Pith. sign in

REVIEW 5 major objections 5 minor 42 references

Motion imitation learning that matches walking kinematics does not guarantee biomechanically plausible joint kinetics; adding ground reaction force and center of pressure rewards brings simulated joint moments much closer to inverse dynamic

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

2026-08-02 18:17 UTC pith:SHNBPMW3

load-bearing objection Useful, well-controlled ablation of kinetic rewards for gait IL, but the 'three speeds' claim in the abstract doesn't match the single-speed experiment, and the coarse contact model makes the strong negative conclusion conditional. the 5 major comments →

arxiv 2603.12408 v2 pith:SHNBPMW3 submitted 2026-03-12 cs.RO cs.LG

Human Motion Data Alone Does Not Guarantee Plausible Gait Biomechanics

Xinyi Liu , Jangwhan Ahn , Edgar Lobaton , Jennie Si , He Huang This is my paper
classification cs.RO cs.LG
keywords gait biomechanicsimitation learningreinforcement learningjoint momentsground reaction forcecenter of pressureforward dynamicsinverse dynamics
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 whether a reinforcement-learning agent trained only to imitate human joint motion can also infer realistic internal forces like joint moments. In walking simulations, the authors find that motion-only imitation (MOIL) reproduces kinematics well but produces substantially inaccurate ground reaction forces, center of pressure locations, and joint moments. Adding explicit kinetic reward terms for ground reaction force and center of pressure (a kinetics-aware IL, or KAIL) significantly improves kinetic plausibility while keeping kinematic tracking essentially unchanged. The authors conclude that motion-only imitation is under-constrained for biomechanical analyses and that force-related measurements remain necessary when imitation learning is used in human biomechanics or wearable robotics.

Core claim

The central discovery is that kinematic fidelity and kinetic plausibility decouple in imitation-learned gait policies. When trained solely on motion rewards, a policy can reproduce hip, knee, and ankle angle trajectories within acceptable error limits while generating unphysiological ground reaction forces, large center-of-pressure errors, and joint moments that deviate markedly from inverse-dynamics references. Introducing rewards that penalize mismatch in sagittal-plane ground reaction forces and anterior-posterior center-of-pressure position systematically reduces these kinetic errors: the full KAIL condition cuts CoP RMSE by 44.3% relative to MOIL and raises joint-moment correlations wit

What carries the argument

The key mechanism is the reward decomposition in the imitation-learning objective. The total reward is Rk + Rdyn, where Rk contains kinematic imitation terms (joint pose, end-effector positions, root orientation and velocity) and Rdyn = w_grf R_grf + w_cop R_cop adds exponential penalties for mismatches in vertical/anterior-posterior ground reaction forces and anterior-posterior center-of-pressure position for each foot. These kinetic rewards are what push the simulated foot-ground interaction toward measured quantities, thereby regularizing the joint moments produced by the impedance controller. The ablation study—running the same pre-trained policy with Rk only, Rk+R_grf, Rk+R_cop, or all

Load-bearing premise

The conclusion rests on the assumption that the simplified foot-ground contact model used in the simulation is accurate enough that the observed kinetic differences between MOIL and KAIL are caused by the rewards, not by the contact physics itself.

What would settle it

Replace the box-foot contact model with a more realistic deformable or calibrated contact model, retrain both MOIL and KAIL, and check whether the gap in joint-moment RMSE between them persists; if MOIL's kinetic errors become comparable to KAIL's under a better contact model, the claim that motion imitation alone is 'fundamentally under-constrained' would be weakened.

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

If this is right

  • Motion-only imitation pipelines used to build 'digital twin' gait models will likely misestimate joint moments and muscle demands, which could bias clinical diagnoses or rehabilitation planning.
  • Wearable exoskeleton or prosthesis controllers optimized using joint moments from motion-only imitation will inherit these kinetic errors, potentially producing incorrect assistive torques.
  • Kinetic reward shaping (GRF and CoP) should become a standard component in biomechanically oriented imitation learning, since it improves internal and external kinetic realism without sacrificing kinematic fidelity.
  • Imitation learning, when augmented with kinetic constraints, can complement inverse dynamics by providing stable forward simulations that generate continuous joint moments for cause-effect and human-robot interaction analysis.

Where Pith is reading between the lines

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

  • The superior CoP-only improvement (43.2% CoP RMSE reduction) versus GRF-only (14.5%) suggests that the point of application of the ground reaction force is the more informative kinetic signal for joint-moment accuracy; a testable extension would be to weight CoP more heavily or add mediolateral CoP terms.
  • If the foot-ground contact model were refined (e.g., deformable foot, time-varying contact parameters), the MOIL-vs-KAIL gap might shrink or grow; co-optimizing contact parameters with policy learning, as the authors suggest, is a concrete next step to test whether 'under-constrained' is fundamental or an artifact of contact simplification.
  • The finding likely generalizes to other simulated locomotion tasks (running, stair negotiation) and to musculoskeletal models with muscle actuators, where the underlying force-generation mechanism differs but the need for kinetic ground truth remains.
  • The relatively large knee-moment errors in all conditions, consistent with known sensitivity of knee moments to CoP uncertainty, imply that clinical conclusions about knee loading from imitation learning should be especially cautious until kinetic rewards or better contact models are in place.

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

5 major / 5 minor

Summary. The paper addresses whether motion imitation learning (IL) can recover biomechanically plausible joint kinetics without explicit kinetic information. It compares a motion-only IL baseline (MOIL) with a kinetics-aware IL (KAIL) that adds ground reaction force (GRF) and center of pressure (CoP) rewards, in a MuJoCo/PPO forward-dynamics simulation of treadmill walking. The authors report that both methods achieve comparable kinematic tracking, but KAIL reduces GRF/CoP errors and brings hip, knee, and ankle joint moments closer to Visual3D inverse dynamics estimates. They conclude that motion-only IL is fundamentally limited for estimating human-like joint kinetics and that kinetic reward shaping is necessary. The study includes an ablation of the GRF and CoP terms and uses inverse dynamics as an external reference for joint moments, which is a strength.

Significance. If the central claim is established, this is a valuable cautionary result for the growing use of motion-only RL imitation in biomechanics and wearable robotics. The paper's strengths are the ablation design, the explicit comparison of joint moments against inverse dynamics (which is not part of the reward), and the use of shared anthropometric parameters between the inverse-dynamics and forward-dynamics models. The main limitation is that the demonstration rests on one participant, one speed, and one simplified contact model; the title and abstract overstate the generality. The claims are important for the community, provided the confounds below are addressed.

major comments (5)
  1. [§IV-A vs. Abstract/Title] The abstract states that experiments were conducted at three speeds (0.9, 1.2, and 1.5 m/s), but the Methods section reports a single 2-minute trial at 1.2 m/s. No results for 0.9 or 1.5 m/s appear anywhere in the paper. Section VI concedes that the study evaluated 'only a single walking condition.' The central claim of a fundamental limitation of MOIL should be reframed as a single-speed, single-participant simulation observation, or additional speeds must be provided.
  2. [§III-B, Eq. (5)-(6); §V-C] The GRF and CoP rewards directly minimize distance to the exact expert GRF/CoP signals that are later used as evaluation metrics in Fig. 6 and Table I. The improvement in these external-kinetics metrics is therefore a fitting effect, not evidence of generalization. The joint-moment comparison is less circular because inverse dynamics is not part of the reward, but the inverse-dynamics reference is itself computed from the measured GRF and CoP. With kinematics closely matched, a policy that reproduces the experimental GRF/CoP will almost mechanically produce joint moments closer to the inverse-dynamics reference. The paper should state this and provide a stronger test, e.g., evaluation on held-out gait cycles or under perturbed contact conditions.
  3. [§II and §VI] The foot-ground contact model is a box-shaped foot with four discrete contact points, contact margin epsilon=0.05, time constant tau=0.02, and damping ratio c=1. Section VI explicitly calls this model 'an oversimplification, a known bottleneck in sim-to-real transfer.' The MOIL-vs-KAIL gap could therefore be caused by the contact model's inability to transduce joint torques into biological GRF patterns, with the kinetic reward compensating for contact-model error rather than adding genuinely missing information. A concrete test would be to vary contact parameters or use a more realistic foot model and check whether MOIL's kinetic errors persist. Without such a test, the phrase 'fundamental limitation' is not supported.
  4. [§III-A, Eq. (1)-(2)] The policy outputs a residual wrench ξ_t applied to the floating base. This is a non-physical external wrench that can help achieve kinematic tracking without realistic joint torques. The R_vf term only penalizes, not prohibits, large residual wrenches. The paper does not report residual-wrench magnitudes for MOIL vs. KAIL. If MOIL relies on large residual forces, the poor joint moments reflect a property of the chosen residual-force-control architecture, not of motion data alone. Please report ||ξ|| over time and, ideally, include an ablation with zero or heavily penalized residual wrench to support the causal attribution.
  5. [§III-B, §IV-B, §V] The reward weights and kernel constants (w_p, w_ee, w_rp, w_rv, w_vf, w_grf, w_cop, k_grf, k_cop) and the contact parameters are not reported. Since the ablation conclusions depend on these weights, this prevents reproducibility and makes it impossible to assess sensitivity. In addition, the paired t-tests in Section V do not report the number of independent training seeds or the unit of pairing. If the 'pairs' are gait cycles from a single policy, the p-values do not support generalization across policies or participants. Please provide hyperparameter values and seed information.
minor comments (5)
  1. [§IV-A] Typo: 'Fourty-two' should be 'Forty-two.'
  2. [§III-B] Notation is inconsistent: Eq. (3) uses 'Rk' while the text discusses 'R_k'; please unify notation throughout.
  3. [Fig. 4] The text refers to 'the left pane of Fig. 4(b)' but the figure is not clearly labeled; using explicit panel labels would improve readability.
  4. [§II] The symbol τ is overloaded: it denotes both the actuator joint torques in Eq. (2) and the contact time constant in the ground-contact model. Please use separate symbols to avoid confusion.
  5. [§III-A] The state definition contains 'q −t' with an awkward superscript; please typeset the excluded-root coordinates cleanly.

Circularity Check

1 steps flagged

Minor circularity: GRF/CoP improvements are partly the training objective; the central joint-moment result is independently grounded against inverse dynamics.

specific steps
  1. fitted input called prediction [Section III-B, Eqs. (5)-(6); Section V-C, Fig. 6(b) and Table I]
    "Ground reaction force (GRF) reward Rgrf = exp(−k_grf ||Δf_grf||^2), ... Δf_grf stacks the difference of sagittal plane GRFs of learned interaction forces with the expert profiles from treadmill measurement ... Center of Pressure (CoP) reward Rcop = exp(−k_cop ∥Δx_cop∥^2), where ... Δx_cop ... represents the difference between learned and expert anterior–posterior CoP position. ... (b) CPCC evaluation of GRF prediction accuracy ... Table I: Rk only 38.3±2.7 ... ALL 21.4±2.3."

    The GRF and CoP rewards directly minimize the squared differences between simulated and measured GRF/CoP signals. Section V-C then reports these same error metrics as GRF prediction accuracy and CoP RMSE reduction. Because the policy is explicitly trained to maximize these rewards, the improved GRF/CoP agreement under KAIL is a fitted objective, not an emergent prediction. However, the central joint-moment result is evaluated against inverse dynamics, which is not a reward term, so the core claim is not fully circular.

full rationale

The paper's derivation chain is largely self-contained and not driven by self-citation. It builds on the external residual-force-control baseline of Yuan and Kitani [7], and the central comparison of MOIL versus KAIL is implemented in MuJoCo and evaluated against Visual3D inverse-dynamics references. There is no load-bearing self-citation and no imported uniqueness claim. The only reduction-by-construction element is the GRF/CoP evaluation: Eqs. (5)-(6) define the reward as exponentially weighted squared errors to the exact expert GRF and CoP signals, and Section V-C reports those same signals as prediction accuracy. Thus the external-kinetics improvements under KAIL are partly a fitting effect. The joint-moment comparison in Fig. 7 is not directly a training objective, so the core claim that adding kinetic rewards moves simulated moments closer to inverse dynamics retains independent support, even though the moments are dynamically coupled to the fitted GRF/CoP inputs through Eq. (1). The acknowledged foot-ground contact oversimplification is a correctness risk, not a circularity. Overall score 2 reflects one minor fitting-effect step with an otherwise independent central result.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

Free parameters are dominated by hand-set reward weights, impedance gains, and contact-model constants; none are fitted to the inverse-dynamics moments, which keeps the joint-moment comparison partially independent. All listed axioms are standard biomechanical or RL modeling assumptions; no new physical entities are introduced.

free parameters (4)
  • Kinetic reward weights w_grf, w_cop and kernel constants k_grf, k_cop = Not reported
    Equations (5)-(6) define the GRF and CoP rewards with these weights and kernel constants; values are not given, so the balance between kinematic and kinetic rewards is not reproducible. The ablation varies which terms are active but not their magnitudes.
  • Kinematic reward weights w_p, w_ee, w_rp, w_rv, w_vf = Not reported
    Equation (3) aggregates kinematic rewards with these weights; they are inherited from Yuan et al. [7] but values are not listed in this paper.
  • Impedance gains K_p, K_d = Not reported (from [7])
    Equation (2) generates joint torques via impedance control; stiffness and damping coefficients are preset following [7] but not reported, and they directly shape the joint moment profiles being evaluated.
  • Contact model parameters epsilon, tau, damping c = 0.05, 0.02, 1
    Section II states these values were chosen to obtain reasonably smooth foot-ground interactions; they are hand-set and directly influence GRF and CoP outputs.
axioms (6)
  • domain assumption Ground-truth joint moments are correctly given by inverse dynamics (Recursive Newton-Euler) using measured kinematics, GRFs, and a matching multi-segment model.
    Section IV-A and Fig. 7 use Visual3D inverse dynamics as the reference; if this reference is biased by soft-tissue artifacts or model mismatch, the reported kinetic errors are not meaningful.
  • domain assumption The MuJoCo floating-base rigid-body model and the Visual3D model have identical segment lengths, masses, and inertias and are adequate representations of the participant.
    Section II states the models shared anthropometric configurations to ensure consistency; no validation of this equality is provided.
  • domain assumption The box-foot contact model with four contact points per foot and parameters (epsilon=0.05, tau=0.02, damping c=1) is accurate enough to compare simulated GRF/CoP with experimental measurements.
    Section II describes the contact model, while Section VI admits it is an oversimplification. The MOIL-vs-KAIL difference could be an artifact of this model.
  • domain assumption The residual wrench xi applied to the floating base is an acceptable modeling tool and does not mask the biomechanical quantities being evaluated in a way that biases the comparison.
    Section III-A and Eq. (1) include the RFC residual wrench; the paper's conclusion that MOIL is under-constrained depends on residual forces being a legitimate way to exploit contact dynamics.
  • domain assumption After 700 iterations of pre-training and 200 iterations of fine-tuning, the policies are converged representatives of each reward condition.
    Section IV-B relies on the same checkpoint and fixed training duration; no convergence criterion or seed variance is reported.
  • domain assumption A single participant's 30 gait cycles at one speed are sufficient to support the stated conclusion about MOIL versus KAIL.
    Section IV-A reports one healthy participant and one treadmill speed; Section VI acknowledges this limits generalizability. The abstract's three-speed claim is not supported by these data.

pith-pipeline@v1.3.0-alltime-deepseek · 10518 in / 14674 out tokens · 145990 ms · 2026-08-02T18:17:33.821927+00:00 · methodology

0 comments
read the original abstract

Motion imitation learning (IL) is increasingly used in robotics and human gait modeling, yet its ability to recover biomechanically consistent joint moments without explicit kinetic information remains unclear. In this study, we examined whether motion imitation alone can estimate reasonable biological joint moments. We compare motion-only IL (MOIL) against a kinetics-aware IL (KAIL) framework that incorporates ground reaction forces (GRF) and center of pressure (CoP) in imitation rewards, with an ablation study to examine the contribution of each kinetic term. Experiments were conducted using walking data from a non-disabled participant at three speeds (0.9, 1.2, and 1.5 m/s). While both MOIL and KAIL achieved comparable kinematic tracking accuracy, MOIL exhibited substantially larger errors in GRF, CoP, and joint moment estimates relative to inverse dynamics references. In contrast, KAIL produced kinetics more consistent with biomechanical values. These findings highlight a fundamental limitation of MOIL approaches, which may lead to erroneous interpretations of gait biomechanics and downstream applications by failing to estimate consistent human-like gait kinetics.

Figures

Figures reproduced from arXiv: 2603.12408 by Edgar Lobaton, He Huang, Jangwhan Ahn, Jennie Si, Xinyi Liu.

Figure 1
Figure 1. Figure 1: Comparison of (a) a biomechanics inverse dynamics pipeline [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Models used for the inverse dynamics and imitation learning. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic of the actor-critic architecture for PPO-based [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reward ablation scheme and corresponding reward curves. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Joint angle trajectories and tracking accuracy across ablation [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ground reaction force (GRF) vector trajectories and complex [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Joint kinetics performance. (a) Joint moment estimation errors at the hip, knee, and ankle joints across ablation conditions. (b) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] 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

42 extracted references · 4 canonical work pages

  1. [1]

    A survey of robot learning from demonstration

    Brenna D Argall et al. “A survey of robot learning from demonstration”. In:Robotics and autonomous systems57.5 (2009), pp. 469–483

    2009

  2. [2]

    Reinforcement learn- ing for robot research: A comprehensive review and open issues

    Tengteng Zhang and Hongwei Mo. “Reinforcement learn- ing for robot research: A comprehensive review and open issues”. EN. In:International Journal of Advanced Robotic Systems18.3 (May 2021). Publisher: SAGE Publications, p. 17298814211007305.ISSN: 1729-8806.DOI:10.1177/ 17298814211007305

    2021

  3. [3]

    A Survey of Imitation Learning: Al- gorithms, Recent Developments, and Challenges

    Maryam Zare et al. “A Survey of Imitation Learning: Al- gorithms, Recent Developments, and Challenges”. In:IEEE Transactions on Cybernetics54.12 (2024), pp. 7173–7186. DOI:10.1109/TCYB.2024.3395626

    arXiv 2024

  4. [4]

    Siheng Zhao et al.ResMimic: From General Motion Track- ing to Humanoid Whole-body Loco-Manipulation via Resid- ual Learning. 2025. arXiv:2510.05070 [cs.RO]

    arXiv 2025

  5. [5]

    Yuke Zhu et al.Reinforcement and Imitation Learning for Diverse Visuomotor Skills. 2018. arXiv:1802 . 09564 [cs.RO]

    2018

  6. [6]

    DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills

    Xue Bin Peng et al. “DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills”. In:CoRRabs/1804.02717 (2018). arXiv:1804.02717

    Pith/arXiv arXiv 2018

  7. [7]

    Residual force control for agile human behavior imitation and extended motion synthesis

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

    2020

  8. [8]

    Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety

    Jianyu Chen, Bodi Yuan, and Masayoshi Tomizuka. “Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety”. In:2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Macau, China: IEEE Press, 2019, pp. 2884–2890. DOI:10.1109/IROS40897.2019.8968225

    arXiv 2019

  9. [9]

    Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration

    Jinning Li et al. “Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration”. In:2024 IEEE International Conference on Robotics and Automation (ICRA). 2024, pp. 7447–7454.DOI:10.1109/ ICRA57147.2024.10611123

    arXiv 2024

  10. [10]

    arXiv:2509.26633 [cs]

    Lujie Yang et al.OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction. arXiv:2509.26633 [cs]. Oct. 2025. DOI:10.48550/arXiv.2509.26633

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.26633 2025

  11. [11]

    Humanoid locomotion and manipula- tion: Current progress and challenges in control, planning, and learning

    Zhaoyuan Gu et al. “Humanoid locomotion and manipula- tion: Current progress and challenges in control, planning, and learning”. In:arXiv preprint arXiv:2501.02116(2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Hongjie Fang et al.AirExo-2: Scaling up Generalizable Robotic Imitation Learning with Low-Cost Exoskeletons

  13. [13]

    Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation

    Liyiming Ke et al. “Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation”. In:2021 IEEE International Conference on Robotics and Automation (ICRA). Xi’an, China: IEEE Press, 2021, pp. 6185–6191.DOI:10 . 1109 / ICRA48506 . 2021.9561662

    arXiv 2021

  14. [14]

    Cam- bridge University Press, 2017

    Kevin M Lynch and Frank C Park.Modern robotics. Cam- bridge University Press, 2017

    2017

  15. [15]

    Comparison of markerless and marker-based motion cap- ture systems using 95% functional limits of agreement in a linear mixed-effects modelling framework

    Kishor Das, Thiago de Paula Oliveira, and John Newell. “Comparison of markerless and marker-based motion cap- ture systems using 95% functional limits of agreement in a linear mixed-effects modelling framework”. In:Scientific Reports13.1 (2023), p. 22880

    2023

  16. [16]

    OpenCap: Human movement dy- namics from smartphone videos

    Scott D Uhlrich et al. “OpenCap: Human movement dy- namics from smartphone videos”. In:PLoS computational biology19.10 (2023), e1011462

    2023

  17. [17]

    Present and future of gait assessment in clinical practice: Towards the application of novel trends and technologies

    Abdul Aziz Hulleck et al. “Present and future of gait assessment in clinical practice: Towards the application of novel trends and technologies”. In:Frontiers in medical technology4 (2022), p. 901331

    2022

  18. [18]

    Deep reinforcement learning for modeling human locomotion control in neuromechanical simulation

    Seungmoon Song et al. “Deep reinforcement learning for modeling human locomotion control in neuromechanical simulation”. In:Journal of NeuroEngineering and Rehabili- tation18.1 (). Ed. by null.DOI:10.1186/s12984-021- 00919-y

    work page doi:10.1186/s12984-021-

  19. [19]

    Simulating the effect of ankle plantarflexion and inversion-eversion exoskeleton torques on center of mass kinematics during walking

    Nicholas A Bianco et al. “Simulating the effect of ankle plantarflexion and inversion-eversion exoskeleton torques on center of mass kinematics during walking”. In:PLoS Computational Biology19.8 (2023), e1010712

    2023

  20. [20]

    Automatically customizing a powered knee prosthesis with human in the loop using adaptive dynamic programming

    Yue Wen et al. “Automatically customizing a powered knee prosthesis with human in the loop using adaptive dynamic programming”. In:2017 International Symposium on Wear- able Robotics and Rehabilitation (WeRob). 2017, pp. 1–2. DOI:10.1109/WEROB.2017.8383835

    arXiv 2017

  21. [21]

    R. James Cotton.KinTwin: Imitation Learning with Torque and Muscle Driven Biomechanical Models Enables Precise Replication of Able-Bodied and Impaired Movement from Markerless Motion Capture. en-US. arXiv:2505.13436 [cs]. May 2025.DOI:10.48550/arXiv.2505.13436

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.13436 2025

  22. [22]

    DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Muscu- loskeletal Systems

    Pierre Schumacher et al. “DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Muscu- loskeletal Systems”. In:The Eleventh International Confer- ence on Learning Representations. 2023

    2023

  23. [23]

    AI in therapeutic and assistive exoskeletons and exosuits: Influences on perfor- mance and autonomy

    Herman Van Der Kooij et al. “AI in therapeutic and assistive exoskeletons and exosuits: Influences on perfor- mance and autonomy”. en. In:Science Robotics10.104 (July 2025), eadt7329.ISSN: 2470-9476.DOI:10.1126/ scirobotics.adt7329

    2025

  24. [24]

    arXiv:2311.02496 [cs]

    Firas Al-Hafez et al.LocoMuJoCo: A Comprehen- sive Imitation Learning Benchmark for Locomotion. arXiv:2311.02496 [cs]. Nov. 2023.DOI:10 . 48550 / arXiv.2311.02496

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.02496 2023

  25. [25]

    Learning Speed-Adaptive Walking Agent Using Imitation Learning with Physics-Informed Sim- ulation

    Yi-Hung Chiu et al. “Learning Speed-Adaptive Walking Agent Using Imitation Learning with Physics-Informed Sim- ulation”. In:2025 International Conference On Rehabili- tation Robotics (ICORR). 2025, pp. 835–841.DOI:10 . 1109/ICORR66766.2025.11063216

    arXiv 2025

  26. [26]

    MuJoCo: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. “MuJoCo: A physics engine for model-based control”. In:2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. 2012, pp. 5026–5033.DOI:10 . 1109 / IROS . 2012 . 6386109

    2012

  27. [27]

    A simple state-determined model reproduces entrainment and phase-locking of human walking

    Jooeun Ahn and Neville Hogan. “A simple state-determined model reproduces entrainment and phase-locking of human walking”. In:PloS one7.11 (2012), e47963

    2012

  28. [28]

    Simultaneous positive and negative external mechanical work in human walking

    J.Maxwell Donelan, Rodger Kram, and Arthur D. Kuo. “Simultaneous positive and negative external mechanical work in human walking”. In:Journal of Biomechanics35.1 (2002), pp. 117–124.ISSN: 0021-9290.DOI:https : / / doi.org/10.1016/S0021-9290(01)00169-5

    work page doi:10.1016/s0021-9290(01)00169-5 2002

  29. [29]

    The sensory role of the sole of the foot: Review and update on clinical perspectives

    Frederic JF Viseux. “The sensory role of the sole of the foot: Review and update on clinical perspectives”. In:Neu- rophysiologie Clinique50.1 (2020), pp. 55–68

    2020

  30. [30]

    Thomas, Joseph A

    Stephen J. Thomas, Joseph A. Zeni, and David A. Winter. Winter’s biomechanics and motor control of human move- ment / Stephen J. Thomas, Joseph A. Zeni, David A. Winter (deceased).eng. Fifth edition. Hoboken, NJ: John Wiley & Sons, Inc., 2023 - 2023.ISBN: 9781119827023

    2023

  31. [31]

    Analysis of different uncertainties in the inverse dynamic analysis of human gait

    Rosa P `amies-Vil`a et al. “Analysis of different uncertainties in the inverse dynamic analysis of human gait”. In:Mecha- nism and machine theory58 (2012), pp. 153–164

    2012

  32. [32]

    Errors in alignment of center of pressure and foot coordinates affect predicted lower extremity torques

    Steven T McCaw and Paul DeVita. “Errors in alignment of center of pressure and foot coordinates affect predicted lower extremity torques”. In:Journal of biomechanics28.8 (1995), pp. 985–988

    1995

  33. [33]

    A Quantitative Evaluation of the Frequency

    Skeletal Muscle In Vivo. “A Quantitative Evaluation of the Frequency”. In:Journal of Biomechanical Engineering101 (1979), p. 28

    1979

  34. [34]

    Is my model good enough? Best practices for verification and validation of musculoskeletal models and simulations of movement

    Jennifer L Hicks et al. “Is my model good enough? Best practices for verification and validation of musculoskeletal models and simulations of movement”. In:Journal of biome- chanical engineering137.2 (2015), p. 020905

    2015

  35. [35]

    Influence of center of pressure estimation errors on 3D inverse dynamics solutions during gait at different velocities

    Franklin Camargo-Junior et al. “Influence of center of pressure estimation errors on 3D inverse dynamics solutions during gait at different velocities”. In:Journal of applied biomechanics29.6 (2013), pp. 790–797

    2013

  36. [36]

    Learning Physically Simulated Tennis Skills from Broadcast Videos

    Haotian Zhang et al. “Learning Physically Simulated Tennis Skills from Broadcast Videos”. In:ACM Trans. Graph.(). DOI:10.1145/3592408

    doi:10.1145/3592408

  37. [37]

    Reinforcement learning-based motion imitation for physiologically plausible musculoskeletal motor control

    Merkourios Simos, Alberto Silvio Chiappa, and Alexander Mathis. “Reinforcement learning-based motion imitation for physiologically plausible musculoskeletal motor control”. In:arXiv(2025).DOI:10.48550/arXiv.2503.14637

    work page doi:10.48550/arxiv.2503.14637 2025

  38. [38]

    A neural cir- cuitry that emphasizes spinal feedback generates diverse behaviours of human locomotion

    Seungmoon Song and Hartmut Geyer. “A neural cir- cuitry that emphasizes spinal feedback generates diverse behaviours of human locomotion”. In:The Journal of Physi- ology593.16 (2015), pp. 3493–3511.DOI:https://doi. org/10.1113/JP270228

    work page doi:10.1113/jp270228 2015

  39. [39]

    Learning a Control Policy for Fall Prevention on an Assistive Walking Device

    Visak C.V . Kumar et al. “Learning a Control Policy for Fall Prevention on an Assistive Walking Device”. en-US. In: 2020 IEEE International Conference on Robotics and Au- tomation (ICRA). ISSN: 2577-087X. May 2020, pp. 4833– 4840.DOI:10.1109/ICRA40945.2020.9196798

    arXiv 2020

  40. [40]

    A review on foot-ground contact mod- eling strategies for human motion analysis

    Leonor Saraiva et al. “A review on foot-ground contact mod- eling strategies for human motion analysis”. In:Mechanism and Machine Theory177 (2022), p. 105046

    2022

  41. [41]

    The Neuromusculoskeletal Mod- eling Pipeline: MATLAB-based model personalization and treatment optimization functionality for OpenSim

    Claire V Hammond et al. “The Neuromusculoskeletal Mod- eling Pipeline: MATLAB-based model personalization and treatment optimization functionality for OpenSim”. In:Jour- nal of NeuroEngineering and Rehabilitation22.1 (2025), p. 112

    2025

  42. [2025]

    arXiv:2503.03081 [cs.RO]

    Pith/arXiv arXiv