Pith. sign in

REVIEW 6 major objections 8 minor 1 cited by

SMAP: Self-supervised Motion Adaptation for Physically Plausible Humanoid Whole-body Control

T0 review · 6 major / 8 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 A learned motion adapter turns retargeted human motion into physically plausible robot goals, and SMAP reports that this makes humanoid whole-body imitation faster, more accurate, and more stable.

desk verdict SMAP's core idea is solid, but its evaluation compares SMAP against its own adapted goals while baselines track retargeted human motion, so the reported gains are not yet trustworthy. read the letter →

arxiv 2505.19463 v1 pith:UZI5H6UO submitted 2025-05-26 cs.RO

Haoyu Zhao , Sixu Lin , Qingwei Ben , Minyue Dai , Hao Fei , Jingbo Wang , Hua Zou , Junting Dong This is my paper
classification cs.RO
keywords humanoidwhole-bodycontrolmotionadaptationvector-quantizedperiodicautoencoderteacher-studentdistillationreinforcementlearningsim-to-realtransferretargetingphysicalplausibility
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

SMAP claims that a humanoid controller learns to imitate humans much more reliably when the imitation goal is first translated from human motion into physically plausible humanoid motion by a learned adapter, rather than using the raw retargeted human motion as the goal. The paper reports that this translation improves tracking accuracy and velocity tracking, reduces fall counts, and roughly halves the number of training iterations needed to reach a given performance level, with further gains coming from distilling a privileged teacher into a student policy and from a reward that weights upper-body precision against lower-body balance. If true, the framework lets roboticists use large unpaired human motion datasets for whole-body humanoid control without hand-crafting per-motion retargets. The method is evaluated in simulation and demonstrated on a real humanoid robot.

What carries the argument

The central object is the Humanoid-Adapter, a vector-quantized periodic autoencoder built around a phase manifold: an encoder predicts a phase, a frequency, and a quantized amplitude for an input motion sequence, and a decoder reconstructs the motion from embeddings $p = \Psi(\alpha,\phi) = \alpha_0 \sin(2\pi\phi) + \alpha_1 \cos(2\pi\phi)$. Two such autoencoders, one for human motion and one for humanoid robot motion, share a single codebook whose discrete amplitude vectors act as atomic behavior codes, so semantically similar motions are clustered along the same curve of the shared phase manifold. At inference the human encoder is composed with the robot decoder, yielding an adapted humanoid motion without paired supervision. This object carries the argument by converting the imitation goal from a heterogeneous retargeted human trajectory into a robot-space trajectory that the downstream reinforcement-learning policy can track more easily.

What would settle it

Collect paired source-human and adapted-robot motions, run the policy in simulation, and compare the executed robot pose against the original human pose after canonical alignment, alongside speed and amplitude distributions of adapted versus source motions; if robot poses track adapted goals closely but diverge from the human poses, or if the adapter systematically slows or smooths the motion, then the claimed mimicry is partly an artifact of an easier goal.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the heterogeneity between human and humanoid action spaces is a bottleneck for whole-body imitation, and that it can be removed by pretraining a vector-quantized periodic autoencoder on both human motion and simulator-generated humanoid motion, with a shared codebook that forces semantically similar motions to occupy nearby latent codes. At inference, human motion is encoded with the human encoder and decoded with the humanoid decoder, producing an adapted motion that becomes the imitation goal for reinforcement learning. SMAP then trains a privileged teacher policy on this goal with a decoupled upper/lower body reward, and distills it into a student policy that uses only observations available on real hardware. In simulation on the H1 humanoid, SMAP reports the lowest velocity error and the fewest failure terminations among the compared whole-body control methods, and the ablations attribute these gains to the adapter, to teacher-student distillation, and to the decoupled reward. The authors also show qualitative real-world demonstrations of expressive motions including waving, fast walking, and one-legged jumping.

Load-bearing premise

The load-bearing premise is that tracking the adapted humanoid motion counts as mimicking the original human motion; the paper evaluates against the adapted goal and never checks whether it preserves the source human pose.

Editorial extensions

If this is right

  • Using the adapted motion as the imitation goal reaches a given tracking performance in about 2,500 training iterations, roughly half the 4,500 iterations needed with directly retargeted human motion.
  • Feeding the adapter's output into a baseline method reduces failure terminations on novel motion samples, suggesting the adapted goals help policies beyond SMAP itself.
  • Distilling a privileged teacher into a student policy with a decoupled upper/lower body reward improves velocity tracking and upper-body precision while preserving lower-body balance.
  • A codebook size of 32 and an observation history of 10 steps give the best reported velocity error and failure count in the ablation study.
  • Because the adapter is trained on unpaired data, the same pipeline can absorb arbitrary human motion datasets without per-motion retargeting annotations.

Reading between the lines

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

  • Inference: the reported gains should be re-tested with error measured against the original human pose; if the adapter merely makes goals easier by smoothing or slowing motion, the advantage over baselines may shrink.
  • Inference: the shared codebook trained on unpaired human and robot motion could be retrained for other robot morphologies, turning the adapter into a general cross-morphology motion translator, a step the paper does not demonstrate.
  • Inference: the adapted motion could be sampled to generate synthetic training data for downstream tasks beyond goal-conditioned imitation, such as learning motion priors for manipulation; the paper mentions data augmentation only in passing.
  • Inference: the real-world results are qualitative demonstrations, while the quantitative claims are simulation-based; a quantitative real-robot evaluation over many motion samples would be needed to confirm the reported advantages transfer.
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

6 major / 8 minor

Summary. The paper introduces SMAP, a framework for whole-body humanoid control that uses a vector-quantized periodic autoencoder (Humanoid-Adapter) to map human motion into a physically plausible humanoid action space, followed by teacher-student policy distillation with a decoupled reward. The authors evaluate in IsaacGym on Unitree H1 against several state-of-the-art humanoid tracking methods using CMU MoCap data, report ablations of the main components, and show qualitative real-world demonstrations on the H1 robot.

Significance. The paper targets a timely problem—stable whole-body control of humanoid robots from large-scale human motion data—and the central idea is attractive: instead of using retargeted human motion directly as the imitation goal, learn an adapted, robot-friendly goal with a periodic autoencoder. The teacher-student distillation with a decoupled upper/lower-body reward is well-motivated, and the supplement provides useful details on rewards and hardware. If the contributions are validated, the motion-adaptation component could be reused as a data augmentation tool for other humanoid controllers. However, the current evaluation does not yet substantiate the headline claim of 'accurate motion mimicry,' because the adapted goal is both the training target and the evaluation reference, and no measure connects the adapted motion back to the source human behavior. The physical-plausibility claim is also supported only by the provenance of the training data. These gaps are addressable but require additional experiments.

major comments (6)
  1. [Table 1] The evaluation targets are asymmetric: SAMP is trained and evaluated on motion adapted by Humanoid-Adapter, while all baselines are evaluated on retargeted human motion. This means the reported improvements in Evel, Empkpe, and Empjpe could reflect a more easily trackable goal rather than better mimicry of the original human motion. Please evaluate SAMP on the same retargeted human reference as the baselines (e.g., by giving the retargeted human motion as the goal at test time) and report those numbers next to the current results; additionally, report the discrepancy between the adapted motion and the original retargeted motion to show that the adaptation preserves the intended behavior.
  2. [Section 4.1 / Supp. Table 3] The training objective in Eq. (3) applies the reconstruction loss separately to the human and robot datasets, so the shared codebook does not by itself enforce cross-domain semantic alignment; the codebook could partition the two domains independently. The only quantitative adapter check, Supp. Table 3, is a paired reconstruction error on human/robot tracking data, not a measure of semantic or temporal preservation on the CMU motions used in Table 1. The paper's limitation section acknowledges a 'lack of explicit joint correspondence' but does not quantify the resulting mismatch. Please provide a quantitative fidelity check on the actual adapted motions used for training and evaluation, such as per-motion MPJPE between adapted and retargeted human poses, action classification accuracy, or a speed/timing correlation.
  3. [Table 1] The headline margins are small: SAMP vs. SAMP w/o Humanoid-Adapter is 0.1698 vs. 0.1743 on trained motions and 0.2331 vs. 0.2465 on novel motions. No error bars or multiple seeds are reported, so it is unclear whether these differences are significant. Please report mean and standard deviation over at least three seeds and provide a statistical test (e.g., paired bootstrap) for the main comparisons, especially the Humanoid-Adapter ablation.
  4. [Section 4.2] The progressive curriculum is a claimed contributor to convergence speed and stability, but the schedule is not specified: the paper does not state the initial ratio of adapted to retargeted motion, how the ratio is annealed, or any exploration incentives. This omission makes it impossible to reproduce the 'SAMP w/o progressive' ablation or the training-convergence comparison in Section 5.5. Please provide a full curriculum description (or a reference to a detailed supplement).
  5. [Section 5.2] The retargeting method used to produce the 'retargeted human motion' is not described or cited. Since this reference is used for all baselines and is also the input to Humanoid-Adapter, the choice of retargeting affects every number in Table 1. Please state the retargeting algorithm, the joint/keypoint correspondence, and whether the same retargeted data are fed to baselines and to the adapter.
  6. [Section 5.4 / Figure 8] Real-world experiments are presented only qualitatively (photographs and video stills), despite the paper's claim of demonstrating 'superior stability and performance' on hardware. To support this claim, please include quantitative real-world metrics, such as tracking errors or termination counts over a fixed number of trials, or explicitly state that the real-world results are demonstrations rather than controlled evaluations.
minor comments (8)
  1. [Abstract/Section 5] The framework name is introduced as SMAP but appears as SAMP in Table 1 and in several places in Section 5; please unify the abbreviation.
  2. [Eq. (3)] The objective omits the codebook commitment and reinitialization terms that are mentioned in the text; please write the full loss used in training.
  3. [Figure 2] The t-SNE plot would benefit from axis labels, a legend, and a quantitative description of how the distributions were computed.
  4. [Table 1] The first row 'Privileged policy' has lower Evel and Empkpe than SAMP; this row should be clearly identified as the teacher (upper bound) and discussed, since otherwise the reader may wonder why it is not the main result.
  5. [Table 1] The row 'Exbody + Humanoid-Adapter' has dashes for trained-motion results; please explain why these numbers are missing.
  6. [Section 5.5] The phrase 'SAMP w/o teacher-student distillation' is ambiguous; please clarify whether this is the privileged teacher policy or a single-stage policy without distillation.
  7. [Section 5.5] The description of 'SAMP w/o progressive' as 'directly using the final weight' is not precise; please specify what is varied in this ablation.
  8. [Section 4.1] The phrase 'ensuring kinematically feasible and dynamically stable motion' is an unquantified claim; please either soften it or provide measurements (foot slip, ground penetration, joint torque feasibility) on adapted trajectories.

Circularity Check

1 steps flagged · score 6.0 of 10

SMAP's headline accuracy metrics use the adapter's own output as the demonstration/reference while baselines are measured against retargeted human motion, so the reported gains may reflect an easier self-generated goal rather than real mimicry of the source motion.

  1. self definitional [Sec. 5.3 Evaluation Metrics; Sec. 5.5 Ablation Study; Table 1]
    "We evaluate the policy's performance using several metrics computed across all motion sequences in the dataset. The mean linear velocity error (Evel) measures the discrepancy between the robot's root linear velocity and the demonstration ... To evaluate the impact of Humanoid-Adapter, we replace it with directly retargeted human motion (SAMP w/o Humanoid-Adapter) during both training and inference."

    By default SMAP uses the Humanoid-Adapter output as the imitation goal at both training and inference, and the quantitative metrics compare the robot's motion against 'the demonstration' (the same goal). Baselines and SAMP w/o Humanoid-Adapter are instead evaluated against retargeted human motion. Because the adapter is fitted to simulator-generated robot motion, its output can be robot-friendly (e.g., smoothed or slowed), making the self-generated goal systematically easier to track. The reported Evel/Empkpe/Empjpe improvements therefore measure tracking of the adapter's own output, not fidelity to the source human motion; no external metric connects the adapted motion to the original human pose.

full rationale

Score 6: the central comparison in Table 1 is self-referential. For SMAP the demonstration used in all three tracking metrics is the adapted motion produced by Humanoid-Adapter; for baselines and the w/o ablation it is retargeted human motion. Since the adapter is trained on robot motion and can output easier goals, the reported gains are not an independent test of human-motion mimicry. There is no load-bearing self-citation chain: the few author self-citations ([3], [29], [52], [54], [56]) support general context or reward weighting, not the central claim. The 'physically plausible' statement is asserted from training-data provenance rather than demonstrated by foot-slip or torque checks, but that is an unsupported claim, not a circular derivation. The teacher-student distillation and decoupled reward components are evaluated by ablations with the same reference and are not reduced to inputs, but the overall accuracy claim remains compromised by the self-referential evaluation target.

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

The central empirical claim depends on several unvalidated assumptions about the motion adapter's ability to produce physically plausible, semantically faithful motions, plus several hand-tuned hyperparameters (codebook size, history length, reward weights). The paper provides no physics validation of the adapted motion and no measure of human-likeness preservation.

free parameters (4)
  • Codebook size = 32
    Ablation in Fig. 10 shows codebook size 32 yields best Empkpe/faill; part of the central architecture.
  • Observation history length = 10
    Ablation in Fig. 10; history of 10 steps gives best metrics.
  • Decoupled reward weights = Upper DoF 3.0, lower DoF 1.0, upper keypoint 2.0, lower keypoint 1.0, etc.
    Hand-chosen weights in the tracking reward table (Supp Mat Tab. 2) distinguish upper/lower body; the claimed benefit of decoupled reward depends on these weights.
  • Curriculum schedule = Not specified (gradual introduction of retargeted motion)
    The progressive curriculum is claimed to improve training, but the schedule (iteration counts, mixing ratio) is not reported.
assumptions (4)
  • domain assumption The shared codebook between human and humanoid robot motion captures semantically aligned atomic behaviors across the two morphologies.
    The alignment is learned from unpaired reconstruction loss (Eq. 3); no paired supervision or metric confirms that the same code index corresponds to the same behavior.
  • domain assumption The simulator-generated humanoid robot motion dataset S_r is representative of the physically plausible humanoid action space.
    S_r is collected from an RL policy (Exbody) tracking retargeted human data; its coverage of the plausible manifold is untested.
  • ad hoc to paper Adapted motion is kinematically feasible and dynamically stable for the real robot.
    The paper asserts this from being in the robot manifold, but provides no physics validation (contact forces, ZMP, etc.) of the adapted motions.
  • domain assumption VQ-PAE reconstruction error (10.6 cm joint error) is low enough for the adapted goal to be a good imitation target.
    Supp Mat Tab. 3 reports ~10 cm error, which is non-negligible relative to the reported tracking errors (Empkpe ~0.06); its effect on learning is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SMAP: Self-supervised Motion Adaptation for Physically Plausible Humanoid Whole-body Control." pith.science (2026). https://pith.science/paper/UZI5H6UO

@misc{pith2026250519463,
  author       = {Pith},
  title        = {Pith review of: SMAP: Self-supervised Motion Adaptation for Physically Plausible Humanoid Whole-body Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZI5H6UO}},
  note         = {Machine review of arXiv:2505.19463}
}
read the original abstract

This paper presents a novel framework that enables real-world humanoid robots to maintain stability while performing human-like motion. Current methods train a policy which allows humanoid robots to follow human body using the massive retargeted human data via reinforcement learning. However, due to the heterogeneity between human and humanoid robot motion, directly using retargeted human motion reduces training efficiency and stability. To this end, we introduce SMAP, a novel whole-body tracking framework that bridges the gap between human and humanoid action spaces, enabling accurate motion mimicry by humanoid robots. The core idea is to use a vector-quantized periodic autoencoder to capture generic atomic behaviors and adapt human motion into physically plausible humanoid motion. This adaptation accelerates training convergence and improves stability when handling novel or challenging motions. We then employ a privileged teacher to distill precise mimicry skills into the student policy with a proposed decoupled reward. We conduct experiments in simulation and real world to demonstrate the superiority stability and performance of SMAP over SOTA methods, offering practical guidelines for advancing whole-body control in humanoid robots.

Figures

Figures reproduced from arXiv: 2505.19463 by the authors.

Figure 1
Figure 1. Our framework enables humanoid robot execute various expressive whole-body motions. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of the distribution of retargeted human motion, humanoid robot motion (recorded within the simulator), and mo￾tion adapted by Humanoid-Adapter on the CMU MoCap dataset [5]. Traditional approaches, which decompose the prob￾lem into perception, planning, and tracking while modularizing arm and leg control separately [8, 13, 27], are time-consuming to design, limited in scope. These limitations make… view at source ↗
Figure 3
Figure 3. Pipeline of SMAP . Given human motion, we use the proposed Humanoid-Adapter (details shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Humanoid-Adapter. To align heterogeneous human motion S h and humanoid robot motion S r , we train two VQ-PAEs ( ) on them to learn a shared phase manifold using a codebook. We use an encoder to predict the amplitude α, phase ϕ and frequency f, with vector quantization…
Figure 5
Figure 5. Figure 5: Qualitative results on the H1 robot in simulation. w/o Humanoid-Adapter w/ Humanoid-Adapter [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation study. Visualization performance in the simulation on challeng￾ing motion sample. w/o teacher-student w/ teacher-student [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Qualitative result on the H1 robot in real world. 5.4 Comparison with SOTA Methods We evaluate our method on motion tracking in simulation across Unitree H1, comparing it with three state-of-the-art methods: HumanPlus [15], H2O [21], OmniH2O [20], and Exbody [6]. This …
Figure 10
Figure 10. Figure 10: Ablation Study Results (Best values in bold) grades significantly under novel and challenging motion samples, ultimately reducing robot stability. Humanoid-Adapter adapts complex motion into more robot-friendly forms, enabling more stable and accurate motion tracking …
Figure 1
Figure 1. Figure 1: Details about Unitree H1 robot. 9 B More Details of Humanoid Adapter 10 We then assume the two properties hold for the whole input sequence X and extrapolate the phase 11 linearly with the predicted frequency to the whole sequence. We calculate the embeddings using 12 …
Figure 2
Figure 2. Figure 2: Expressive motion evaluation in the real world.. [PITH_FULL_IMAGE:figures/full_fig_p015_2.png]

Discussion (0). Sign in 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. RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation

    cs.RO 2026-07 conditional novelty 6.5 of 10

    A real-time robot-centric video world model driven by depth-aware hand skeletons generates imitation-learning trajectories that support zero-shot real-robot transfer and improve policies when mixed with real data.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    Aberman, P

    K. Aberman, P. Li, D. Lischinski, O. Sorkine-Hornung, D. Cohen-Or, and B. Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62–1, 2020. 3

    work page 2020

  2. [2]

    Agarwal, A

    A. Agarwal, A. Kumar, J. Malik, and D. Pathak. Legged locomotion in challenging terrains using egocentric vision. In Proc. of Conf. on Robot Learning, pages 403–415, 2023. 3

    work page 2023

  3. [3]

    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. 3, 6

    arXiv 2025

  4. [4]

    Brygo, I

    A. Brygo, I. Sarakoglou, N. Garcia-Hernandez, and N. Tsagarakis. Humanoid robot teleoperation with vibrotactile based balancing feedback. In Haptics: Neuroscience, Devices, Modeling, and Applications: 9th International Conference, EuroHaptics 2014, Versailles, France, June 24-26, 2014, Proceedings, Part II 9, pages 266–275. Springer, 2014. 3

    work page 2014

  5. [5]

    Carnegie-Mellon mocap database

    Carnegie Mellon University. Carnegie-Mellon mocap database. http://mocap.cs.cmu.edu/, Mar

    work page

  6. [6]

    Cheng, Y

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang. Expressive whole-body control for humanoid robots. In Proc. of Robotics: Science and Systems, 2024. 2, 3, 4, 6, 8

    work page 2024

  7. [7]

    Cheng, A

    X. Cheng, A. Kumar, and D. Pathak. Legs as manipulator: Pushing quadrupedal agility beyond locomotion. In Proc. of IEEE Int. Conf. on Robotics and Automation, pages 5106–5112. IEEE, 2023. 3

    work page 2023

  8. [8]

    Chestnutt, M

    J. Chestnutt, M. Lau, G. Cheung, J. Kuffner, J. Hodgins, and T. Kanade. Footstep planning for the honda asimo humanoid. In Proc. of IEEE Int. Conf. on Robotics and Automation, pages 629–634, 2005. 2

    work page 2005

Show all 57 references
  1. [9]

    Dariush, M

    B. Dariush, M. Gienger, B. Jian, C. Goerick, and K. Fujimura. Whole body humanoid control from human motion descriptors. In Proc. of IEEE Int. Conf. on Robotics and Automation, pages 2677–2684. IEEE,

  2. [10]

    Darvish, Y

    K. Darvish, Y . Tirupachuri, G. Romualdi, L. Rapetti, D. Ferigo, F. J. A. Chavez, and D. Pucci. Whole-body geometric retargeting for humanoid robots. In IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids), pages 679–686. IEEE, 2019. 3

    2019

  3. [11]

    A. D. Dragan, K. C. Lee, and S. S. Srinivasa. Legibility and predictability of robot motion. In ACM/IEEE International Conference on Human-Robot Interaction, pages 301–308. IEEE, 2013. 3

    2013

  4. [12]

    Escontrela, X

    A. Escontrela, X. B. Peng, W. Yu, T. Zhang, A. Iscen, K. Goldberg, and P. Abbeel. Adversarial motion priors make good substitutes for complex reward functions. In Proc. of IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pages 25–32. IEEE, 2022. 3

    2022

  5. [13]

    S. Feng, E. Whitman, X. Xinjilefu, and C. G. Atkeson. Optimization based full body control for the atlas robot. In IEEE-RAS International Conference on Humanoid Robots, pages 120–127, 2014. 2

    2014

  6. [14]

    Z. Fu, X. Cheng, and D. Pathak. Deep whole-body control: learning a unified policy for manipulation and locomotion. In Proc. of Conf. on Robot Learning, pages 138–149. PMLR, 2023. 3

    2023

  7. [15]

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn. HumanPlus: Humanoid shadowing and imitation from humans. In Proc. of Conf. on Robot Learning, 2024. 2, 3, 6, 8

    2024

  8. [16]

    Fuchioka, Z

    Y . Fuchioka, Z. Xie, and M. Van de Panne. Opt-mimic: Imitation of optimized trajectories for dynamic quadruped behaviors. In Proc. of IEEE Int. Conf. on Robotics and Automation, pages 5092–5098. IEEE,

  9. [17]

    J. W. Grizzle, J. Hurst, B. Morris, H.-W. Park, and K. Sreenath. Mabel, a new robotic bipedal walker and runner. In American Control Conference, pages 2030–2036. IEEE, 2009. 3

    2009

  10. [18]

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng. Generating diverse and natural 3d human motions from text. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 5152–5161,

  11. [19]

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan, et al. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills. arXiv preprint arXiv:2502.01143, 2025. 2, 3

    2025 arXiv

  12. [20]

    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. In Proc. of Conf. on Robot Learning, 2024. 2, 3, 6, 8

    2024

  13. [21]

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi. Learning human-to-humanoid real-time whole-body teleoperation. In Proc. of IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, 2024. 2, 3, 6, 8

    2024

  14. [22]

    Hirai, M

    K. Hirai, M. Hirose, Y . Haikawa, and T. Takenaka. The development of honda humanoid robot. InProc. of IEEE Int. Conf. on Robotics and Automation, volume 2, pages 1321–1326. IEEE, 1998. 3

    1998

  15. [23]

    Holden, J

    D. Holden, J. Saito, and T. Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016. 3 10

    2016

  16. [24]

    H. Ito, K. Yamamoto, H. Mori, and T. Ogata. Efficient multitask learning with an embodied predictive model for door opening and entry with whole-body control. Science Robotics, 7(65):eaax8177, 2022. 3

    2022

  17. [25]

    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. 2, 3

    2024 arXiv

  18. [26]

    S. Kim, M. Sorokin, J. Lee, and S. Ha. Humanconquad: human motion control of quadrupedal robots using deep reinforcement learning. In SIGGRAPH Asia 2022 Emerging Technologies, pages 1–2. 2022. 3

    2022

  19. [27]

    Kuindersma, R

    S. Kuindersma, R. Deits, M. Fallon, A. Valenzuela, H. Dai, F. Permenter, T. Koolen, P. Marion, and R. Tedrake. Optimization-based locomotion planning, estimation, and control design for the atlas humanoid robot. Autonomous robots, 40:429–455, 2016. 2

    2016

  20. [28]

    Lee and S

    J. Lee and S. Y . Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 39–48,

  21. [29]

    H. Li, Q. Zhao, H. Xu, X. Jiang, Q. Ben, F. Jia, H. Zhao, L. Xu, J. Zeng, H. Wang, et al. Teleopbench: A simulator-centric benchmark for dual-arm dexterous teleoperation. arXiv preprint arXiv:2505.12748, 2025. 2

    2025

  22. [30]

    P. Li, S. Starke, Y . Ye, and O. Sorkine-Hornung. Walkthedog: Cross-morphology motion alignment via phase manifolds. In ACM SIGGRAPH 2024 Conference Papers, pages 1–10, 2024. 3, 4

    2024

  23. [31]

    T. Li, J. Won, A. Clegg, J. Kim, A. Rai, and S. Ha. Ace: Adversarial correspondence embedding for cross morphology motion retargeting from human to nonhuman characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1–11, 2023. 3

    2023

  24. [32]

    C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang. Mobile-television: Predictive motion priors for humanoid whole-body control. arXiv preprint arXiv:2412.07773, 2024. 2, 3

    2024 arXiv

  25. [33]

    Mahmood, N

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black. Amass: Archive of motion capture as surface shapes. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 5442–5451,

  26. [34]

    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. 3, 6

    2021 arXiv

  27. [35]

    Miura and I

    H. Miura and I. Shimoyama. Dynamic walk of a biped. The International Journal of Robotics Research, 3(2):60–74, 1984. 3

    1984

  28. [36]

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (TOG), 40(4):1–20, 2021. 6, 8

    2021

  29. [37]

    Peternel and J

    L. Peternel and J. Babiˇc. Learning of compliant human–robot interaction using full-body haptic interface. Advanced Robotics, 27(13):1003–1012, 2013. 3

    2013

  30. [38]

    A. R. Punnakkal, A. Chandrasekaran, N. Athanasiou, A. Quiros-Ramirez, and M. J. Black. Babel: Bodies, action and behavior with english labels. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 722–731, 2021. 2

    2021

  31. [39]

    S. Raab, I. Leibovitch, P. Li, K. Aberman, O. Sorkine-Hornung, and D. Cohen-Or. Modi: Unconditional motion synthesis from diverse data. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 13873–13883, 2023. 3

    2023

  32. [40]

    Radosavovic, T

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath. Real-world humanoid locomotion with reinforcement learning. Science Robotics, 9(89):eadi9579, 2024. 3

    2024

  33. [41]

    Radosavovic, B

    I. Radosavovic, B. Zhang, B. Shi, J. Rajasegaran, S. Kamat, T. Darrell, K. Sreenath, and J. Malik. Humanoid locomotion as next token prediction. In Proc. of Advances in Neural Information Processing Systems ,

  34. [42]

    Ramos and S

    J. Ramos and S. Kim. Dynamic locomotion synchronization of bipedal robot and human operator via bilateral feedback teleoperation. Science Robotics, 4(35):eaav4282, 2019. 3

    2019

  35. [43]

    S. Ross, G. Gordon, and D. Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, 2011. 6

    2011

  36. [44]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4

    2017 arXiv

  37. [45]

    Starke, I

    S. Starke, I. Mason, and T. Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1–13, 2022. 3, 4

    2022

  38. [46]

    Tak and H.-S

    S. Tak and H.-S. Ko. A physically-based motion retargeting filter. ACM Transactions on Graphics (ToG), 24(1):98–117, 2005. 3

    2005

  39. [47]

    Villegas, J

    R. Villegas, J. Yang, D. Ceylan, and H. Lee. Neural kinematic networks for unsupervised motion retargetting. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 8639–8648, 2018. 3

    2018

  40. [48]

    E. R. Westervelt, J. W. Grizzle, and D. E. Koditschek. Hybrid zero dynamics of planar biped walkers. IEEE transactions on automatic control, 48(1):42–56, 2003. 3

    2003

  41. [49]

    R. Yang, G. Yang, and X. Wang. Neural volumetric memory for visual locomotion control. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 1430–1440, 2023. 3 11

    2023

  42. [50]

    S. Yang, Z. Wu, M. Li, Z. Zhang, L. Hao, W. Bao, and H. Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pages 2321–2330, 2023. 3

    2023

  43. [51]

    K. Yin, K. Loken, and M. Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105–es, 2007. 3

    2007

  44. [52]

    H. Zhao, H. Wang, C. Yang, and W. Shen. Chase: 3d-consistent human avatars with sparse inputs via gaussian splatting and contrastive learning. arXiv preprint arXiv:2408.09663, 2024. 2

    2024 arXiv

  45. [53]

    H. Zhao, H. Wang, X. Zhao, H. Wang, Z. Wu, C. Long, and H. Zou. Automated 3d physical simulation of open-world scene with gaussian splatting. arXiv preprint arXiv:2411.12789, 2024. 2

    2024 arXiv

  46. [54]

    H. Zhao, C. Yang, H. Wang, X. Zhao, and W. Shen. Sg-gs: Photo-realistic animatable human avatars with semantically-guided gaussian splatting. arXiv preprint arXiv:2408.09665, 2024. 2

    2024 arXiv

  47. [55]

    Zheng and A

    C. Zheng and A. Vedaldi. Online clustered codebook. In Proc. of IEEE Intl. Conf. on Computer Vision, pages 22798–22807, 2023. 5

    2023

  48. [56]

    Zhuang, S

    Z. Zhuang, S. Yao, and H. Zhao. Humanoid parkour learning. arXiv preprint arXiv:2406.10759, 2024. 3 12 SMAP : Self-supervised Motion Adaptation for Physically Plausible Humanoid Whole-body Control Anonymous Author(s) Affiliation Address email A Real Robot System Setup1 Our rea...

    2024 arXiv

  49. [2007]

    2, 4, 6, 8

    [Online]. 2, 4, 6, 8

Pith tools

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