. I and III-B to justify learning residuals on top of the reference.
  • domain assumption Simulation with domain randomization is faithful enough for zero-shot sim-to-real.
    Sec. III-F and Sec. IV-C rely on sim-to-sim validation in MuJoCo to predict real behavior, without quantifying the sim-to-real gap.
  • domain assumption PD control at 50 Hz with the same gains across platforms is sufficient.
    Sec. IV-A states a low-level joint-space PD controller running at 50 Hz is used; the policy is expected to compensate dynamics mismatch via residuals.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RobotDancing: Residual-Action Reinforcement Learning Enables Robust Long-Horizon Humanoid Motion Tracking." pith.science (2026). https://pith.science/paper/NPKCWRO7

@misc{pith2026250920717,
  author       = {Pith},
  title        = {Pith review of: RobotDancing: Residual-Action Reinforcement Learning Enables Robust Long-Horizon Humanoid Motion Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPKCWRO7}},
  note         = {Machine review of arXiv:2509.20717}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Long-horizon, high-dynamic motion tracking on humanoids remains brittle: retargeted reference motions are typically kinematically plausible but dynamically inconsistent with the robot, so small tracking errors accumulate and eventually destabilize control. We present RobotDancing, a practical single-stage reinforcement learning recipe that tracks retargeted motions by predicting reference-conditioned residual joint targets. By parameterizing actions as residual corrections on top of the reference, the policy can focus on compensating reference--robot dynamics mismatch (e.g., actuation limits, latency, friction, inertia) rather than re-synthesizing the motion. We train one policy per reference sequence while reusing a common training and deployment recipe across motions and platforms. We evaluate RobotDancing on eight LAFAN1 dance motions on Unitree G1 and conduct cross-platform experiments on H1 and H1-2. The resulting G1 policies execute long-horizon, high-energy behaviors on hardware without test-time modifications, and qualitative H1/H1-2 clips further illustrate cross-platform feasibility.

Figures

Figures reproduced from arXiv: 2509.20717 by Alois Knoll, Bo-Sheng Huang, Xinlong Wang, Xukun Li, Yibo Peng, Yuan Meng, Zhenguo Sun, Zhenshan Bing.

Figure 1
Figure 1. Figure 1: Framework of RobotDancing. (a) The pipeline from motion data processing to RL policy learning. (b) The RL training framework for long￾horizon high-dynamic motion tracking tasks through residual action learning. (c) The deployment framework enables Sim2Sim and Sim2Real transfers. reference-conditioned policies can reproduce diverse skills under physics constraints. Subsequent systems extended this paradigm … view at source ↗
Figure 2
Figure 2. Figure 2: Real-World High-Dynamic Motions across Platforms. Multiple platforms perform long-horizon, high-dynamic dance sequences. Motion trails visualize stable tracking over time; the same residual-action framework applied on all platforms without task-specific retuning. where spen(t) ∈ [smin, smax] is a curriculum-controlled penalty scale, as described in [4], [5]. Task (Motion Tracking). We follow a DeepMimic-st… view at source ↗
Figure 3
Figure 3. Figure 3: Mean and Standard Deviation of Lower-Body Joint Angles Across Dance Motions. Long-horizon, high-dynamic dance motions exhibit pronounced inter-motion-segment differences and large variance in key lower-body joint positions. therefore over-revisits easy segments and starves rare, error￾prone ones, which inflates gradient noise and slows progress on the true bottlenecks. We treat difficulty as a streaming si… view at source ↗
Figure 5
Figure 5. Figure 5: Learning Curves under Residualization Strategies. Reward versus training steps for NONE, ALL, and SELECTIVE [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learning Curves for Sampling Ablations. Failure+Balanced converges faster and to a higher reward than Failure. C. Performance We evaluated all dance sequences in total eight from [32]. The evaluation results are listed in Table II. These sequences are long-horizon (≈3 min each) and highly dynamic, con￾taining jumps, spins (up to 360◦ ), sprints, abrupt stops, and rapid transitions—challenging even for huma… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Stubborn: A Streamlined and Unified Reinforcement Learning Framework for Robust Motion Tracking and Fall Recovery for Humanoids

    cs.RO 2026-06 unverdicted novelty 6.0

    Stubborn introduces a unified RL framework with yaw-aligned representation, Bernoulli probabilistic termination, and adaptive sampling for robust humanoid motion tracking and fall recovery.

Reference graph

Works this paper leans on

35 extracted references · 19 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,”ACM Transactions On Graphics (TOG), vol. 37, no. 4, pp. 1–14, 2018

    2018

  2. [2]

    Universal humanoid motion representations for physics-based control,

    Z. Luo, J. Cao, J. Merel, A. Winkler, J. Huang, K. Kitani, and W. Xu, “Universal humanoid motion representations for physics-based control,”arXiv preprint arXiv:2310.04582, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Masked- mimic: Unified physics-based character control through masked mo- tion inpainting,

    C. Tessler, Y . Guo, O. Nabati, G. Chechik, and X. B. Peng, “Masked- mimic: Unified physics-based character control through masked mo- tion inpainting,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–21, 2024

    2024

  4. [4]

    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,”arXiv preprint arXiv:2506.12851, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    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

    Pith/arXiv arXiv 2025

  6. [6]

    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

    Pith/arXiv arXiv 2024

  7. [7]

    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

    Pith/arXiv arXiv 2024

  8. [8]

    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

    Pith/arXiv arXiv 2024

  9. [9]

    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,”arXiv preprint arXiv:2506.14770, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Unitracker: Learning universal whole-body motion tracker for humanoid robots,

    K. Yin, W. Zeng, K. Fan, Z. Wang, Q. Zhang, Z. Tian, J. Wang, J. Pang, and W. Zhang, “Unitracker: Learning universal whole-body motion tracker for humanoid robots,”arXiv preprint arXiv:2507.07356, 2025

    arXiv 2025

  11. [11]

    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,”arXiv preprint arXiv:2506.12779, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    I-ctrl: Imitation to control humanoid robots through constrained reinforcement learning,

    Y . Yan, E. V . Mascaro, T. Egle, and D. Lee, “I-ctrl: Imitation to control humanoid robots through constrained reinforcement learning,”arXiv preprint arXiv:2405.08726, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Residual policy learning for perceptive quadruped control using differentiable simulation,

    J. Y . Luo, Y . Song, V . Klemm, F. Shi, D. Scaramuzza, and M. Hutter, “Residual policy learning for perceptive quadruped control using differentiable simulation,”arXiv preprint arXiv:2410.03076, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    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,” inProceedings of the fourteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2011, pp. 627–635

    2011

  15. [15]

    Attention-based map encoding for learning generalized legged loco- motion,

    J. He, C. Zhang, F. Jenelten, R. Grandia, M. B ¨acher, and M. Hutter, “Attention-based map encoding for learning generalized legged loco- motion,”Science Robotics, vol. 10, no. 105, p. eadv3604, 2025

    2025

  16. [16]

    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. Sreenathet al., “Hub: Learning extreme humanoid balance,”arXiv preprint arXiv:2505.07294, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    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,”arXiv preprint arXiv:2502.08378, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Real-world humanoid locomotion with reinforcement learning,

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath, “Real-world humanoid locomotion with reinforcement learning,”Science Robotics, vol. 9, no. 89, p. eadi9579, 2024

    2024

  19. [19]

    Perpetual humanoid control for real-time simulated avatars,

    Z. Luo, J. Cao, K. Kitani, W. Xuet al., “Perpetual humanoid control for real-time simulated avatars,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 895– 10 904

    2023

  20. [20]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Ara ˜Aˇsjo, Z.-a. Cao, X. B. Peng, J. Wu, and C. K. Liu, “Twist: Teleoperated whole-body imitation system,”arXiv preprint arXiv:2505.02833, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Clone: Closed-loop whole-body humanoid teleoperation for long- horizon tasks,

    Y . Li, Y . Lin, J. Cui, T. Liu, W. Liang, Y . Zhu, and S. Huang, “Clone: Closed-loop whole-body humanoid teleoperation for long- horizon tasks,”arXiv preprint arXiv:2506.08931, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    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

    Pith/arXiv arXiv 2025

  23. [23]

    Amo: Adaptive motion optimization for hyper-dexterous humanoid whole- body control,

    J. Li, X. Cheng, T. Huang, S. Yang, R.-Z. Qiu, and X. Wang, “Amo: Adaptive motion optimization for hyper-dexterous humanoid whole- body control,”arXiv preprint arXiv:2505.03738, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    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,”arXiv preprint arXiv:2508.08241, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Residual reinforcement learning for robot control,

    T. Johannink, S. Bahl, A. Nair, J. Luo, A. Kumar, M. Loskyll, J. A. Ojea, E. Solowjow, and S. Levine, “Residual reinforcement learning for robot control,” in2019 international conference on robotics and automation (ICRA). IEEE, 2019, pp. 6023–6029

    2019

  26. [26]

    Feedback control for cassie with deep reinforcement learning,

    Z. Xie, G. Berseth, P. Clary, J. Hurst, and M. Van de Panne, “Feedback control for cassie with deep reinforcement learning,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1241–1246

    2018

  27. [27]

    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,”arXiv preprint arXiv:2507.06905, 2025

    arXiv 2025

  28. [28]

    Proximal policy optimization algorithms,

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

    Pith/arXiv arXiv 2017

  29. [29]

    Humanoid whole-body locomotion on narrow terrain via dynamic balance and reinforcement learning,

    W. Xie, C. Bai, J. Shi, J. Yang, Y . Ge, W. Zhang, and X. Li, “Humanoid whole-body locomotion on narrow terrain via dynamic balance and reinforcement learning,”arXiv preprint arXiv:2502.17219, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Sim-to-real learning of all common bipedal gaits via periodic reward composition,

    J. Siekmann, Y . Godse, A. Fern, and J. Hurst, “Sim-to-real learning of all common bipedal gaits via periodic reward composition,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 7309–7315

    2021

  31. [31]

    Robust mo- tion in-betweening,

    F. G. Harvey, M. Yurick, D. Nowrouzezahrai, and C. Pal, “Robust mo- tion in-betweening,”ACM Transactions on Graphics (TOG), vol. 39, no. 4, pp. 60–1, 2020

    2020

  32. [32]

    Lafan1 retargeting dataset (unitree g1/h1/h1- 2),

    H. Lv, “Lafan1 retargeting dataset (unitree g1/h1/h1- 2),” Hugging Face, 2025. [Online]. Available: https://huggingface.co/datasets/lvhaidong/LAFAN1 Retargeting Dataset

    2025

  33. [33]

    Rectified linear units improve restricted boltzmann machines,

    V . Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” inProceedings of the 27th International Con- ference on Machine Learning (ICML), 2010

    2010

  34. [34]

    Lcm: Lightweight communications and marshalling,

    A. S. Huang, E. Olson, and D. C. Moore, “Lcm: Lightweight communications and marshalling,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2010, pp. 4057–4062

    2010

  35. [35]

    Walk these ways: Tuning robot control for generalization with multiplicity of behavior,

    G. B. Margolis and P. Agrawal, “Walk these ways: Tuning robot control for generalization with multiplicity of behavior,” inConference on Robot Learning. PMLR, 2023, pp. 22–31

    2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.