Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

In controlled comparisons on a real humanoid, including reference joint velocities and a 10-step observation history improves whole-body motion tracking the most, while residual actions, hand-force randomization, and teacher-student trainin

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 11:20 UTC pith:65Q2ORGJ

load-bearing objection Useful controlled ablation, mostly sound, but missing significance tests and a one-run hardware experiment mean the quantitative rankings are not yet settled. the 4 major comments →

arxiv 2607.19903 v1 pith:65Q2ORGJ submitted 2026-07-22 cs.RO

What Matters in Humanoid General Motion Tracking? An Empirical Study

Fabio Amadio (HUCEBOT) , Enrico Mingo Hoffman (HUCEBOT) This is my paper
classification cs.RO
keywords humanoid whole-body controlmotion trackingreinforcement learningmotion imitationablation studyobservation historyresidual actionssim-to-real transfer
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.

This paper asks which practical design choices actually matter when training a humanoid robot to track diverse whole-body motions from human references. To answer it, the authors built a modular framework that lets them change one design factor at a time around a fixed nominal policy, then evaluated each variant on a large set of test motions in simulation and on a real robot. The central findings are that adding reference joint velocities to the motion command and giving the policy a 10-step observation history both reduce tracking errors substantially, while other common choices such as residual actions, stiffer actuation gains, hand-force randomization, and teacher-student training have smaller or mixed effects. The paper also shows that hand-force randomization mainly improves the robot's ability to sustain external loads rather than free-space tracking, and that a mechanics-based actuation profile avoids large torque peaks. A sympathetic reader would take away that the cheapest, most transferable gains come from richer command inputs and temporal context, not from more complex training schemes.

Core claim

The paper claims that, for humanoid general motion tracking, the representation of the motion command and the presence of a short observation history are the design factors with the clearest and largest tracking effects. Removing the reference joint velocities from the motion command degrades all reported tracking metrics (base errors up by 7–14%, key-body errors by 9%, joint-velocity error by 15%), and removing the history buffer degrades them even more (base errors up by 38–50%, joint-velocity error by 17%). Extending history from 10 to 20 steps gives no consistent benefit. Residual actions (offsets around the reference joint positions instead of a default posture) modestly improve key-bod

What carries the argument

The central mechanism is a controlled one-factor-at-a-time ablation protocol built around a nominal policy configuration, with each variant changing a single design axis. The two design elements that carry the main argument are the motion command, which in the nominal setting contains reference joint positions and velocities plus base and key-body references, and a 10-step observation history buffer that concatenates recent proprioception and command vectors. The comparison is made meaningful by a fixed test set of retargeted motions, identical rewards, and the same control frequency, with success rate as a coarse check (all policies succeed) and tracking-error distributions as the fine-grai

Load-bearing premise

The rankings of design choices rest on the assumption that the 1,024 retargeted test motions, split from the same source datasets and subjects as training, are representative of 'general motion' — if those test motions are too similar in style or subject to the training motions, the relative improvements may not transfer to other motion distributions or other humanoid platforms.

What would settle it

Retraining the nominal policy and the no-history / pos-ref-only variants on a held-out test set assembled from subjects and motion styles completely disjoint from training, and showing that the tracking-error gaps shrink or reverse, would falsify the claim that reference joint velocities and a 10-step history are the design choices that matter most.

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

If this is right

  • Including reference joint velocities in the motion command should become a standard practice in humanoid motion-tracking pipelines, since the history encoder does not fully recover the phase and velocity information they carry.
  • A short observation history (around 10 steps) is a robust default: removing it substantially hurts tracking, while going to 20 steps adds little and slightly degrades key-body and joint-space tracking.
  • Residual actions provide only a modest improvement in key-body and joint tracking, so their use should be weighed against algorithmic complexity for a given application.
  • A mechanics-based selection of PD gains and action scales achieves tracking comparable to stiffer hand-tuned gains while avoiding large torque peaks and, on hardware, reducing oscillation after fast transitions.
  • Hand-force randomization during training is valuable specifically for enabling energetic physical interaction with the hands, while leaving free-space tracking performance nearly unaffected.

Where Pith is reading between the lines

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

  • Because removing the history buffer hurt base tracking (38–50%) more than removing reference velocities (7–14%), deployers facing input-size constraints on embedded hardware should prioritize keeping the history buffer over the reference velocity channel.
  • The fact that all policies achieved 100% success on the test set suggests that success rate alone is a weak discriminator for general motion tracking; error distributions, as used here, reveal distinctions that would otherwise be missed, and similar evaluation practices could improve comparability across future systems.
  • The hand-force randomized policy's ability to sustain 6 kg at the hand without a dedicated force controller hints that training-time force randomization could serve as an alternative to impedance-based stiffness tuning for physical interaction, though the paper only demonstrates this for a fixed-arm hold and a single lifting replay.

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 / 4 minor

Summary. The paper proposes YAHMP, an open-source modular framework for training and evaluating whole-body motion tracking policies on the Unitree G1, and uses it to ablate six design axes: motion-command content, observation history length, residual actions, actuation profile, hand-force randomization, and teacher-student training. The nominal policy and its variants are trained with PPO on retargeted AMASS/OMOMO motions (11,151 train / 1,024 test) and evaluated in MuJoCo; TWIST2 is retrained on the same set as an external baseline. Hardware deployment tests zero-shot tracking, balance under perturbation, hand-load capacity, and actuation-profile effects. The main stated findings are that reference joint velocities and a 10-step history improve tracking; residual actions and teacher-student training give modest or mixed changes; a mechanics-based actuation profile limits torque peaks; and hand-force randomization mainly improves interaction capability rather than free-space tracking.

Significance. If the findings are statistically robust, this is a practically useful study for a field where ablations are usually tied to a specific pipeline and individual design choices are hard to isolate. The controlled one-variable-at-a-time protocol, identical reward and evaluation across variants, the retrained external TWIST2 baseline, the open-source implementation, and the zero-shot hardware deployment are genuine strengths. The fixed 1,024-motion test set and the stated pre-registration-like split procedure also support reproducibility. However, the quantitative ranking of choices is not yet established at the level claimed because most effect sizes are small and are reported without confidence intervals, significance tests, or per-seed values. The hardware hand-load result, which carries the paper's strongest interaction claim, is based on a single execution per policy.

major comments (4)
  1. [§III-A, Table II] The headline claims that reference joint velocity and history improve tracking are supported by consistent relative changes across several metrics, but Table II gives no significance tests, confidence intervals, or per-seed values. Several decisive magnitudes are small (4–8%), e.g., No residual key-body pos +5%, History-20 base pos -6%, Hand-force rand. +4%, and all relative changes are computed against a single nominal run. The text says an extra seed gave consistent trends but does not report those numbers. Please report per-seed errors, paired 95% bootstrap intervals across the 1,024 test motions, and a statement on whether each effect replicates across seeds. Without this, the ranking of choices cannot be distinguished from seed-to-seed noise.
  2. [§IV-B, Table IV] The central claim that hand-force randomization mainly improves interaction capability rests on one controlled execution per policy. Loads were increased until the arm no longer reached a steady hold; this stopping criterion is subjective, load order was not randomized, and no repeated trials or automated measurements are reported. At 1 kg the difference is 3.2 vs 1.5 degrees, and at 2 kg 8.3 vs 3.0 degrees, which could plausibly be run-to-run variation. Please provide multiple executions per policy and load, randomized load order, an objective steady-hold criterion, and per-trial data with error bars. Alternatively, the claim should be explicitly downgraded to a qualitative demonstration.
  3. [§II-H and §V] The test set is held out from the same filtered AMASS/OMOMO collection, and Section II-H explicitly allows train and test motions to originate from the same source datasets and subjects. The ranking of design choices may therefore be specific to this in-distribution setting, which is narrower than the 'general motion tracking' claimed in the abstract and introduction. The conclusion's caveat about one robot platform is good, but it does not address dataset/subject overlap. Please add a cross-dataset or held-out-subject evaluation (e.g., whole subjects or datasets excluded from training), or explicitly restrict the conclusions to within-distribution ranking on the Unitree G1.
  4. [§III-B, Table II] The TWIST2 comparison is based on a single retrained run, and the results are mixed: TWIST2 achieves the lowest key-body position error (-10%) but much higher joint-velocity error (+81%). With no repeated seeds or confidence intervals, the comparison does not support strong statements about which complete pipeline is better under the common protocol. At minimum, provide bootstrap intervals across motions and, ideally, multiple training seeds for both pipelines, or soften the comparative conclusion accordingly.
minor comments (4)
  1. [§IV-B, Table VII] The hardware hand-load test reaches 5–6 kg, which exceeds the 20 N cap used for hand-force randomization in Table VII. Please clarify whether this is intentional extrapolation and how the force is transmitted through the arm during the hold.
  2. [§II-H] The statement that the test split was fixed before running ablations is good practice. For full reproducibility, please include the exact dataset snapshot and retargeting version used, so that the claim is verifiable.
  3. [Figure 2] The caption says outliers are hidden, but the display rule is not stated. Please indicate that points beyond the 1.5-IQR whiskers are not plotted and specify the size of the random subset shown as translucent points.
  4. [Table I] The names 'Pos-ref-only' and 'No residual' are defined in the text but Table I alone is ambiguous. A short parenthetical definition in the table caption would help readers who skim the table.

Circularity Check

0 steps flagged

No significant circularity: the paper is a controlled empirical ablation with an external retrained baseline, not a derivation that reduces to its own inputs.

full rationale

The paper's central claims are empirical comparisons among policy variants within a common framework, evaluated on a held-out test set of 1,024 retargeted motions. The variant definitions (motion command with/without reference joint velocities, history length, residual actions, actuation profile, hand-force randomization, teacher-student training) are substantive design choices; none of the reported metrics is constructed from the fitted parameters of the model being evaluated. The external baseline TWIST2 is retrained from its original code on the same training set and evaluated with the same metrics, providing an independent comparison point rather than a self-citation. Self-citations appear only in the future-work section ([19], [20]) and are not load-bearing for any empirical claim. The 'mechanics-based' actuation profile follows BeyondMimic and the residual-action choice follows ResMimic, but these are external prior works and the paper explicitly tests rather than assumes their benefit. The closest concern to circularity would be that the training reward and evaluation metrics both include joint-velocity tracking, but that is a shared objective across all variants and does not by construction favor one variant over another; the ablation result is an empirical finding. The lack of confidence intervals and the single hardware trial per policy are important rigor limitations, but they concern statistical evidence and experimental control, not definitional or self-citational circularity. Under the required standard of quoting a specific reduction from equations or a load-bearing self-citation, no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claims rest on hand-chosen nominal settings (history length, actuation bandwidth, reward weights, force randomization cap) and on the assumption that the AMASS/OMOMO-based test set is a representative 'general motion' distribution. These are not fitted to data, but they define the space in which the ablations are meaningful.

free parameters (6)
  • Actuation bandwidth and damping (ω0 = 2π·10 rad/s, ζ=2) = ω0=62.8 rad/s, ζ=2
    Chosen for the mechanics-based PD gains (Eq. 7); affects the tracking vs torque-peak trade-off that the paper compares.
  • History buffer length H=10 = 10
    Nominal history buffer; the study compares 0, 10, 20 and concludes 10 is a good balance.
  • Hand-force randomization ranges (cap 20 N, Table VII) = max 20 N
    Hand-picked; directly controls the force-interaction capability result in Sec IV-B.
  • Action-scale factor 0.25 in Eq. 8 = 0.25
    Hand-chosen scaling from torque limits; defines the mechanics-based action scales.
  • Reward weights and stds (Table VIII) = listed in Table VIII
    Hand-chosen across all variants; defines the tracking objective and could affect relative design-factor importance.
  • PPO hyperparameters (learning rate, entropy, policy std) = lr 1e-3, entropy 0.005, std 1.0
    Hand-chosen; not fitted, but could interact with design factors; teacher-student uses different lr and std.
axioms (6)
  • domain assumption MuJoCo with 0.005 s timestep and 50 Hz control faithfully models Unitree G1 dynamics for sim-to-real transfer
    Used throughout; zero-shot hardware success supports it but is not a proof across all conditions.
  • domain assumption Retargeted AMASS/OMOMO motions are representative of 'general human motion' for training and evaluation
    Sec II-H; the study's general claims depend on this distribution.
  • domain assumption The train/test split's same-subject overlap does not materially inflate performance
    Sec II-H; not tested by the paper; if false, the ranking could be distribution-specific.
  • domain assumption PPO finds near-optimal policies for each variant given the shared reward
    Sec II-H; standard RL assumption; no guarantee of global optimality, but consistent seed check.
  • standard math PD low-level tracking (Eq. 6) with zero desired velocity is an adequate actuation model
    Common robotics assumption; defines the action interpretation.
  • domain assumption Hand-force randomization forces approximate real interaction loads
    Sec IV-B; hardware test is one execution per policy, so the mapping from training forces to real loads is not validated statistically.

pith-pipeline@v1.3.0-alltime-deepseek · 11127 in / 10472 out tokens · 105021 ms · 2026-08-01T11:20:15.016363+00:00 · methodology

0 comments
read the original abstract

Humanoid general motion tracking requires policies that can follow diverse whole-body references while maintaining balance. Building such policies involves many practical design choices, and their individual effects are often hard to assess. We address this issue with an empirical study of common modeling and training factors used in recent humanoid motion-imitation pipelines. To make the study controlled and reproducible, we developed YAHMP, an open-source modular framework for training, evaluating, and deploying whole-body motion tracking policies on the Unitree G1. Within YAHMP, we define a nominal configuration and compare variants that differ in motion-command representation, observation history, action representation, actuation profile, hand-force randomization during training, and training approach. We evaluate the resulting policies on a test set of retargeted human motions and compare the nominal policy with TWIST2 as an external baseline trained on the same motion set. The results distinguish choices with clear tracking effects from choices that mainly change actuation effort, training complexity, or physical interaction capability. Finally, we deploy YAHMP policies zero-shot on the real Unitree G1, demonstrating diverse whole-body motion tracking, balance under external perturbations, and forceful interaction.

Figures

Figures reproduced from arXiv: 2607.19903 by Enrico Mingo Hoffman (HUCEBOT), Fabio Amadio (HUCEBOT).

Figure 1
Figure 1. Figure 1: Overview of the YAHMP experimental pipeline: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tracking-error distributions over 1,024 test motions. Boxes show median, interquartile range, and 1.5-IQR whiskers. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Execution of the squat motion on a soft mattress [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 7
Figure 7. Figure 7: Measured right ankle pitch angle during a crouching [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Frames from the full-body box-lifting replay. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] 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. Learning Reusable Hybrid Motion Priors for Humanoid Locomotion from Motion Imitation

    cs.LG 2026-07 conditional novelty 5.0

    A frozen hybrid motion prior, distilled from a humanoid motion-imitation expert via residual vector quantization, lets task policies choose discrete codebook entries to perform multiple locomotion tasks and transfers ...

Reference graph

Works this paper leans on

20 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Deepmimic: example-guided deep reinforcement learning of physics-based character skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “Deepmimic: example-guided deep reinforcement learning of physics-based character skills,”ACM Trans. Graph., vol. 37, no. 4, 2018

    2018

  2. [2]

    Learning agile robotic locomotion skills by imitating animals,

    X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine, “Learning agile robotic locomotion skills by imitating animals,”arXiv preprint arXiv:2004.00784, 2020

    Pith/arXiv arXiv 2004

  3. [3]

    Whole-body humanoid robot locomotion with human reference,

    Q. Zhang, P. Cui, D. Yan, J. Sun, Y . Duan, G. Han, W. Zhao, W. Zhang, Y . Guo, A. Zhang, and R. Xu, “Whole-body humanoid robot locomotion with human reference,”arXiv preprint arXiv:2402.18294, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    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

  5. [5]

    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

  6. [6]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Araújo, Z. 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

  7. [7]

    Twist2: Scalable, portable, and holistic humanoid data collection system,

    Y . Ze, S. Zhao, W. Wang, A. Kanazawa, R. Duan, P. Abbeel, G. Shi, J. Wu, and C. K. Liu, “Twist2: Scalable, portable, and holistic humanoid data collection system,”arXiv preprint arXiv:2511.02832, 2025

    arXiv 2025

  8. [8]

    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,” inInternational Conference on Computer Vision, Oct. 2019, pp. 5442–5451

    2019

  9. [9]

    Object motion guided human motion synthesis,

    J. Li, J. Wu, and C. K. Liu, “Object motion guided human motion synthesis,”ACM Trans. Graph., vol. 42, no. 6, 2023

    2023

  10. [10]

    Retargeting matters: General motion retargeting for humanoid motion tracking,

    J. P. Araujo, Y . Ze, P. Xu, J. Wu, and C. K. Liu, “Retargeting matters: General motion retargeting for humanoid motion tracking,”arXiv preprint arXiv:2510.02252, 2025

    arXiv 2025

  11. [11]

    Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control,

    Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath, “Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control,”The International Journal of Robotics Research, vol. 44, no. 5, pp. 840–888, 2025

    2025

  12. [12]

    Resmimic: From general motion tracking to humanoid whole-body loco-manipulation via residual learning,

    S. Zhao, Y . Ze, Y . Wang, C. K. Liu, P. Abbeel, G. Shi, and R. Duan, “Resmimic: From general motion tracking to humanoid whole-body loco-manipulation via residual learning,”arXiv preprint arXiv:2510.05070, 2025

    arXiv 2025

  13. [13]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,

    Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu, “Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,”arXiv preprint arXiv:2508.08241, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Falcon: Learning force-adaptive humanoid loco-manipulation,

    Y . Zhang, Y . Yuan, P. Gurunath, I. Gupta, S. Omidshafiei, A.-a. Agha-mohammadi, M. Vazquez-Chanlatte, L. Pedersen, T. He, and G. Shi, “Falcon: Learning force-adaptive humanoid loco-manipulation,” arXiv preprint arXiv:2505.06776, 2025

    arXiv 2025

  15. [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

    Pith/arXiv arXiv 2024

  16. [17]

    mjlab: A lightweight framework for gpu-accelerated robot learning,

    K. Zakka, Q. Liao, B. Yi, L. L. Lay, K. Sreenath, and P. Abbeel, “mjlab: A lightweight framework for gpu-accelerated robot learning,” 2026

    2026

  17. [18]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2012, pp. 5026–5033

    2012

  18. [19]

    Target-referred dmps for learning bimanual tasks from shared-autonomy telemanipulation,

    F. Amadio, M. Laghi, L. Raiano, F. Rollo, A. Zunino, G. Raiola, and A. Ajoudani, “Target-referred dmps for learning bimanual tasks from shared-autonomy telemanipulation,” in2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), 2022, pp. 496–503

    2022

  19. [20]

    From vocal instructions to household tasks: The inria tiago++ in the eurobin service robots coopetition,

    F. Amadio, C. Donoso, D. Totsila, R. Lorenzo, Q. Rouxel, O. Rochel, E. M. Hoffman, J.-B. Mouret, and S. Ivaldi, “From vocal instructions to household tasks: The inria tiago++ in the eurobin service robots coopetition,”IEEE Robotics and Automation Practice, 2026. APPENDIXA IMPLEMENTATIONDETAILS The following tables summarize the PPO hyperparameters, actuat...

    2026

  20. [2017]

    Available: http://arxiv.org/abs/1707.06347

    [Online]. Available: http://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv