Pith. sign in

REVIEW 2 major objections 5 minor 35 references

A dynamics correction computed from the parallel-to-serial coordinate transform plus hardware frequency-response measurements restores the missing inertia and damping of a parallel-link leg in a serial-tree simulator, cutting joint-position

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 ↗

Simulator-side normalization that adds actuator inertia redistribution and residual linkage inertia to serial-tree models of parallel-link legs reduces sim-to-real motion, torque, and force errors by 60-82%.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely useful engineering correction for serial-tree simulators of parallel-link legs, with a clean derivation and strong hardware evals, but the inertia-dominant approximation is explicitly untested and should be quantified before full trust. the 2 major comments →

arxiv 2608.01697 v1 pith:VIJCVALE submitted 2026-08-03 cs.RO cs.SYeess.SY

Bridging the Sim-to-Real Gap in Parallel-Link Leg Mechanisms via Simulator-Side Dynamics Normalization

Jinsong Hong , Jangho Kim , Jihwan Lee , Donghyun Kim , Sehoon Oh This is my paper
classification cs.RO cs.SYeess.SY
keywords sim-to-real transferparallel-link mechanismsserial-tree surrogatedynamics normalizationfrequency-response identificationinertia couplingquadruped locomotionreinforcement learning
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

This paper targets the sim-to-real gap that appears when a parallel-link leg, a high-stiffness closed-chain mechanism, is simulated as a simpler serial-chain surrogate. Its claim is that the gap is structural rather than a parameter-tuning issue: the parallel-to-serial coordinate change redistributes actuator inertia and damping into coupling terms the surrogate lacks, and the linkage's own inertia is dropped entirely. The proposed Simulator-Side System Normalization (S3N) computes correction terms from the coordinate transformation and from actuator- and leg-level frequency-response measurements, then reinjects them into the serial simulator without changing its tree topology. In the paper's tests, the full version reduced joint-position and torque errors by about 80%, ground-reaction-force error by more than 60%, and the command-normalized velocity gap of a learned circular-locomotion policy from 17.3% to 9.9%. If correct, S3N lets researchers keep their serial-tree simulator and reinforcement-learning pipeline while training policies that respond dynamically like the physical parallel mechanism.

Core claim

The paper's central claim is Eq. (6): expressed in serial coordinates, the structure-induced dynamics gap decomposes as ΔM_s(α) = ΔM_act + ΔM_link(α) and ΔD_s = ΔD_act. The first term is the inertia and damping of the physical parallel actuators, pulled back into serial coordinates through Jᵀ(·)J, minus the diagonal values the serial simulator assigns; the second is the residual inertia of the parallel linkage that the serial-tree surrogate omits. S3N-Act restores the coordinate-induced redistribution; S3N-Full additionally restores the residual linkage inertia, identified from leg-level MIMO frequency responses after subtracting the separately identified actuator contribution so it is not d

What carries the argument

The load-bearing identity is the pull-back of actuator inertia and damping under the constant Jacobian J = [[1,0],[1,1]] that maps the two serial joints to the two parallel actuation coordinates: M_p→s_act = Jᵀ M_p_act J turns the diagonal actuator inertia (J̄₁, J̄₂) into a fully coupled serial matrix with an off-diagonal J̄₂ term, whereas the serial simulator assumes a diagonal assignment — the difference, Eqs. (12)–(13), is exactly what the parallel transmission adds to the effective dynamics. The residual linkage inertia ΔM_link(α) uses the four-bar kinetic-energy coupling cos α, with the total leg inertia identified from a leg-level MIMO frequency-response measurement at a single posture

Load-bearing premise

The method corrects only inertia and damping: it assumes the gravity, Coriolis, and centrifugal effects of the parallel linkage omitted from the serial surrogate are small enough to ignore, and it identifies the leg's total inertia at a single posture (α = π/3) and extrapolates it with a cos α model across the whole configuration range.

What would settle it

Run the same 2-DoF comparison on a high-speed, large-range trajectory where the neglected Coriolis/centrifugal and gravity terms of the residual linkage are large: if joint-position and torque RMSE return toward Kin-Only levels at high speed despite S3N-Full, the inertia-dominant approximation is the binding limit. Separately, measure the leg-level FRF at several postures (α = 30°, 60°, 90°) to test whether the single-posture cos α extrapolation of ΔM_link holds.

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

If this is right

  • A policy trained in an S3N-normalized serial-tree simulator experiences hip–knee inertia coupling and damping that mimic the physical parallel linkage, so it transfers to hardware with much closer joint tracking (position and torque RMSE down about 80% in the 2-DoF test).
  • Force fidelity does not follow automatically from motion fidelity: with nearly identical closed-loop pitch motions, the GRF-norm sim-to-real RMSE dropped from 27.1 N to about 10 N only when the dynamics normalization was present.
  • Broad unstructured randomization over leg-link parameters did not reproduce S3N's improvement in locomotion transfer, so structure-aware dynamics correction and domain randomization play complementary roles rather than substitutable ones.
  • S3N requires no loop-closure constraints and no online identification at deployment: identification is performed once, and the deployed policy needs no S3N force injection on hardware.
  • The formulation applies leg-wise to all four legs of the quadruped, so the reported gains should compose across the whole machine's dynamics.

Where Pith is reading between the lines

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

  • If the inertia-dominant approximation is the binding constraint, S3N's advantage should shrink at high-speed, large-range motions where the neglected Coriolis/centrifugal and gravity terms of the residual linkage grow; repeating the same 2-DoF comparison at increasing speeds would map that boundary.
  • The cos α inertia model is identified at a single posture (α = π/3); measuring the leg-level FRF at several postures (e.g., 30°, 60°, 90°) would test whether the extrapolation holds, and would suggest a posture-scheduled ΔM_link if it does not.
  • The decomposition is generic: any closed-chain mechanism reduced to a tree surrogate with a known coordinate Jacobian admits the same Jᵀ M J pull-back correction, so S3N could be applied to other parallel linkages (humanoid legs, manipulation wrists) without altering their learning pipeline.
  • Because the correction couples hip and knee inertially, policies trained with S3N should exhibit different hip–knee acceleration correlations than Kin-Only policies — an observable, testable prediction from the rollout data itself.
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

2 major / 5 minor

Summary. The paper proposes S3N, a method to improve sim-to-real dynamic consistency when a parallel-link leg is simulated as a serial-tree surrogate. It formalizes the dynamics gap as a coordinate-induced redistribution of actuator inertia/damping (Delta M_act, Delta D_act) plus residual linkage inertia (Delta M_link(alpha)) (Eq. 6), derives these terms by pulling back actuator and link inertia from parallel to serial coordinates (Eqs. 9-16), and constructs S3N-Act and S3N-Full by adding correction terms to the serial-tree equations of motion (Eqs. 20-22). Parameters are identified from actuator-level and leg-level frequency-response measurements (Tables I and II). The method is evaluated in three settings: contact-free 2-DoF chirp tracking (position/torque RMSE reductions of 80.9%/82.1% vs. Kin-Only), GRF during pitch-in-place (RMSE reductions 65.1%/62.4%), and circular locomotion (phase-averaged command-normalized gap from 17.3% to 9.9%).

Significance. Strengths: the analytic derivation is transparent and algebraically consistent; the Jacobian pull-back in Eqs. (9)-(13) is the correct kinetic-energy/virtual-work transformation, and Eq. (19) correctly avoids double-counting actuator inertia. The evaluation is genuinely out-of-sample: the FRF identification experiments differ from the chirp, GRF, and locomotion tasks, and no task metric is used to fit the reported parameters. The reductions are large and directionally consistent. If the normalized simulator actually reproduces the dominant inertial/damping response of the parallel mechanism, this is a useful practical contribution for TBCM-based RL pipelines, preserving serial-tree topology while improving force-level fidelity. The main reservation is that the inertia-dominant approximation underlying Eq. (22) is not quantitatively supported.

major comments (2)
  1. [§III-B, Eq. (22), §VII-B] The normalized equation of motion (Eq. 22) omits the Coriolis/centrifugal and gravitational terms associated with the residual parallel linkage. Since M^{p→s}(α) depends on α=q^s_2 through cos α, the Euler-Lagrange equations that follow from the identified inertia (Eq. 14) contain velocity-dependent terms proportional to ∂M^{p→s}/∂α times q̇_i q̇_j (in particular terms involving M12 sinα q̇_1 q̇_2 and q̇_2^2). In addition, g^s does not include the gravitational torque of the parallel links discarded in the serial-tree reduction. The paper lists these as limitations in §VII-B, but gives no estimate of their magnitude. This is load-bearing because the central claim is that S3N restores the dominant dynamics; if the omitted terms are comparable to the residual 0.9 N·m torque RMSE or to the damping torques during the 25-Hz chirp, even perfect identification of inertia and damping will not ma
  2. [§IV-B, Eq. (14)] The full-leg inertia is identified at a single posture α*=π/3 and then extrapolated with the model M_p(α)=[[M11, M12 cosα],[M12 cosα, M22]] (Eq. 14). This assumes the planar four-bar coupling form holds over the entire configuration range and that M11 and M22 are posture-independent. For a four-bar linkage, diagonal inertias generally vary with configuration as well; the cosα model may capture only part of the posture dependence. Since locomotion and the chirp traverse a wide α range, an incorrect extrapolation would bias ΔM_link(α) in Eq. (19) at configurations away from the identification posture. Please justify the form with the linkage geometry, or identify at several postures and show the fitted model remains valid; at minimum, report the α-range covered in the experiments and the residuals of the cosα fit.
minor comments (5)
  1. [§IV-A] The identification procedure is described qualitatively ('damping primarily affects the low-frequency magnitude...'). The paper should state the actual fit criterion, the frequency range used for the fit, and the fit residuals for the actuator-level FRF.
  2. [Eq. (25)] M_s(α_n) is used for each leg, but the notation is not explicitly defined in that equation. Clarify that it is the serial-tree inertia with α_n substituted.
  3. [Fig. 10] The four colored simulation curves overlap and may hide the black mean hardware curve. Consider plotting simulation results as a shaded band or using a separate panel.
  4. [Table II] The +0.0% entry for M^s_22 under S3N-Act is consistent with Eq. (12) but may confuse readers; a brief footnote explaining that the actuator pull-back does not change the (2,2) entry would help.
  5. [Algorithm 1] In Step 4, τ^s_S3N is not explicitly defined in the pseudocode. Add a reference to Eqs. (25)–(27) for the compensation-torque computation.

Circularity Check

0 steps flagged

No significant circularity: S3N corrections are hardware-identified and the validation tasks are out-of-sample; the only self-citation is non-load-bearing background.

full rationale

The derivation of the S3N corrections is self-contained. The actuator-side increments (Eqs. 9-13) follow algebraically from the fixed Jacobian J via kinetic-energy invariance; they are not fitted to the evaluation data. The residual-linkage increment ΔM_link is constructed from separately measured actuator- and leg-level FRFs (Eqs. 14-19) at a fixed posture α*=π/3, and the validation tasks—2-DoF chirp tracking, pitch-in-place GRF, and circular locomotion—use different references and are not used to fit any parameter. The reported RMSE reductions are therefore out-of-sample with respect to the identification. The only self-citation is [3] by author S. Oh, used in the introduction as general background on biarticular actuation; it is not load-bearing for the S3N construction. The explicit limitation in Section VII-B that Coriolis/centrifugal and gravitational terms of the residual linkage are neglected is a disclosed modeling approximation and a correctness risk, not a circular step, because Eq. (22) is presented as approximate and the omitted terms are not silently re-introduced as predictions. No step of the derivation reduces to its own input by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper's central claim rests on two identified parameter sets (actuator-side inertia/damping and leg-level inertia) and on the inertia-dominant approximation. The parameters are physical and measured, not free tuning knobs, but their identification carries uncertainty that is not reported. The main nonstandard assumption is the neglect of velocity-dependent and gravitational residual-linkage terms, plus the extrapolation of a single-posture identification to the full configuration space.

free parameters (6)
  • Actuator inertia J_bar = N^2 J_m = 0.0162 kg m^2
    Identified from actuator torque-to-velocity FRF (Table I). Used to compute S3N-Act correction terms.
  • Actuator damping B_bar = N^2 B_m = 0.0972 N m s/rad
    Identified from the same actuator FRF (Table I). Used in S3N-Act and S3N-Full damping corrections.
  • Command-path delay T_d = 3.0 ms
    Identified from phase lag in the actuator FRF; implemented as a one-step torque delay.
  • Total parallel-leg inertia M11 = 0.04793 kg m^2
    Identified from leg-level MIMO FRF at α*=π/3 (Eq. 31).
  • Total parallel-leg coupling inertia M12 = 0.005413 kg m^2
    Identified from closed-loop sum/difference-mode FRFs (Eq. 31).
  • Total parallel-leg inertia M22 = 0.02368 kg m^2
    Identified from leg-level MIMO FRF (Eq. 31).
axioms (5)
  • domain assumption The parallel-to-serial coordinate mapping is q_p = J q_s with constant J = [[1,0],[1,1]].
    Assumed for the specific 2-DoF parallel-link leg (Eq. 1-2).
  • standard math Kinetic energy is invariant under the coordinate transformation, justifying the pull-back M = J^T M_p J.
    Standard result from the operational space formulation [26].
  • domain assumption The total parallel-leg inertia has the form M_p(α) = [[M11, M12 cos α],[M12 cos α, M22]].
    Based on planar four-bar coupling [2], [25]; constants estimated at a single posture.
  • ad hoc to paper Coriolis/centrifugal and gravitational terms of the residual linkage are negligible (inertia-dominant approximation).
    Explicitly stated in Section III-B and listed as a limitation in Section VII-B.
  • domain assumption Actuator inertia and damping are diagonal and constant in parallel coordinates.
    Assumed for the actuators; identified from the actuator FRF.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging the Sim-to-Real Gap in Parallel-Link Leg Mechanisms via Simulator-Side Dynamics Normalization." pith.science (2026). https://pith.science/paper/VIJCVALE

@misc{pith2026260801697,
  author       = {Pith},
  title        = {Pith review of: Bridging the Sim-to-Real Gap in Parallel-Link Leg Mechanisms via Simulator-Side Dynamics Normalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VIJCVALE}},
  note         = {Machine review of arXiv:2608.01697}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper addresses the sim-to-real gap in dynamics arising when a parallel-link mechanism is represented by a serial-tree surrogate in simulation. Conventional Jacobian-based state and torque mappings preserve consistency with the kinematic and virtual-work relations but do not account for the coordinate-induced redistribution of actuator inertia and damping and the linkage inertia omitted during serial-tree reduction. To address this gap, Simulator-Side System Normalization (S3N) is proposed to normalize the serial-tree simulator's effective dynamics while preserving its tree topology. S3N-Act incorporates actuator inertia and damping into the serial-coordinate dynamics through coordinate transformation, whereas S3N-Full restores residual linkage inertia by separately identifying actuator- and leg-level frequency responses. In the 2-DoF validation, S3N-Full reduced the joint-position and torque RMSEs by 80.9% and 82.1%, respectively, relative to the Jacobian-mapping baseline. During pitch-in-place motion, S3N-Act and S3N-Full reduced the RMSE of the ground reaction force norm by 65.1% and 62.4%, respectively. During circular locomotion, S3N-Full reduced the phase-averaged, command-normalized sim-to-real gap from 17.3% to 9.9%. These results show that simulator-side normalization improves motion- and force-level sim-to-real consistency. It enables policy training in a serial-tree framework with hardware-consistent dynamics that better represent the physical parallel-link mechanism.

Figures

Figures reproduced from arXiv: 2608.01697 by Donghyun Kim, Jangho Kim, Jihwan Lee, Jinsong Hong, Sehoon Oh.

Figure 2
Figure 2. Figure 2: Coordinate definitions for the serial-tree simulator and the physical [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Simulator-side dynamics normalization with S3N. All variants retain the serial-tree simulator and use the same coordinate mapping. S3N-Act additionally [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Measured and fitted actuator torque-to-velocity FRFs with parameter [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulation-to-hardware joint-position error, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Simulation-to-hardware joint-torque error, [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Summary of joint-position and joint-torque RMSEs for the 2-DoF leg [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of closed-loop pitch-velocity responses. Each hardware [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Rear-right GRF-norm sim-to-real gap under similar closed-loop pitch [PITH_FULL_IMAGE:figures/full_fig_p007_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Phase-aligned, command-normalized sim-to-real velocity gaps during circular locomotion. The angular coordinate denotes the circular-command [PITH_FULL_IMAGE:figures/full_fig_p008_12.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

35 extracted references · 28 canonical work pages · 2 internal anchors

  1. [1]

    Merlet,Parallel Robots, 2nd ed

    J.-P. Merlet,Parallel Robots, 2nd ed. Springer, 2006

    work page 2006

  2. [2]

    Briot and W

    S. Briot and W. Khalil,Dynamics of Parallel Robots: From Rigid Bodies to Flexible Elements. Springer, 2015

    work page 2015

  3. [3]

    Mathematical and experi- mental verification of efficient force transmission by biarticular muscle actuator,

    S. Oh, V . Salvucci, Y . Kimura, and Y . Hori, “Mathematical and experi- mental verification of efficient force transmission by biarticular muscle actuator,” inProceedings of the 18th IFAC World Congress, ser. IFAC Proceedings V olumes, vol. 44, no. 1. IFAC Secretariat, 2011, pp. 13 516–13 521

    work page 2011

  4. [4]

    A parallel actuated pantograph leg for high-speed locomotion,

    W. Guo, C. Cai, M. Li, F. Zha, P. Wang, and K. Wang, “A parallel actuated pantograph leg for high-speed locomotion,”Journal of Bionic Engineering, vol. 14, no. 2, pp. 202–217, 2017

    work page 2017

  5. [5]

    Proprioceptive actuator design in the MIT cheetah: Impact mitigation and high-bandwidth physical interaction for dynamic legged robots,

    P. M. Wensing, A. Wang, S. Seok, D. Otten, J. Lang, and S. Kim, “Proprioceptive actuator design in the MIT cheetah: Impact mitigation and high-bandwidth physical interaction for dynamic legged robots,” IEEE Transactions on Robotics, vol. 33, no. 3, pp. 509–522, 2017

    2017

  6. [6]

    Modeling and Numerical Analysis of Kangaroo Lower Body based on Constrained Dynamics of Hybrid Serial-Parallel Floating-Base Systems

    E. M. Hoffman, A. Curti, N. Miguel, S. K. Kothakota, A. Molina, A. Roig, and L. Marchionni, “Modeling and numerical analysis of kangaroo lower body based on constrained dynamics of hybrid serial– parallel floating-base systems,”Robotics and Autonomous Systems, vol. 182, p. 104827, 2024, arXiv:2312.04161

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Mechanical intelligence- aware curriculum reinforcement learning for humanoids with parallel actuation,

    Y . Tanaka, A. Zhu, Q. Wang, and D. Hong, “Mechanical intelligence- aware curriculum reinforcement learning for humanoids with parallel actuation,” in2025 IEEE-RAS 24th International Conference on Hu- manoid Robots (Humanoids), 2025, pp. 882–889. IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. XX, NO. X, MONTH YEAR 10

    work page 2025

  8. [8]

    High dynamic position control for a typical hydraulic quadruped robot leg based on virtual decomposition control,

    K. Zhang, J. Zhang, H. Zong, L. Fang, J. Shen, M. Cheng, and B. Xu, “High dynamic position control for a typical hydraulic quadruped robot leg based on virtual decomposition control,”IEEE/ASME Transactions on Mechatronics, vol. 30, no. 4, pp. 2473–2484, Aug. 2025

    work page 2025

  9. [9]

    MuJoCo: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “MuJoCo: A physics engine for model-based control,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 5026–5033

    work page 2012

  10. [10]

    Learning to walk with hybrid serial– parallel linkages: A case study on the kangaroo robot,

    F. Amadio, H. Li, L. Uttini, D. Kanoulas, S. Ivaldi, V . Modugno, and E. M. Hoffman, “Learning to walk with hybrid serial– parallel linkages: A case study on the kangaroo robot,” inResults from the 13th International Conference on Robot Intelligence Technology and Applications, 2025, in press. [Online]. Available: https://discovery.ucl.ac.uk/id/eprint/10216433/

    work page arXiv 2025

  11. [11]

    MuJoCo documentation: Computation,

    Google DeepMind, “MuJoCo documentation: Computation,” https://mujoco.readthedocs.io/en/stable/computation/index.html, accessed: 2026-03-19

    work page 2026

  12. [12]

    MuJoCo documentation: Modeling,

    ——, “MuJoCo documentation: Modeling,” https://mujoco.readthedocs. io/en/stable/modeling.html, accessed: 2026-03-19

    work page 2026

  13. [13]

    NimbRo-OP2: Grown- up 3d printed open humanoid platform for research,

    G. Ficht, P. Allgeuer, H. Farazi, and S. Behnke, “NimbRo-OP2: Grown- up 3d printed open humanoid platform for research,” in2017 IEEE-RAS 17th International Conference on Humanoid Robotics (Humanoids), 2017, pp. 669–675

    work page 2017

  14. [14]

    Learning to Hop for a Single-Legged Robot with Parallel Mechanism

    H. Zhang, X. Chu, Y . Chen, Y . Tang, L. Yue, Y .-H. Liu, and K. W. S. Au, “Learning to hop for a single-legged robot with parallel mechanism,” arXiv preprint arXiv:2501.11945, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Cat-like jumping and landing of legged robots in low gravity using deep reinforcement learning,

    N. Rudin, H. Kolvenbach, V . Tsounis, and M. Hutter, “Cat-like jumping and landing of legged robots in low gravity using deep reinforcement learning,”IEEE Transactions on Robotics, vol. 38, no. 1, pp. 317–328, 2022

    2022

  16. [16]

    Booster gym: An end-to-end reinforcement learning framework for humanoid robot locomotion,

    Y . Wang, P. Chen, X. Han, F. Wu, and M. Zhao, “Booster gym: An end-to-end reinforcement learning framework for humanoid robot locomotion,”arXiv preprint arXiv:2506.15132, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Real-world humanoid locomotion with reinforcement learning,

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath, “Real-world humanoid locomotion with reinforcement learning,”Sci- ence Robotics, vol. 9, no. 89, p. eadi9579, 2024

    2024

  18. [18]

    LiPS: Large-scale humanoid robot reinforcement learning with parallel-series structures,

    Q. Zhang, G. Han, J. Sun, W. Zhao, J. Cao, J. Wang, H. Cheng, L. Zhang, Y . Guo, and R. Xu, “LiPS: Large-scale humanoid robot reinforcement learning with parallel-series structures,”arXiv preprint arXiv:2503.08349, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Control of humanoid robots with parallel mechanisms using differential actuation models,

    V . Lutz, L. de Matte ¨ıs, V . Batto, and N. Mansard, “Control of humanoid robots with parallel mechanisms using differential actuation models,” in Proceedings of the 2026 IEEE International Conference on Robotics and Automation (ICRA), Vienna, Austria, Jun. 2026, paper TuI1I.90

    work page 2026

  20. [20]

    Model sim- plification for dynamic control of series–parallel hybrid robots: A representative study on the effects of neglected dynamics,

    S. Kumar, J. Martensen, A. Mueller, and F. Kirchner, “Model sim- plification for dynamic control of series–parallel hybrid robots: A representative study on the effects of neglected dynamics,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019, pp. 5701–5708

    work page 2019

  21. [21]

    Sim-to-real transfer of robotic control with dynamics randomization,

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel, “Sim-to-real transfer of robotic control with dynamics randomization,” inProceedings of the IEEE International Conference on Robotics and Automation, 2018, pp. 3803–3810

    work page 2018

  22. [22]

    RMA: Rapid motor adaptation for legged robots,

    A. Kumar, Z. Fu, D. Pathak, and J. Malik, “RMA: Rapid motor adaptation for legged robots,” inRobotics: Science and Systems, 2021

    2021

  23. [23]

    An efficient learning control framework with sim-to-real for string-type artificial muscle- driven robotic systems,

    J. Tao, Y . Zhang, S. K. Rajendran, and F. Zhang, “An efficient learning control framework with sim-to-real for string-type artificial muscle- driven robotic systems,”IEEE/ASME Transactions on Mechatronics, vol. 31, no. 1, pp. 444–455, Feb. 2026

    work page 2026

  24. [24]

    NVIDIA Isaac Sim,

    NVIDIA, “NVIDIA Isaac Sim,” Software, version 5.1.0, 2025, released October 30, 2025. [Online]. Available: https://docs.isaacsim.omniverse. nvidia.com/5.1.0/

    work page 2025

  25. [25]

    K. M. Lynch and F. C. Park,Modern Robotics: Mechanics, Planning, and Control. Cambridge University Press, 2017

    work page 2017

  26. [26]

    A unified approach for motion and force control of robot manipulators: The operational space formulation,

    O. Khatib, “A unified approach for motion and force control of robot manipulators: The operational space formulation,”IEEE Journal on Robotics and Automation, vol. 3, no. 1, pp. 43–53, Feb. 1987

    work page 1987

  27. [27]

    Synthesis of low-peak-factor signals and binary sequences with low autocorrelation,

    M. R. Schroeder, “Synthesis of low-peak-factor signals and binary sequences with low autocorrelation,”IEEE Transactions on Information Theory, vol. 16, no. 1, pp. 85–89, 1970

    work page 1970

  28. [28]

    Pintelon and J

    R. Pintelon and J. Schoukens,System Identification: A Frequency Domain Approach, 2nd ed. Wiley-IEEE Press, 2012

    work page 2012

  29. [29]

    Experimental determination of frequency response function estimates for flexible joint industrial manipulators with serial kinematics,

    F. Saupe and A. Knoblach, “Experimental determination of frequency response function estimates for flexible joint industrial manipulators with serial kinematics,”Mechanical Systems and Signal Processing, vol. 52– 53, pp. 60–72, 2015

    work page 2015

  30. [30]

    Accurate FRF identification of LPV systems: nD-LPM with application to a medical X-ray system,

    R. van der Maas, A. van der Maas, R. J. V oorhoeve, and T. A. E. Oomen, “Accurate FRF identification of LPV systems: nD-LPM with application to a medical X-ray system,”IEEE Transactions on Control Systems Technology, vol. 25, no. 5, pp. 1724–1735, 2017

    work page 2017

  31. [31]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  32. [32]

    Sim-to-real: Learning agile locomotion for quadruped robots,

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, and V . Vanhoucke, “Sim-to-real: Learning agile locomotion for quadruped robots,” inProceedings of Robotics: Science and Systems, 2018

    work page 2018

  33. [33]

    Using data- driven domain randomization to transfer robust control policies to mobile robots,

    M. Sheckells, G. Garimella, S. Mishra, and M. Kobilarov, “Using data- driven domain randomization to transfer robust control policies to mobile robots,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 3224–3230

    work page 2019

  34. [34]

    Domain randomization via entropy maximization,

    G. Tiboni, P. Klink, J. Peters, T. Tommasi, C. D’Eramo, and G. Chalvatzaki, “Domain randomization via entropy maximization,” in International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=GXtmuiVrOM

    work page 2024

  35. [35]

    Closing the sim-to-real loop: Adapting simula- tion randomization with real world experience,

    Y . Chebotar, A. Handa, V . Makoviychuk, M. Macklin, J. Issac, N. D. Ratliff, and D. Fox, “Closing the sim-to-real loop: Adapting simula- tion randomization with real world experience,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 8973–8979

    work page 2019

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.