Pith. sign in

REVIEW 3 major objections 5 minor 43 references

The paper claims that the classical capture point becomes deployable once it is expressed relative to the support foot, because the unmeasurable base linear velocity cancels exactly, and that a policy given this support-relative dynamic-CoM

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 →

A support-relative dynamic capture-point observation, reconstructible without base linear velocity, lets a humanoid policy hold clean single-leg balance at 86/90 in simulation and deploy on a Unitree G1 without distillation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Support-relative capture point is a genuine deployability trick and the sim2sim benchmark is a useful contribution, but the paper asserts rather than proves that the on-board reconstruction matches the training-time observation. the 3 major comments →

arxiv 2608.00500 v1 pith:QYIFOVU2 submitted 2026-08-01 cs.RO

First Deployable Dynamic-CoM: A Unified Policy and Method-Agnostic Benchmark for Humanoid Single-Leg Balance

classification cs.RO
keywords humanoid single-leg balancecapture pointsupport-relative dynamic CoMsim2sim benchmarkpostural controlreinforcement learningsim-to-real transferhumanoid balance
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

Single-leg balance is the skill general humanoid policies lack most: on the paper's benchmark, eight released generalist policies score zero clean single-leg holds out of 90 motions, staying up only by stepping or hopping. The paper's central claim is that the reason is an observation gap, not a training gap. The classical capture point, the CoM extrapolated by its velocity, needs the body's absolute linear velocity, which no onboard sensor provides, so it has only ever fed a training reward or privileged critic. Expressed relative to the support foot, that velocity cancels exactly, leaving a dynamic-CoM state, position and velocity of the CoM with respect to the foot, reconstructible from joint encoders and an IMU. Putting this support-relative dynamic-CoM observation directly into the deployed actor, surrounded by rewards translated from human postural control, yields a policy that holds a clean single-leg stance on 86 of 90 held-out motions and transfers to a real humanoid without teacher-student distillation; the paper also releases a sim2sim benchmark that makes such claims measurable.

Core claim

The paper establishes that the support-relative capture point state (r^B, dot r^B), the CoM's horizontal position and velocity relative to the support foot in the torso frame, is the deployable form of the dynamic balance signal. In this frame the base linear velocity v_b enters the CoM velocity and the support-center velocity identically and cancels (Appendix A, Eq. 6), so dot r^B = omega^B x (d_c - d_s) + (J_c - J_s) dot q, a function of only the gyro, joint encoders, and the kinematic/mass model. Therefore a policy trained on the exact simulator value can consume the on-robot reconstruction unchanged. The evidence: ablations that remove the dynamic observation are the largest single drive

What carries the argument

The load-bearing object is the support-relative dynamic-CoM observation o_bal = (r^B, dot r^B) in R^4: the horizontal CoM position and velocity relative to the support foot, expressed in the base frame. It is the capture point xi = c + dot c / omega_0, with omega_0 = sqrt(g/h), rewritten as xi - s approximately r + dot r / omega_0; position alone cannot disambiguate an edge-of-foot CoM moving inward from one moving outward. The identity in Eq. 6 does the deployability work: because v_b cancels, dot r^B needs only the torso gyro, encoder positions and velocities, and the mass model, so the same vector appears in simulation and on hardware. Around this observation the paper places a reward lib

Load-bearing premise

The on-board reconstruction of the support-relative CoM velocity, via encoder velocities, gyro, mass model, and the gravity-aligned support-foot rule, matches the exact simulator state used in training; this equivalence is asserted from the cancellation identity, not measured on hardware.

What would settle it

On a real G1, run the deployed checkpoint through several of the 90 test motions while a motion-capture rig records true torso and foot positions; compute the per-step difference between the on-board reconstructed (r^B, dot r^B) and motion-capture ground truth, and compare it with the observation noise used in training, capture-point noise 0.015 units and dof-velocity noise 0.5. If the measured reconstruction error is larger than the noise envelope, or if clean-hold success in that run falls below the sim2sim rate, the deployability claim fails.

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

If this is right

  • Actors, not just critics, can be trained on the dynamic balance signal: any future humanoid balance policy can consume the support-relative dynamic-CoM state directly from standard sensors, removing the distillation step.
  • Single-leg balance becomes a measurable, reproducible capability: the sim2sim benchmark gives a common yardstick with Perfect/Marginal/Fall tiers plus MoS and TTB diagnostics, so published balance claims can be compared under byte-identical conditions.
  • General motion-tracking scale by itself does not confer clean single-leg balance: eight strong released generalists all score 0/90 Perfect, so the architecture needs a balance-specific observation and reward block.
  • The dynamic-CoM observation is a prerequisite for learning the task, not only for deployment: without it training plateaus near 380-420 of 500 steps, so other policies trained without this state are unlikely to fit single-leg balance motions.
  • Checkpoint selection by measured sim2sim competence rather than training reward is necessary: training reward saturates and is essentially uncorrelated with Perfect success, so the deployed checkpoint is neither the final nor the highest-reward one.

Where Pith is reading between the lines

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

  • Because the cancellation is frame-based rather than task-specific, the same support-relative dynamic-CoM observation should extend to other single-support phases such as walking stance, stance on uneven or soft terrain, and push recovery, as long as a support reference can be defined; the paper does not test these.
  • A direct test of the deployability claim would be a motion-capture ground-truth comparison on hardware: the paper reports only qualitative deployment, and Appendix I notes that reconstruction error is not quantified, so the equivalence of simulator and on-board vectors is an assumption to verify.
  • The support-foot selection rule, a gravity-aligned foot-height threshold with a 3cm double-support cutoff, is the main source of train/deploy mismatch; perturbing that threshold in the benchmark could predict how much contact transients degrade clean holds.
  • The reward library terms are imported from human postural control; the same terms could be evaluated on quiet two-leg stance or sit-to-stand to see which postural principles transfer across tasks.
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

3 major / 5 minor

Summary. The paper introduces FDDC, a single policy for humanoid single-leg balance trained with asymmetric FastSAC, and a method-agnostic sim2sim benchmark. The central claim is that the support-relative dynamic-CoM state (r^B, dot r^B), in which the base linear velocity cancels exactly, is reconstructible from joint encoders and an IMU alone, so the deployed actor can consume the same balance state it saw in training without teacher-student distillation. On the benchmark, FDDC achieves 86/90 Perfect holds on held-out stratified motions, while eight released generalist policies achieve 0/90. The paper reports ablations isolating the dynamic-CoM observation as the main driver and a qualitative real-robot deployment on a Unitree G1.

Significance. If the deployability claim holds, this is a meaningful advance: it gives the first actor-side, on-board-reconstructible capture-point observation for humanoid balance, with a clean kinematic derivation in Appendix A and a reproducible benchmark protocol. The held-out test set, stratified pose grid, and monotone ablations are strengths, and the 0/90 baseline result is a striking, falsifiable finding. However, the benchmark's reproducibility currently rests on a promised release without an artifact link, and the hardware evidence is qualitative. The core correctness of the deployment-time observation equivalence is not quantitatively verified, which is the main risk to the paper's central claim.

major comments (3)
  1. [Appendix A, Eq. (3)-(6); Sec. 3.1] The cancellation of the base linear velocity is exact only if the support center s is a material point whose velocity is the weighted sum of foot-link velocities. In simulation, C and the weights w_i in Eq. (3) come from the physics engine's contact set; on hardware, the support foot is selected by a gravity-aligned foot-height rule (lower foot; feet within 3 cm count as double support, Appendix I). The center of pressure is not a fixed material point, so its velocity contains a contribution from the pressure-distribution shift that cannot be reconstructed from joint encoders and gyro alone. During double-support transients or when the swing foot is near the ground, the height rule and the physical contact centroid can disagree. The statement in Sec. 3.1 that 'the simulator value and the on-robot reconstruction are the same vector' is therefore not established. The paper should quantify
  2. [Sec. 5.6; Appendix I] The real-robot deployment is reported only qualitatively ('the transfer is stable'), and the paper itself states in Sec. 6 that real-robot results are qualitative. Since the central contribution is a 'deployable' dynamic-CoM observation, the hardware evidence should include at least per-pose success counts, stance durations, and, ideally, a comparison of the reconstructed observation with an independent estimate. Without such numbers, the claim that the dynamic-CoM observation transfers to hardware without distillation is not quantitatively supported.
  3. [Abstract; Sec. 1; Sec. 6] The paper repeatedly states 'We release the full stack—data, code, policy, and benchmark' and describes the benchmark as 'reproducible,' but no artifact link, repository identifier, or DOI appears in the manuscript. For a benchmark paper, reproducibility is a central claim, and the absence of the artifact makes the benchmark unusable and unverifiable. A URL or anonymous repository link must be provided.
minor comments (5)
  1. [Tables 1 and 2] All success rates are point estimates from n=90 clips (or n=10 for per-class cells). Exact binomial confidence intervals or standard errors across seeds would be helpful, especially for the smaller ablation gaps (e.g., ankle action-rate penalty, -4.5 pt).
  2. [Figure 3] The claim that training reward is 'essentially uncorrelated' with Perfect success is not quantified. Report the Spearman correlation or a scatter plot with the coefficient.
  3. [Sec. 4.3] The definition of Failure includes 'holds a leg but drifts outside the 0.5 m tracking gate,' which is described as an edge case that is empirically empty. It would be clearer to state that no trial fell into this sub-case, rather than folding it into the tier definition without evidence.
  4. [Appendix A, Eq. (6)] The derivation uses the notation J_c and J_s but does not explicitly define the Jacobians for the support center when the support region changes continuously (e.g., CoP motion within the foot). Clarify the assumption that s is a material combination of foot links, and state that CoP motion is neglected in the deployable observation.
  5. [Appendix D, Table 6] The table lists 'Action delay (training) disabled' but no discussion of whether the deployed system has any latency that could induce a distribution shift. A brief comment on this would be useful.

Circularity Check

1 steps flagged

No fatal circularity; the dynamic-CoM derivation is self-contained. One evaluative overlap: the benchmark's continuous diagnostics are the same quantities the reward library optimizes, so the 'biomechanically genuine' claim is partly self-confirming.

specific steps
  1. other [Sec. 3.2 (Table 5) vs Sec. 4.3 / Appendix F; invoked in Sec. 5.3]
    "Alongside the tiers we report continuous diagnostics grounded in the same postural-control physics as our method: Margin of Stability and time-to-boundary on the balance axis, plus HuB tracking errors, support-foot slippage, jerk, and time-to-fall. ... These metrics share the human-postural-control grounding of our observations and rewards (Sec. 3.1–3.2)."

    The reward library (Sec. 3.2, Table 5) directly penalizes capture-point margin (MoS), time-to-boundary (TTB), and action jerk. The benchmark's continuous diagnostics report exactly these same quantities as evidence that FDDC's hold is 'biomechanically genuine on every axis' (Sec. 5.3). Since the policy was trained to optimize these reward terms, reporting the matching diagnostics as independent validation is partly self-confirming: the metric values reflect the training objective rather than an external test of it. The outcome tiers (Perfect/Marginal/Fail) are defined independently and do not reduce to the reward terms, so the circularity is partial and does not affect the mathematical derivation of the dynamic-CoM observation.

full rationale

The paper's core derivation chain is not circular. Appendix A derives the support-relative dynamic-CoM observation (r^B, dot r^B) as a change of frame from the classical capture point; the base linear velocity cancels by algebra, and the result is reconstructible from encoders, gyro, and forward kinematics. No fitted parameter is renamed as a prediction, and no load-bearing self-citation chain forces the conclusion. The ablation that removes the observation is an empirical experiment, not a definitional tautology, and the held-out test protocol is standard. The one genuine circularity concern is evaluative: the benchmark's continuous diagnostics (MoS, TTB, jerk) are literally the same postural-control quantities encoded in the reward library, so claiming biomechanical genuineness from those metrics is partly self-confirming. However, the main success criterion (clean single-leg balance: no hop, no touchdown, no fall, tracking within gate) is independent of the reward terms, and the hardware transfer claim, while only qualitative, is a separate empirical matter rather than a logical reduction. The score of 3 reflects this partial self-confirmation without alleging that the central derivation is circular.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim depends on the LIP/capture-point model, the support-contact estimate, the foreign-simulator fairness assumption, and a hand-tuned reward/benchmark configuration. None of these is independently measured or verified in the manuscript; the most sensitive undisclosed choice is the LIP height h, with the reward weights and bin edges also hand-chosen.

free parameters (5)
  • LIP height h (or omega0 = sqrt(g/h)) = not disclosed
    Defines the capture point and the MoS/TTB rewards; the numerical h used in training and evaluation is not listed in Table 6.
  • MoS safety margin and TTB reaction threshold = 2 cm/3 cm support; 0.3 s/0.2 s
    Hand-chosen reward thresholds in Table 5 and Appendix C; no sensitivity analysis is reported.
  • Reward weight vector = e.g., MoS -20, TTB -15, ankle -0.3, knee -0.1, jerk -0.1
    All weights in Table 6 are manually set; the ablation shows several materially affect performance, so the configuration is tuned by hand.
  • Stratified pose bin edges = pelvis 0.38/0.51/0.65/0.78 m; swing foot 0.09/0.28/0.48/0.67 m
    Hand-chosen grid edges determine class balance, difficulty, and the reported 90-clip test set.
  • Domain randomization noise magnitudes = e.g., dof velocity 0.5, capture point 0.015
    Training-time noise levels are hand-set and directly affect robustness numbers, especially the noisy K=10 column.
axioms (6)
  • domain assumption Single-leg balance is well approximated by a linear inverted pendulum with constant CoM height h, so the capture point condition determines balance.
    Invoked in Sec. 3.1 with Eq. (1); h is never specified and squat-depth variation challenges the constant-height assumption.
  • standard math Support center is computable from contact-weighted foot positions and whole-body CoM from kinematic/mass model; base linear velocity cancels in the relative difference.
    Appendix A Eqs. (3)-(6); relies on rigid-body kinematics and known link masses.
  • domain assumption The shared MuJoCo kernel with a 50 Hz LowState to LowCmd contract is a fair and deployment-predictive evaluation for policies trained in different simulators.
    Sec. 4.2; the 0/90 baseline result depends on this premise, which is not validated in the paper.
  • domain assumption On hardware, the support foot can be inferred by the gravity-aligned lower-foot rule and matches the training-time support center.
    Appendix I; the paper does not compare the height-rule support estimate to the engine's contact-weighted support center.
  • domain assumption The AMS synthetic balance-motion dataset provides feasible, representative reference motions for single-leg balance.
    Sec. 4.1 and Appendix E; 186 clips are discarded for ground penetration, and the pose grid is cut from this one source.
  • domain assumption Human postural-control principles (MoS, TTB, ankle-to-hip hierarchy, jerk) transfer to RL reward terms without per-robot recalibration.
    Sec. 3.2 and Table 5; central to the prevention-over-repair design and to the benchmark's diagnostic suite.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of First Deployable Dynamic-CoM: A Unified Policy and Method-Agnostic Benchmark for Humanoid Single-Leg Balance." pith.science (2026). https://pith.science/paper/QYIFOVU2

@misc{pith2026260800500,
  author       = {Pith},
  title        = {Pith review of: First Deployable Dynamic-CoM: A Unified Policy and Method-Agnostic Benchmark for Humanoid Single-Leg Balance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYIFOVU2}},
  note         = {Machine review of arXiv:2608.00500}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unified humanoid policies handle agile whole-body motion, yet stumble on a simple demand: staying balanced on one leg. On our single-leg-balance benchmark, eight released state-of-the-art general policies hold a clean single-leg stance on 0 of 90 test motions; they stay up only by stepping or hopping, recovering from imbalance rather than preventing it. Prevention needs the capture point (xCoM), the center of mass (CoM) extrapolated by its velocity, which has never driven a hardware policy because it requires a base linear velocity no on-board sensor provides; expressed relative to the support foot, that velocity cancels exactly, leaving an observation reconstructible from encoders and IMU alone. We put this first deployable dynamic-CoM observation directly into the actor that runs on hardware, and pair it with a reward library translated term by term from human postural control, under one principle: prevention over repair. Trained by asymmetric FastSAC with a privileged critic and no distillation, the resulting policy, FDDC (First Deployable Dynamic-CoM), holds clean single-leg balance on 86 of 90 held-out motions across nine stratified pose classes and transfers to a real Unitree G1; in ablation, the dynamic-CoM observation is the single largest driver: removing it alone costs 40 points of clean single-leg balance. We release the full stack with the first method-agnostic, reproducible sim2sim benchmark for humanoid single-leg balance, scoring each policy in a simulator distinct from its training one, a step toward turning balance from a per-task trick into a capability the field can measure.

Figures

Figures reproduced from arXiv: 2608.00500 by Bozhou Chen, Jieming Cui, Wenxin Li, Xingyun Wang, Yikai Fan, Yikai Zhou, Yixin Zhu.

Figure 1
Figure 1. Figure 1: FDDC produces stable single-leg balance across a wide range of poses that vary in squat depth, swing-foot height, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. One unified single-leg policy is trained by asymmetric FastSAC: a deployable actor (proprioception, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Checkpoint selection on the benchmark. Top: train [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training-time mean episode length. The full policy [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-robot deployment. The unified single-leg pol [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-class Failure (fall) rate over the 3 × 3 pose grid (pelvis × swing-foot height; 10 held-out clips/class; clean). (a) FDDC falls in only one of the nine classes, in 10% of its clips (1/10), and never in the other eight. (b) The eight-baseline mean rises steeply from the easy corner (high pelvis, low swing foot) to the hard one (deep squat, high swing foot): difficulty emerges along both axes, and where … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 20 canonical work pages · 2 internal anchors

  1. [1]

    Al-Hafez, F.; Zhao, G.; Peters, J.; and Tateo, D. 2023. LocoMuJoCo : A Comprehensive Imitation Learning Benchmark for Locomotion. arXiv preprint arXiv:2311.02496. NeurIPS 2023 Robot Learning Workshop

  2. [2]

    Amazon FAR . 2025. Holosoma . https://github.com/amazon-far/holosoma. Software framework

  3. [3]

    A.; Humphreys, J.; Mingo Hoffman, E.; Fernandez Talavera, N.; Rodriguez Sanchez, M

    Castano, J. A.; Humphreys, J.; Mingo Hoffman, E.; Fernandez Talavera, N.; Rodriguez Sanchez, M. C.; and Zhou, C. 2022. Benchmarking Dynamic Balancing Controllers for Humanoid Robots. Robotics, 11(5): 114

  4. [4]

    Chen, M.; Wang, K.; Zhang, B.; Ma, X.; Yang, Z.; Ren, Y.; Huang, Q.; Zhu, Z.; Wang, Y.; and Su, Z. 2026. HoloMotion-1 Technical Report. arXiv preprint arXiv:2605.15336

  5. [5]

    B.; and Wang, X

    Chen, Z.; Ji, M.; Cheng, X.; Peng, X.; Peng, X. B.; and Wang, X. 2025. GMT : General Motion Tracking for Humanoid Whole-Body Control. arXiv preprint arXiv:2506.14770

  6. [6]

    Cheng, X.; Ji, Y.; Chen, J.; Yang, R.; Yang, G.; and Wang, X. 2024. Expressive Whole-Body Control for Humanoid Robots. arXiv preprint arXiv:2402.16796. Robotics: Science and Systems (RSS) 2024

  7. [7]

    Fu, Z.; Zhao, Q.; Wu, Q.; Wetzstein, G.; and Finn, C. 2024. HumanPlus : Humanoid Shadowing and Imitation from Humans. arXiv preprint arXiv:2406.10454. Conference on Robot Learning (CoRL) 2024

  8. [8]

    He, T.; Luo, Z.; He, X.; Xiao, W.; Zhang, C.; Zhang, W.; Kitani, K.; Liu, C.; and Shi, G. 2024. OmniH2O : Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning. arXiv preprint arXiv:2406.08858. Conference on Robot Learning (CoRL) 2024

  9. [9]

    Hertel, J.; and Olmsted-Kramer, L. C. 2007. Deficits in time-to-boundary measures of postural control with chronic ankle instability. Gait & Posture, 25(1): 33--39

  10. [10]

    L.; Gazendam, M

    Hof, A. L.; Gazendam, M. G. J.; and Sinke, W. E. 2005. The condition for dynamic stability. Journal of Biomechanics, 38(1): 1--8

  11. [11]

    Kajita, S.; Kanehiro, F.; Kaneko, K.; Yokoi, K.; and Hirukawa, H. 2001. The 3D Linear Inverted Pendulum Mode: A simple modeling for a biped walking pattern generation. In Proceedings of the 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), volume 1, 239--246

  12. [12]

    Kim, J.; Kim, W.; Lu, Y.; Cheng, J.; Zargarbashi, F.; Zeng, Z.; Qi, Z.; Dou, Z.; Sontakke, N.; Baek, D.; Ha, S.; and Li, T. 2025. Switch-JustDance: Benchmarking Whole Body Motion Tracking Controllers Using a Commercial Console Game. arXiv preprint arXiv:2511.17925

  13. [13]

    Lei, Z.; Cao, L.; Wang, X.; Qian, T.; Liu, J.; and Li, X. 2026. A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking. arXiv preprint arXiv:2602.13656

  14. [14]

    E.; Huang, X.; Gao, Y.; Tevet, G.; Sreenath, K.; and Liu, C

    Liao, Q.; Truong, T. E.; Huang, X.; Gao, Y.; Tevet, G.; Sreenath, K.; and Liu, C. K. 2025. BeyondMimic : From Motion Tracking to Versatile Humanoid Control via Guided Diffusion. arXiv preprint arXiv:2508.08241

  15. [15]

    Luo, Z.; Wang, J.; Liu, K.; Zhang, H.; Tessler, C.; Wang, J.; Yuan, Y.; Cao, J.; Lin, Z.; Wang, F.; Hodgins, J.; and Kitani, K. 2024. SMPLOlympics : Sports Environments for Physically Simulated Humanoids. arXiv preprint arXiv:2407.00187

  16. [16]

    Luo, Z.; Yuan, Y.; Wang, T.; Li, C.; Casta \ n eda, F.; Chen, S.; Cao, Z.-A.; Li, J.; Minor, D.; Ben, Q.; Park, J.; Sami, D.; Wang, Z.; Da, X.; Ding, R.; Hogg, C.; Song, L.; Lim, E.; Jeong, E.; He, T.; Xue, H.; Xiao, W.; Yuen, S.; Kautz, J.; Chang, Y.; Iqbal, U.; Fan, L.; and Zhu, Y. 2025. SONIC : Supersizing Motion Tracking for Natural Humanoid Whole-Bod...

  17. [17]

    Massion, J. 1992. Movement, posture and equilibrium: interaction and coordination. Progress in Neurobiology, 38(1): 35--56

  18. [18]

    O.; and Hertel, J

    McKeon, P. O.; and Hertel, J. 2008 a . Systematic Review of Postural Control and Lateral Ankle Instability, Part I : Can Deficits Be Detected With Instrumented Testing? Journal of Athletic Training, 43(3): 293--304

  19. [19]

    O.; and Hertel, J

    McKeon, P. O.; and Hertel, J. 2008 b . Systematic Review of Postural Control and Lateral Ankle Instability, Part II : Is Balance Training Clinically Effective? Journal of Athletic Training, 43(3): 305--315

  20. [20]

    Mergner, T.; and Lippi, V. 2018. Posture Control---Human-Inspired Approaches for Humanoid Robot Benchmarking: Conceptualizing Tests, Protocols and Analyses. Frontiers in Neurorobotics, 12: 21

  21. [21]

    G.; Bicchi, A.; and Garabini, M

    Monteleone, S.; Negrello, F.; Grioli, G.; Catalano, M. G.; Bicchi, A.; and Garabini, M. 2023. A method to benchmark the balance resilience of robots. Frontiers in Robotics and AI, 9: 817870

  22. [22]

    Pan, Y.; Qiao, R.; Chen, L.; Chitta, K.; Pan, L.; Mai, H.; Bu, Q.; Zhao, H.; Zheng, C.; Luo, P.; and Li, H. 2025. Agility Meets Stability: Versatile Humanoid Control with Heterogeneous Data. arXiv preprint arXiv:2511.17373

  23. [23]

    B.; Abbeel, P.; Levine, S.; and van de Panne, M

    Peng, X. B.; Abbeel, P.; Levine, S.; and van de Panne, M. 2018. DeepMimic : Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills. ACM Transactions on Graphics, 37(4): 143:1--143:14

  24. [24]

    B.; Ma, Z.; Abbeel, P.; Levine, S.; and Kanazawa, A

    Peng, X. B.; Ma, Z.; Abbeel, P.; Levine, S.; and Kanazawa, A. 2021. AMP : Adversarial Motion Priors for Stylized Physics-Based Character Control. ACM Transactions on Graphics, 40(4): 144:1--144:20

  25. [25]

    E.; and Griffin, R

    Poddar, N.; McCrory, S.; Penco, L.; Clark, G.; Svil, H. E.; and Griffin, R. 2026. Embedding Classical Balance Control Principles in Reinforcement Learning for Humanoid Recovery. arXiv preprint arXiv:2603.08619

  26. [26]

    Pratt, J.; Carff, J.; Drakunov, S.; and Goswami, A. 2006. Capture Point: A Step toward Humanoid Push Recovery. In 2006 6th IEEE-RAS International Conference on Humanoid Robots (Humanoids), 200--207

  27. [27]

    Qi, Z.; Chen, X.; Liu, D.; Lin, C.; Lian, Y.; Liang, S.; Zhang, Z.; Guan, Y.; Wang, J.; Zhang, W.; Yu, X.; Wang, H.; and Yi, L. 2026. Humanoid-GPT : Scaling Data and Structure for Zero-Shot Motion Tracking. arXiv preprint arXiv:2606.03985

  28. [28]

    L.; Myers, J

    Riemann, B. L.; Myers, J. B.; and Lephart, S. M. 2003. Comparison of the ankle, knee, hip, and trunk corrective action shown during single-leg stance on firm, foam, and multiaxial surfaces. Archives of Physical Medicine and Rehabilitation, 84(1): 90--95

  29. [29]

    R.; Schwartz, J.; and Heise, G

    Semak, M. R.; Schwartz, J.; and Heise, G. D. 2020. Examining Human Unipedal Quiet Stance: Characterizing Control through Jerk. Computational and Mathematical Methods in Medicine, 2020: 5658321

  30. [30]

    Seo, Y.; Sferrazza, C.; Chen, J.; Shi, G.; Duan, R.; and Abbeel, P. 2025. Learning Sim-to-Real Humanoid Locomotion in 15 Minutes. arXiv preprint arXiv:2512.01996

  31. [31]

    Sferrazza, C.; Huang, D.-M.; Lin, X.; Lee, Y.; and Abbeel, P. 2024. HumanoidBench : Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation. arXiv preprint arXiv:2403.10506. Robotics: Science and Systems (RSS) 2024

  32. [32]

    Sun, Z.; Huang, B.-S.; Peng, Y.; Li, X.; Ma, J.; Sun, Y.; Li, Z.; Jiang, H.; Gao, B.; Bing, Z.; Wang, X.; and Knoll, A. 2026. MOSAIC : Bridging the Sim-to-Real Gap in Generalist Humanoid Motion Tracking and Teleoperation with Rapid Residual Adaptation. arXiv preprint arXiv:2602.08594

  33. [33]

    B.; Coumans, E.; Shi, Y.; Zhang, H.; Rempe, D.; Chechik, G.; and Fidler, S

    Tessler, C.; Jiang, Y.; Peng, X. B.; Coumans, E.; Shi, Y.; Zhang, H.; Rempe, D.; Chechik, G.; and Fidler, S. 2025. ProtoMotions3 : An Open-source Framework for Humanoid Simulation and Control. https://github.com/NVLabs/ProtoMotions. GitHub repository

  34. [34]

    Tropp, H.; and Odenrick, P. 1988. Postural control in single-limb stance. Journal of Orthopaedic Research, 6(6): 833--839

  35. [35]

    Wang, T.; Xie, Z.; Yang, B.; Wang, Y.; Yuan, Z.; Fang, Y.; Feng, Y.; Wang, Y.; Chen, X.; Chen, H.; Wu, Q.; Xu, W.; Chen, L.; Li, L.; Zeng, Z.; and Xu, R. 2026 a . HumanoidArena : Benchmarking Egocentric Hierarchical Whole-body Learning. arXiv preprint arXiv:2606.17833

  36. [36]

    Wang, Y.; Zhu, S.; Zhi, P.; Li, Y.; Li, J.; Li, Y.-L.; Xiao, Y.; Wang, X.; Jia, B.; and Huang, S. 2026 b . OmniXtreme : Breaking the Generality Barrier in High-Dynamic Humanoid Control. arXiv preprint arXiv:2602.23843

  37. [37]

    Wang, Z.; Wang, J.; Yao, S.; Zhang, Y.; Ding, Z.; Yang, M.; Wang, Y.; Jiang, H.; Ma, C.; Shi, X.; and Lu, Z. 2026 c . General Humanoid Whole-Body Control via Pretraining and Fast Adaptation. arXiv preprint arXiv:2602.11929

  38. [38]

    Xie, W.; Bai, C.; Shi, J.; Yang, J.; Ge, Y.; Zhang, W.; and Li, X. 2025 a . Humanoid Whole-Body Locomotion on Narrow Terrain via Dynamic Balance and Reinforcement Learning. arXiv preprint arXiv:2502.17219

  39. [39]

    Xie, W.; Han, J.; Zheng, J.; Li, H.; Liu, X.; Shi, J.; Zhang, W.; Bai, C.; and Li, X. 2025 b . KungfuBot : Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills. arXiv preprint arXiv:2506.12851

  40. [40]

    B.; and Ames, A

    Yang, L.; Werner, B.; Ghansah, A. B.; and Ames, A. D. 2025. Bracing for Impact: Robust Humanoid Push Recovery and Locomotion with Reduced Order Models. arXiv preprint arXiv:2505.11495. IEEE-RAS International Conference on Humanoid Robots (Humanoids) 2025

  41. [41]

    P.; Cao, Z.-a.; Peng, X

    Ze, Y.; Chen, Z.; Ara \'u jo, J. P.; Cao, Z.-a.; Peng, X. B.; Wu, J.; and Liu, C. K. 2025. TWIST : Teleoperated Whole-Body Imitation System. arXiv preprint arXiv:2505.02833

  42. [42]

    Zhang, T.; Zheng, B.; Nai, R.; Hu, Y.; Wang, Y.-J.; Chen, G.; Lin, F.; Li, J.; Hong, C.; Sreenath, K.; and Gao, Y. 2025 a . HuB : Learning Extreme Humanoid Balance. arXiv preprint arXiv:2505.07294

  43. [43]

    Zhang, Z.; Guo, J.; Chen, C.; Wang, J.; Lin, C.; Lian, Y.; Xue, H.; Wang, Z.; Liu, M.; Lyu, J.; Liu, H.; Wang, H.; and Yi, L. 2025 b . Track Any Motions under Any Disturbances. arXiv preprint arXiv:2509.13833

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.