Pith. sign in

REVIEW 5 major objections 4 minor 95 references

Tired Actor: Fatigue-Informed Character Control

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Fatigue-aware torque clipping keeps simulated motion coverage while making characters move more naturally.

desk verdict A useful engineering contribution—scaling fatigue-aware control to AMASS-wide imitation—but the naturalness claim is under-supported and the main baseline comparison confounds fatigue with initialization randomization. read the letter →

arxiv 2608.03528 v1 pith:35J5KTLD submitted 2026-08-04 cs.RO

Shengyuan Zhang , Xinpeng Liu , Muchun Niu , Yulong Chen , Lizhuang Ma , Yue Gao , Cewu Lu , Yong-Lu Li This is my paper
classification cs.RO
keywords charactercontrolfatiguemodelingbehavioralenergeticsphysics-basedanimationmotionimitationreinforcementlearningThree-CompartmentControllerTiredActor
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

This paper claims that a fatigue model from exercise physiology can be layered onto a general physics-based character controller to make simulated humans move more naturally without sacrificing how many motions they can imitate. The Tired Actor wraps an existing full-body tracking policy with a Three-Compartment Controller fatigue model: as the modeled fraction of fatigued motor units grows, the maximum torque available to each joint shrinks. On AMASS, it keeps success at 99.1% on the training split and 97.9% on the test split, slightly above the unfatigued baseline, while producing shorter steps, corner-cutting, and a single-leg swing instead of foot-sliding. If this is right, fatigue is a cheap, controllable proxy for finite energy that makes physics-based characters look less robotic.

What carries the argument

The load-bearing operation is the torque-clipping identity tau_t = clip(tau_raw, -RC * tau_max, RC * tau_max), with RC = 1 - MF, where MF evolves via the Three-Compartment Controller differential equations for active, fatigued, and resting motor units. This single operation turns any torque-based controller into a fatigue-aware one. The maximal torque profile tau_max is collected by rolling out the perpetual humanoid control baseline over AMASS at beta=1, and a reweighted fatigue initialization samples the remaining capacity RC from an exponential distribution whose rate grows with training episodes, forcing the policy to learn the high-fatigue regime that uniform sampling would miss.

What would settle it

Rerun the same training with the per-joint torque ceiling defined per sequence, e.g., from the raw torque peaks of the reference motion, and check whether the success-rate cliff still appears near 90% fatigue and whether the corner-cutting and short-step behaviors persist; if the cliff or the behaviors shift or vanish, the global rollout-based tau_max is what creates the reported fatigue effects.

Watch

Extended reading notes

Core claim

The paper's central claim is that fatigue can be carried as an explicit state in a general-purpose character policy: each joint's applied torque is clipped to a fraction of a pre-measured maximal torque profile, where the fraction shrinks as the fatigued motor-unit share grows under the 3CC dynamics. The resulting Tired Actor imitates the full AMASS dataset at essentially the same success rate as the perpetual humanoid control baseline, and slightly better on several metrics, while producing visibly different fatigue-consistent movements: shorter steps, corner-cutting, and single-leg swinging instead of foot sliding. It also responds to external perturbations with more natural compensation,

Load-bearing premise

The load-bearing premise is that the maximal torque profile tau_max, measured by rolling out the baseline controller on AMASS, is a fair ceiling for human motor capacity; every fatigue effect is scaled against this self-measured profile, so if that profile had been defined differently the reported fatigue threshold and behaviors would shift.

Editorial extensions

If this is right

  • Fatigue can be added to an existing general controller as a torque-clipping layer: Tired Actor keeps 99.1% AMASS-Train and 97.9% AMASS-Test success, matching or exceeding the unfatigued baseline.
  • Fatigued control shifts movement style quantitatively: step length drops from 0.407 m to 0.385 m, paths cut corners, and the actor trades strict global pose tracking for balance and local pose fidelity.
  • Fatigue extends the training distribution: with 100 and 1,000 training sequences, the fatigue-informed actor outperforms the baseline on out-of-domain CIRCLE motions, and the gap widens as data grows.
  • Fatigue parameters control robustness: raising the fatigue rate F from 2 to 10 drops test success to 76.5%, and raising the recovery rate R to 0.2 restores it to 93.2%, while the rest-recovery multiplier r has little effect.
  • The policy compensates for external perturbations more naturally than the unfatigued baseline, avoiding falls and using more reasonable recovery gestures when it does fall.
  • Consistent whole-body fatigue is easier to compensate than localized fatigue, and asymmetric left-right fatigue reveals that the inner half of the body works harder during circle running.

Reading between the lines

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

  • Because the fatigue mechanism is a torque clip, the same layer could be dropped onto other torque-based controllers, such as diffusion-policy or residual-force controllers, to give them fatigue-aware behavior without retraining from scratch.
  • The reported result that fatigue barely bites until near 90% is a property of the self-measured torque ceiling; a lower or per-sequence ceiling would move the bite point earlier, so recalibrating tau_max per motion or per individual is a direct way to test whether a universal fatigue threshold exists.
  • The fatigue-induced off-reference motions, such as corner-cutting and shorter steps, could serve as free data augmentation for downstream tasks like pose estimation or motion prediction, producing plausible kinematic variations without manual labels.
  • For humanoid robots, the same clipping mechanism could act as an adjustable energy or safety budget rather than a fatigue model, limiting over-exertion in long-horizon tasks.
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

5 major / 4 minor

Summary. The paper proposes Tired Actor, a physics-based character controller that augments the PHC motion-imitation pipeline with the Three-Compartment Controller (3CC) fatigue model. The fatigue state modulates the maximum applicable torque (Eq. 6), and the policy is trained with PPO using a reweighted fatigue initialization, hard-negative mining, and fall-recovery finetuning. The authors claim that Tired Actor improves the naturalness of simulated human motion while maintaining motion coverage, and they report experiments on AMASS and CIRCLE datasets that show competitive success rates, better root-relative mPJPE-L in several settings, and qualitatively more plausible fatigued behaviors such as shorter steps and corner cutting.

Significance. The direction is timely and relevant: bringing biomechanical/physiological priors such as fatigue into large-scale physics-based character control is a useful step beyond purely data-driven methods, and the paper builds transparently on the 3CC model of Cheema et al. and the PHC baseline. The training pipeline is standard and reproducible, and the released code is a concrete asset. If the naturalness claim could be substantiated and the reported comparisons made robust, the work would be a solid contribution to character animation. However, in its current form, the central claim — improved naturalness — lacks quantitative support, and the controlled comparison is confounded by a simultaneous change in the training distribution. The findings are therefore not yet established at the level required for a strong publication.

major comments (5)
  1. [Sec. 1, 4.2, Table 1] The headline claim 'improving its naturalness' is not supported by any perceptual evaluation or validated naturalness metric. The quantitative results in Table 1 actually show that Tired Actor is worse than Baseline on mPJPE-G, acceleration error, and velocity error on AMASS-Train; only success rate and mPJPE-L improve. The qualitative examples in Fig. 4 and the step-length comparison (0.385 vs 0.407 m) are not substitutes for a user study or a perceptual naturalness metric. I would require either a human-evaluation study or a strong existing naturalness metric to support the central claim.
  2. [Sec. 4.2 and Eq. 7] The baseline comparison is not controlled. The 'Baseline' is obtained by 'eliminating the fatigue modeling of the Tired Actor,' which removes both the torque clip (Eq. 6) and the reweighted fatigue initialization (Eq. 7). Tired Actor is therefore trained with a randomized initial strength distribution, whereas Baseline is not (MF=MA=0, MR=100). This is a domain-randomization confound: the gains in success rate, mPJPE-L, and generalization in Tables 1, 3, and 4 could stem from the stochastic initialization alone, not from the fatigue mechanism itself. The paper needs an ablation that randomizes the initial fatigue state while keeping the torque clip disabled (or vice versa) to attribute the observed behaviors to fatigue.
  3. [Sec. 3.1, reward definition] The behavior findings are partly encoded by the reward. The reward includes r_f = 0.01 * M_F, which explicitly penalizes cumulative fatigue. Consequently, observing that Tired Actor produces reduced action amplitudes and fatigue-avoiding behaviors is not an emergent property of the fatigue dynamics; it is a direct consequence of the training objective. The paper should disentangle the contribution of the fatigue model (Eqs. 3-6) from the explicit fatigue-minimization reward, for example by ablating r_f or by reporting results where fatigue is only a constraint rather than a reward term.
  4. [Sec. 3.2 and Eq. 6] The maximal torque profile tau_max is collected by rolling out PHC on AMASS with beta=1, and every downstream fatigue effect is normalized by this profile via TL = tau_raw / tau_max. The paper's quantitative insight that 'fatigue barely matters until MF around 90%' (Sec. 4.3) is thus a property of the PHC-derived tau_max, not a physiological invariant. If tau_max had been defined per-sequence (as in Cheema et al. [11]) or from raw torque peaks, the fatigue-bite threshold and the reported behaviors would likely shift. I request a sensitivity analysis with alternative tau_max definitions and a report of the TL distribution, since this is the scaling factor for all fatigue-dependent results.
  5. [Tables 1, 3, 4 and Sec. 4.2] All experimental numbers are single-run point estimates with no variance or significance testing. Given that PPO training is stochastic, differences such as 99.1% vs 98.9% (Table 1) or 96.9% vs 98.0% (Table 3) may be within run-to-run noise. The step-length comparison (0.385 vs 0.407 m) likewise lacks error bars or a statistical test. The paper should report means and standard deviations over at least three seeds, and ideally a significance test for the behavioral comparisons, before claims of superiority are made.
minor comments (4)
  1. [Sec. 4.3, Spatial Decision-Making] The text contains '(Eq. ??)' — an unresolved reference to the next-frame imitation target. This needs to be fixed.
  2. [Sec. 3.1, Eqs. 3-5] The paragraph after Eq. (5) contains a sentence fragment: 'the fatigue and recovery rates. 1−R/F also indicates the upper bound of MF.' This appears to be a formatting/editing artifact and should be rewritten for clarity.
  3. [Table 1 and Table 4] The 'Baseline' in Table 1 is described in Sec. 4.2, but Table 4 also uses 'Baseline' without a cross-reference. Clarify that the same baseline is used, and specify whether the 'Tired Actor' rows in Table 4 correspond to the initial-MF=0 setting.
  4. [Fig. 4] The qualitative comparison in Fig. 4 would benefit from a caption describing the exact conditions (e.g., same reference motion, fatigue parameters F, R, r) and noting which differences are visible in the stills versus the supplementary video.

Circularity Check

2 steps flagged · score 4.0 of 10

Fatigued behaviors are partly designed into the reward; the 90% fatigue threshold is a property of the self-collected torque ceiling, but the central coverage claims rest on independent benchmarks.

  1. self definitional [Section 3.1, Rewards; Section 4.2, qualitative results]
    "We further introduce a fatigue reward r_t^f = 0.01M_t^F for explicit regulation on cumulative fatigue minimization."

    In the 3CC model, M_F increases through activation driven by TL = tau_raw / tau_max (Eqs. 3 and 6). Minimizing r_f is therefore, up to the 3CC dynamics, equivalent to minimizing a monotone function of the raw torque magnitude. The paper later reports as a discovered fatigued behavior that 'Tired Actor learns to cut corners ... by reducing action amplitudes.' But reduced action amplitude is the direct quantity optimized by the explicit fatigue reward, so this headline behavioral finding is an input (the objective) recycled as an output (the observed fatigue-informed behavior).

  2. fitted input called prediction [Section 4.3, Different Initial MF; Section 3.2, Maximal Torque Collection]
    "MF introduces a marginal performance drop until reaching about 90%, consistent with the fact that most movements only need less than 20% maximum torque."

    The 'fact' that most movements need less than 20% of maximum torque is not an external physiological constant; it is derived from the paper's own maximal torque collection: 'we directly roll out PHC on AMASS, and record the maximal torques for each joint according to Eq. 2 with beta^t = 1.' Since tau_max is defined as the maximum observed raw torque over PHC rollouts, TL = tau_raw / tau_max is <= 1 by construction and typically much smaller. The 90% fatigue threshold is thus a direct consequence of how tau_max was defined, not an independent prediction about human fatigue. A different tau_max definition (per-sequence peaks, raw torque maxima, or physiological data) would shift the threshold.

full rationale

The paper's central quantitative claims—motion coverage on AMASS-Train/Test and generalization to CIRCLE—are evaluated against external benchmarks and are not circular: the success rates, mPJPE values, and generalization comparisons do not reduce to the fatigue model's own parameters. However, two load-bearing interpretative moves do reduce to the paper's construction. First, the fatigue reward r_f explicitly penalizes cumulative fatigue M_F, and M_F is driven by torque magnitude, so the reported 'reduced action amplitude' behavior is partly an optimized objective rather than an emergent discovery. Second, the explanation that fatigue only matters near MF≈90% relies on the self-collected tau_max profile; because tau_max is the maximum of PHC's own raw torques, the claim that 'most movements only need less than 20% maximum torque' is an artifact of that measurement, not an external validation. The baseline comparison is also confounded by removing both the torque clipping and the reweighted fatigue initialization, which is a correctness risk but not itself a circularity. Weighing these, the central coverage result remains independent, so the circularity is partial rather than total.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central numbers depend on four classes of hand-set or self-measured quantities: fatigue constants (F, R, r), the tau_max torque profile measured from PHC rollouts, the reweighted initialization schedule, and reward weights including the explicit fatigue-minimizing reward. The fatigue model itself and the naturalness definition are inherited domain assumptions. No new physical entities are introduced.

free parameters (4)
  • Fatigue model constants (F, R, r) = F=2, R=0.05, r=1
    Set in Section 4.2 and used for all main experiments; Section 4.3 shows F=10 drops success rate to 76.5%, so results are sensitive to these hand-chosen values taken from prior 3CC work.
  • Maximal torque profile tau_max = Rolled out from PHC on AMASS with beta=1
    Defined in Section 3.2. Sets the scale at which fatigue bites; Table 3's finding that MF up to 90% barely matters follows from this self-measured profile, not from physiology.
  • Reweighted initialization schedule lambda_ep = 1 increasing to 3 over 20K episodes, then fixed at 3
    Section 3.2 and 4.1; hand-chosen schedule controlling how aggressively high-fatigue states are sampled during training.
  • Reward weights (task weights, fatigue reward, power reward) = omega1-4 = 0.5/0.1/0.3/0.1; r_f coefficient 0.01; power coefficient 0.0005
    Section 3.1 and 4.1; hand-set. The fatigue reward r_f = 0.01 * M_F explicitly trains the policy to minimize cumulative fatigue, directly shaping the reported fatigue behaviors.
assumptions (5)
  • domain assumption The 3CC motor-unit compartment model (Eqs. 3-6) validly describes whole-body human fatigue via torque clipping RC = 1 - MF.
    Adopted from [11, 40, 77] in Section 3.1; validity for a 24-joint whole-body SMPL character is assumed, not validated against human fatigue data.
  • domain assumption Fatigue is a serviceable proxy for the finite energy limit that behavioral energetics says shapes movements.
    Motivation from Sections 1 and 2.2; the link from energy optimality to this torque-clipping fatigue model is asserted, not derived.
  • ad hoc to paper tau_max from PHC rollouts with beta=1 represents general human motor capacity.
    Section 3.2 'Maximal Torque Collection'; this is the paper's own operationalization and determines when fatigue constraints become active.
  • domain assumption The AMP discriminator trained on reference motion captures 'naturalness'.
    Section 3.1, r_amp follows AMP [61]; the naturalness claim relies on this learned discriminator plus qualitative visual judgment.
  • standard math Standard RL machinery (PPO, IsaacGym rigid body dynamics, PD torque control of the SMPL skeleton) is reliable.
    Section 3.1; standard in the field, cited [50, 54, 59].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tired Actor: Fatigue-Informed Character Control." pith.science (2026). https://pith.science/paper/35J5KTLD

@misc{pith2026260803528,
  author       = {Pith},
  title        = {Pith review of: Tired Actor: Fatigue-Informed Character Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/35J5KTLD}},
  note         = {Machine review of arXiv:2608.03528}
}
read the original abstract

Replicating human behavior with physics simulation has been a long-expected goal in character animation. Existing efforts have achieved impressive performance in imitating a wide span of general motions. However, most existing efforts could still suffer from unnatural movements due to the lack of biomechanical and physiological priors. Given this, we project our sights to advances in behavioral energetics, which demonstrate how energy use shapes human movements. In contrast, current character controllers typically assume the character is equipped with infinite energy over time. Inspired by these, we propose to adopt fatigue as a proxy of the finite energy limit, inject it into general character animation, and thoroughly investigate how fatigue introduces new characteristics to physics-based character control. Leveraging the Three-Compartment Controller (3CC) model, we managed to obtain a policy for general motion imitation under different fatigue statuses. Furthermore, extensive analyses are conducted to demonstrate how fatigue could influence the naturalness, scalability, and robustness of character animation. Our code will be made public.

Figures

Figures reproduced from arXiv: 2608.03528 by the authors.

Figure 1
Figure 1. Current controllers could still be unnatural, like replicating treadmill paces on the ground. Inspired by behavioral [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Reweighted fatigue initialization. limits, we could now train our fatigue-informed controller with PPO. The criterion network shares the same structure as the policy. At the start of each episode, the initial fatigue states MF , MA, MR, and fatigue parameters F, R, r are randomly sampled. However, this introduces new problems. In [11], MF , MA, MR was uniformly sampled. However, our general￾purpose torque limits cou… view at source ↗
Figure 2
Figure 2. The general pipeline of Tired Actor. We first collect [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Motion imitation with different initial MF s. Init 𝑀𝐹 = 0 Arm Init 𝑀𝐹 = 99 Other parts’ Init 𝑀𝐹 = 0 Init 𝑀𝐹 = 99 Leg Init 𝑀𝐹 = 99 Other parts’ Init 𝑀𝐹 = 0 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Fatigue-induced cross-part coordination. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Fatigue-induced asymmetry. Fatigue on inner-/outer-half body results in diverse behaviors. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Fatigue-induced spatial decision making. Trajectories and curvature distributions under different turning angles are [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Fatigue-induced compensation under different [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 10
Figure 10. Figure 10: External perturbation compensation with different perturbation parts. Perturbations applied to distal segments are [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Tired Actors with and without fall-recovery fine-tuning imitate a forward-walking motion with initial [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

95 extracted references · 55 canonical work pages

  1. [1]

    McNeill Alexander

    R. McNeill Alexander. 1996.Optima for Animals: Re- vised Edition. Princeton University Press. http: //www.jstor.org/stable/j.ctv173f0gj

    1996

  2. [2]

    Karen Liu

    Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander Clegg, and C. Karen Liu. 2023. CIRCLE: Capture In Rich Con- textual Environments . In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, Los Alamitos, CA, USA, 21211– 21221. doi:10.1109/CVPR52729.2023.02032

    arXiv 2023

  3. [3]

    Christopher J Arellano and Rodger Kram. 2011. The effects of step width and arm swing on energetic cost and lateral balance during running.Journal of biomechanics 44, 7 (2011), 1291–1295

    2011

  4. [4]

    Edgar Atzler and Robert Herbst. 1927. Arbeitsphysiolo- gische studien: Iii. teil.Pfl ¨uger’s Archiv f¨ur die gesamte Physiologie des Menschen und der Tiere215 (1927), 291– 328

    1927

  5. [5]

    John EA Bertram and Andy Ruina. 2001. Multiple walking speed–frequency relations are predicted by con- strained optimization.Journal of theoretical Biology209, 4 (2001), 445–453

    2001

  6. [6]

    Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. 2024. Physically Plausi- ble Full-Body Hand-Object Interaction Synthesis . In 2024 International Conference on 3D Vision (3DV). IEEE Computer Society, Los Alamitos, CA, USA, 464–473. doi:10.1109/3DV62453.2024.00109

    arXiv 2024

  7. [7]

    J Brisswalter, C Hausswirth, D Smith, F Vercruyssen, and JM Vallier. 2000. Energetically optimal cadence vs. freely-chosen cadence during cycling: effect of exercise duration.International journal of sports medicine21, 01 (2000), 60–64

    2000

  8. [8]

    GA Brooks, TD Fahey, TP White, and KM Baldwin. 2005. Basics of metabolism.Exercise Physiology Human Bioen- ergetics and its Applications4 (2005), 43–58

    2005

Show all 95 references
  1. [9]

    Geoffrey L Brown, Nidhi Seethapathi, and Manoj Srini- vasan. 2021. A unified energy-optimality criterion pre- dicts human navigation paths and speeds.Proceedings of the National Academy of Sciences118, 29 (2021), e2020327118

    2021

  2. [10]

    Frey-Law, Kourosh Naderi, Jaakko Lehtinen, Philipp Slusallek, and Perttu H¨am¨al¨ainen

    Noshaba Cheema, Laura A. Frey-Law, Kourosh Naderi, Jaakko Lehtinen, Philipp Slusallek, and Perttu H¨am¨al¨ainen. 2020. Predicting Mid-Air Interaction Move- ments and Fatigue Using Deep Reinforcement Learning. InProceedings of the 2020 CHI Conference on Human Factors in Computi...

    2020

  3. [12]

    Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, and Xiaolong Wang. 2024. Expressive Whole- Body Control for Humanoid Robots. InProceedings of Robotics: Science and Systems. Delft, Netherlands. doi:10.15607/RSS.2024.XX.107

    2024 doi

  4. [13]

    Pei Hua Cher, Ian Stewart, and Charles Worringham. 2015. Minimum cost of transport in human running is not ubiq- uitous.Medicine and Science in Sports and Exercise47, 2 (2015), 307–314

    2015

  5. [14]

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. 0. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research0, 0 (0), 02783649241273668. arXiv:https://doi.or...

    doi

  6. [15]

    Katherine AJ Daniels and JF Burn. 2023. Human locomo- tion over obstacles reveals real-time prediction of energy expenditure for optimized decision-making.Proceedings of the Royal Society B290, 2000 (2023), 20230200

    2023

  7. [16]

    Martin De Lasa, Igor Mordatch, and Aaron Hertzmann

  8. [17]

    Frederick J Diedrich and William H Warren Jr. 1995. Why change gaits? Dynamics of the walk-run transition.Jour- nal of Experimental Psychology: Human Perception and Performance21, 1 (1995), 183

    1995

  9. [18]

    Jun Ding, Anthony S Wexler, and Stuart A Binder- Macleod. 2000. A predictive model of fatigue in human skeletal muscles.Journal of applied physiology89, 4 (2000), 1322–1332

    2000

  10. [19]

    J Maxwell Donelan, David W Shipman, Rodger Kram, and Arthur D Kuo. 2004. Mechanical and metabolic re- quirements for active lateral stabilization in human walk- ing.Journal of biomechanics37, 6 (2004), 827–835

    2004

  11. [20]

    Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. 2023. C·ASE: Learning Conditional Adversarial Skill Embeddings for Physics-based Charac- ters. InSIGGRAPH Asia 2023 Conference Papers(Syd- ney, NSW, Australia)(SA ’23). Association for Comput- ing Machinery,...

    2023

  12. [21]

    Herbert Elftman. 1966. Biomechanics of muscle: with particular application to studies of gait.JBJS48, 2 (1966), 363–377

    1966

  13. [22]

    Yusen Feng, Xiyan Xu, and Libin Liu. 2023. MuscleV AE: Model-Based Controllers of Muscle-Actuated Characters. InSIGGRAPH Asia 2023 Conference Papers(Sydney, NSW, Australia)(SA ’23). Association for Computing Machinery, New York, NY , USA, Article 3, 11 pages. doi:10.1145/36105...

    2023

  14. [23]

    Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. 2024. Coohoi: Learning cooperative human-object interaction with manipulated object dynam- ics.Advances in Neural Information Processing Systems 37 (2024),...

    2024

  15. [24]

    Thomas Geijtenbeek, Michiel Van De Panne, and A Frank Van Der Stappen. 2013. Flexible muscle-based locomo- tion for bipedal creatures.ACM Transactions on Graphics (TOG)32, 6 (2013), 1–11

    2013

  16. [25]

    Yohanan Giat, Joseph Mizrahi, and Mark Levy. 1996. A Model of Fatigue and Recovery in Paraplegic’s Quadri- ceps Muscle Subjected to Intermittent FES.Journal of Biomechanical Engineering118, 3 (08 1996), 357–366. arXiv:https://asmedigitalcollection.asme.org/biomechanical/articl...

    1996 doi

  17. [26]

    Yohanan Giat, Joseph Mizrahi, Mark Levy, et al. 1993. A musculotendon model of the fatigue profiles of paralyzed quadriceps muscle under FES.IEEE transactions on biomedical engineering40, 7 (1993), 664–674

    1993

  18. [27]

    Jinger S Gottschall and Rodger Kram. 2003. Energy cost and muscular activity required for propulsion during walking.Journal of Applied Physiology94, 5 (2003), 1766–1772

    2003

  19. [29]

    Tairan He, Jiawei Gao, Wenli Xiao, Yuanhang Zhang, Zi Wang, Jiashun Wang, Zhengyi Luo, Guanqi He, Nikhil Sobanbab, Chaoyi Pan, Zeji Yi, Guannan Qu, Kris Kitani, Jessica Hodgins, Linxi ”Jim” Fan, Yuke Zhu, Changliu Liu, and Guanya Shi. 2025. ASAP: Aligning Simulation and Real-W...

    2025 arXiv

  20. [30]

    Tairan He, Wenli Xiao, Toru Lin, Zhengyi Luo, Zhenjia Xu, Zhenyu Jiang, Jan Kautz, Changliu Liu, Guanya Shi, Xiaolong Wang, Linxi Fan, and Yuke Zhu. 2025. HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots. arXiv:2410.21229 [cs.RO] https://arxiv. org/abs/2410.21229

    2025 arXiv

  21. [31]

    Jonathan Ho and Stefano Ermon. 2016. Gen- erative Adversarial Imitation Learning. InAd- vances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett (Eds.), V ol. 29. Curran Associates, Inc. https://proceedings.neurips. cc/paper_fi...

    2016

  22. [32]

    Kenneth G Holt, Joseph Hamill, and Robert O Andres

  23. [33]

    ALAN Hreljac. 1993. Preferred and energetically optimal gait transition speeds in human locomotion.Medicine and science in sports and exercise25, 10 (1993), 1158–1162

    1993

  24. [34]

    Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. 2022. PADL: Language-Directed Physics-Based Character Control. InSIGGRAPH Asia 2022 Conference Papers(Daegu, Republic of Korea)(SA ’22). Association for Computing Machinery, New York, NY , USA, Article 19, 9 pages. ...

    2022

  25. [36]

    Kider, Kaitlin Pollock, and Alla Safonova

    Joseph T. Kider, Kaitlin Pollock, and Alla Safonova. 2011. A data-driven appearance model for human fatigue. In Proceedings of the 2011 ACM SIGGRAPH/Eurographics Symposium on Computer Animation(Vancouver, British Columbia, Canada)(SCA ’11). Association for Comput- ing Machiner...

    2011

  26. [37]

    Taku Komura, Yoshihisa Shinagawa, and Tosiyasu L Ku- nii. 2000. Creating and retargetting motion by the mus- culoskeletal human body model.The visual computer16 (2000), 254–270

    2000

  27. [38]

    Niranjan Kumar, Irfan Essa, and Sehoon Ha

    K. Niranjan Kumar, Irfan Essa, and Sehoon Ha. 2023. Words into Action: Learning Diverse Humanoid Robot Behaviors using Language Guided Iterative Motion Re- finement. arXiv:2310.06226 [cs.RO]

    2023 arXiv

  28. [39]

    Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. 2023. QuestEnvSim: Environment-Aware Simulated Motion Tracking from 9 Sparse Sensors. InACM SIGGRAPH 2023 Conference Proceedings(Los Angeles, CA, USA)(SIGGRAPH ’23). Association for Computing Machiner...

    2023 doi

  29. [40]

    Jing Z Liu, Robert W Brown, and Guang H Yue. 2002. A dynamical model of muscle activation, fatigue, and recovery.Biophysical journal82, 5 (2002), 2344–2359

    2002

  30. [41]

    Libin Liu and Jessica Hodgins. 2017. Learning to sched- ule control fragments for physics-based characters using deep q-learning.ACM Transactions on Graphics (TOG) 36, 3 (2017), 1–14

    2017

  31. [42]

    Libin Liu and Jessica Hodgins. 2018. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning.Acm transactions on graphics (tog)37, 4 (2018), 1–14

    2018

  32. [43]

    Libin Liu, KangKang Yin, Michiel van de Panne, Tianjia Shao, and Weiwei Xu. 2010. Sampling-based contact- rich motion control.ACM Trans. Graph.29, 4, Article 128 (July 2010), 10 pages. doi: 10.1145/1778765. 1778865

    2010 doi

  33. [44]

    Leroy L Long III and Manoj Srinivasan. 2013. Walking, running, and resting under time, distance, and average speed constraints: optimality of walk–run–rest mixtures. Journal of The Royal Society Interface10, 81 (2013), 20120980

    2013

  34. [45]

    John M Looft, Nicole Herkert, and Laura Frey-Law

  35. [46]

    Matthew Loper, Naureen Mahmood, Javier Romero, Ger- ard Pons-Moll, and Michael J. Black. 2015. SMPL: a skinned multi-person linear model, In ACM Transac- tions on Graphics.ACM Trans. Graph.34, Article 248, 16 pages. doi:10.1145/2816795.2818013

    2015

  36. [47]

    Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. 2024. Omnigrasp: Grasping diverse objects with simulated humanoids.Ad- vances in Neural Information Processing Systems37 (2024), 2161–2184

    2024

  37. [48]

    Zhengyi Luo, Jinkun Cao, Rawal Khirodkar, Alexan- der Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. 2024. Real-Time Simulated Avatar from Head- Mounted Sensors . In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, Los Alamit...

    2024

  38. [49]

    Kitani, and Weipeng Xu

    Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Win- kler, Jing Huang, Kris M. Kitani, and Weipeng Xu

  39. [50]

    Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. 2023. Perpetual Humanoid Control for Real-time Simulated Avatars. In2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, Los Alamitos, CA, USA, 10861–10870. doi: 10.1109/ ICCV51070.2023.01000

    2023

  40. [51]

    Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani

  41. [52]

    Zhengyi Luo, Shun Iwase, Ye Yuan, and Kris Kitani

  42. [53]

    Mahmood, N

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. 2019. AMASS: Archive of Motion Cap- ture As Surface Shapes. In2019 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). IEEE Computer Society, Los Alamitos, CA, USA, 5441–5450. doi:10.1109/ICCV.2019.00554

    2019

  43. [54]

    Viktor Makoviychuk, Lukasz Wawrzyniak, Yun- rong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. 2021. Isaac Gym: High Performance GPU Based Physics Simulation For Robot Learning. InProceedings of the ...

    2021

  44. [55]

    McAllister, Anthony Chen, and Jessica C

    Megan J. McAllister, Anthony Chen, and Jessica C. Selinger. 2025. Behavioural energetics in human locomo- tion: how energy use influences how we move.Journal of Experimental Biology228 (02 2025), JEB248125. arXiv:https://journals.biologists.com/jeb/article- pdf/228/Suppl 1/JEB...

    2025 doi

  45. [56]

    Jacques Mercier, Daniel Le Gallais, Marc Durand, Chris- tine Goudal, Jean Paul Micallef, and Christian Pr ´efaut

  46. [57]

    Josh Merel, Yuval Tassa, Dhruva TB, Sriram Srinivasan, Jay Lemmon, Ziyu Wang, Greg Wayne, and Nicolas 10 Heess. 2017. Learning human behaviors from motion cap- ture by adversarial imitation. arXiv:1707.02201 [cs.RO] https://arxiv.org/abs/1707.02201

    2017 arXiv

  47. [58]

    NH Molen, RH Rozendal, and W Boon. 1972. Graphic representation of the relationship between oxygen- consumption and characteristics of normal gait of the human male.Proceedings of the Koninklijke Nederlandse Akademie Van Wetenschappen. Series C. Biological and Medical Sciences...

    1972

  48. [59]

    Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. 2018. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG)37, 4 (2018), 1– 14

    2018

  49. [60]

    Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. 2022. ASE: large-scale reusable adver- sarial skill embeddings for physically simulated charac- ters.ACM Trans. Graph.41, 4, Article 94 (July 2022), 17 pages. doi:10.1145/3528223.3530110

    2022

  50. [61]

    Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. 2021. Amp: Adversarial motion priors for stylized physics-based character control.ACM Transactions on Graphics (ToG)40, 4 (2021), 1–20

    2021

  51. [62]

    Henry J Ralston. 1958. Energy-speed relation and optimal speed during level walking.Internationale Zeitschrift f ¨ur Angewandte Physiologie Einschliesslich Arbeitsphysiolo- gie17, 4 (1958), 277–283

    1958

  52. [63]

    Joseph K Rathkey and Cara M Wall-Scheffler. 2017. Peo- ple choose to run at their optimal speed.American Journal of Physical Anthropology163, 1 (2017), 85–93

    2017

  53. [64]

    Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. 2023. InsActor: instruction-driven physics-based characters. InProceedings of the 37th In- ternational Conference on Neural Information Process- ing Systems(New Orleans, LA, USA)(NIPS ’23). Cur- ran Asso...

    2023

  54. [65]

    Cristian D Riveros-Matthey, Timothy J Carroll, Mark J Connick, and Glen A Lichtwark. 2025. An in-silico inves- tigation of the effect of changing cycling crank power and cadence on muscle energetics and active muscle volume. Journal of Biomechanics180 (2025), 112530

    2025

  55. [66]

    St´ephane Ross, Geoffrey Gordon, and Drew Bagnell

  56. [67]

    Karen L Steudel-Numbers and Cara M Wall-Scheffler

  57. [68]

    Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. 2024. Maskedmimic: Unified physics- based character control through masked motion inpainting. ACM Transactions on Graphics (TOG)43, 6 (2024), 1– 21

    2024

  58. [69]

    Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. 2023. CALM: Condi- tional Adversarial Latent Models for Directable Virtual Characters. InACM SIGGRAPH 2023 Conference Papers (Los Angeles, CA, USA)(SIGGRAPH ’23). Association for Computing Mach...

    2023

  59. [70]

    Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit Haim Bermano, and Michiel van de Panne. 2025. CLoSD: Closing the Loop between Simulation and Diffusion for multi-task char- acter control. InThe Thirteenth International Confer- ence on Learnin...

    2025

  60. [71]

    Darryl G Thelen. 2003. Adjustment of muscle mechanics model parameters to simulate dynamic contractions in older adults.J. Biomech. Eng.125, 1 (2003), 70–77

    2003

  61. [72]

    Takara Everest Truong, Michael Piseno, Zhaoming Xie, and Karen Liu. 2024. PDP: Physics-Based Character Animation via Diffusion Policy. InSIGGRAPH Asia 2024 Conference Papers(Tokyo, Japan)(SA ’24). Association for Computing Machinery, New York, NY , USA, Article 86, 10 pages. d...

    2024

  62. [73]

    Jiashun Wang, Jessica Hodgins, and Jungdam Won. 2024. Strategy and Skill Learning for Physics-based Table Ten- nis Animation. InACM SIGGRAPH 2024 Conference Pa- pers(Denver, CO, USA)(SIGGRAPH ’24). Association for Computing Machinery, New York, NY , USA, Article 122, 11 pages....

    2024

  63. [74]

    Michael A Willcockson and Cara M Wall-Scheffler. 2012. Reconsidering the effects of respiratory constraints on the optimal running speed.Medicine & Science in Sports & Exercise44, 7 (2012), 1344–1350

    2012

  64. [75]

    Alexander Winkler, Jungdam Won, and Yuting Ye. 2022. QuestSim: Human Motion Tracking from Sparse Sensors with Simulated Avatars. InSIGGRAPH Asia 2022 Confer- ence Papers(Daegu, Republic of Korea)(SA ’22). Associ- ation for Computing Machinery, New York, NY , USA, Ar- ticle 2, ...

    2022

  65. [76]

    Jack M Winters. 1995. An improved muscle-reflex actu- ator for use in large-scale neuromusculoskeletal models. Annals of biomedical engineering23 (1995), 359–374

    1995

  66. [77]

    Ting Xia and Laura A Frey Law. 2008. A theoretical ap- proach for modeling peripheral muscle fatigue and recov- ery.Journal of biomechanics41, 14 (2008), 3046–3052. 11

    2008

  67. [78]

    Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wen- wei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang

  68. [79]

    Michael Xu, Yi Shi, KangKang Yin, and Xue Bin Peng. 2025. PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers. arXiv:2505.04002 [cs.GR]

    2025 arXiv

  69. [80]

    Pei Xu, Xiumin Shang, Victor Zordan, and Ioannis Karamouzas. 2023. Composite motion learning with task control.ACM Transactions on Graphics (TOG)42, 4 (2023), 1–16

    2023

  70. [81]

    Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. 2025. InterMimic: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions. arXiv:2502.20390 [cs.CV]

    2025

  71. [82]

    Xinyu Xu, Yizheng Zhang, Yong-Lu Li, Lei Han, and Cewu Lu. 2024. HumanVLA: Towards Vision- Language Directed Object Rearrangement by Physical Humanoid. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Bel- grave, A. Fan, U. Paquet, J. Tomczak, a...

    2024

  72. [83]

    Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. 2022. Controlvae: Model-based learning of gen- erative controllers for physics-based characters.ACM Transactions on Graphics (TOG)41, 6 (2022), 1–16

    2022

  73. [84]

    Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. 2024. Moconvq: Unified physics-based motion control via scalable discrete repre- sentations.ACM Transactions on Graphics (TOG)43, 4 (2024), 1–21

    2024

  74. [85]

    Runyi Yu, Yinhuai Wang, Qihan Zhao, Hok Wai Tsui, Jingbo Wang, Ping Tan, and Qifeng Chen. 2025. SkillMimic-V2: Learning Robust and Generalizable In- teraction Skills from Sparse and Noisy Demonstrations. arXiv:2505.02094 [cs.LG] https://arxiv.org/ abs/2505.02094

    2025 arXiv

  75. [86]

    Wenhao Yu, Greg Turk, and C Karen Liu. 2018. Learning symmetric and low-energy locomotion.ACM Transac- tions on Graphics (TOG)37, 4 (2018), 1–12

    2018

  76. [87]

    Ye Yuan and Kris Kitani. 2020. Residual force control for agile human behavior imitation and extended motion synthesis.Advances in Neural Information Processing Systems33 (2020), 21763–21774

    2020

  77. [88]

    InThe Twelfth International Confer- ence on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    Unified Human-Scene Interaction via Prompted Chain-of-Contacts. InThe Twelfth International Confer- ence on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net. https: //openreview.net/forum?id=1vCnDyQkjg

    2024

  78. [98]

    Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han. 2023. Neural Categorical Priors for Physics-Based Character Control.ACM Trans. Graph.42, 6, Article 178 (dec 2023), 16 pages. doi:10.1145/3618397 A Additional Qualitative Results The following figures provide additional qualitat...

    2023 doi

  79. [1991]

    Predicting the minimal energy costs of human walking.Medicine and science in sports and exercise23, 4 (1991), 491–498

    1991

  80. [1994]

    Energy expenditure and cardiorespiratory responses at the transition between walking and running.European journal of applied physiology and occupational physiol- ogy69, 6 (1994), 525–529

    1994

  81. [2009]

    Optimal running speed and the evolution of hominin hunting strategies.Journal of Human Evolution56, 4 (2009), 355–360

    2009

  82. [2010]

    Feature-based locomotion controllers.ACM trans- actions on graphics (TOG)29, 4 (2010), 1–10

    2010

  83. [2011]

    InProceedings of the fourteenth international conference on artificial intel- ligence and statistics

    A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the fourteenth international conference on artificial intel- ligence and statistics. JMLR Workshop and Conference Proceedings, 627–635

  84. [2018]

    Modification of a three-compartment muscle fa- tigue model to predict peak torque decline during inter- mittent tasks.Journal of biomechanics77 (2018), 16–25

    2018

  85. [2021]

    Dynamics-regulated kinematic policy for egocen- tric pose estimation.Advances in Neural Information Processing Systems34 (2021), 25019–25032

    2021

  86. [2022]

    Advances in Neural Information Processing Systems35 (2022), 6815–6828

    Embodied scene-aware human pose estimation. Advances in Neural Information Processing Systems35 (2022), 6815–6828

    2022

  87. [2024]

    InThe Twelfth International Con- ference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    Universal Humanoid Motion Representations for Physics-Based Control. InThe Twelfth International Con- ference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net. https: //openreview.net/forum?id=OrOd8PxOO2

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.