Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Shooting for Contact: Contact-Implicit Multiple Shooting for Dynamic Motion Retargeting

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

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

Pith's one-line read DSMS embeds a differentiable simulator inside a nonlinear program to turn kinematically plausible references into whole-body robot trajectories that are dynamically feasible by construction.

desk verdict DSMS is a clean, well-tested retargeting pipeline whose 'dynamic feasibility' is only as strong as MuJoCo's contact model—send to review with requests for code, sensitivity analysis, and quantitative hardware results. read the letter →

arxiv 2608.03116 v1 pith:5CRWRYVF submitted 2026-08-04 cs.RO cs.SYeess.SY

Sergio A. Esteban , Jason H. K. Siu , Derrick Mach , Junheng Li , Vince Kurtz , Joel W. Burdick , Aaron D. Ames This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords motionretargetingcontact-implicittrajectoryoptimizationdifferentiablesimulationmultipleshootingreinforcementlearninghumanoidrobotsim-to-realtransfer
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 introduces a way to convert kinematically plausible reference motions—from human motion capture, animation, or reduced-order models—into whole-body trajectories that respect the robot's actual contact, friction, impact, self-collision, and actuation limits. The method, called DSMS, runs a differentiable simulator as the dynamics model inside a multiple-shooting optimizer, so the optimizer never needs a prescribed contact schedule or explicit contact constraints. The paper claims that the resulting references are dynamically feasible by construction, that policies trained on them learn faster and track more accurately than policies trained on kinematic-only references, and that the approach transfers zero-shot to a real Unitree G1 for crawling and a 180-degree jump-turn. The significance is that it removes the contact-scheduling bottleneck from retargeting and gives reinforcement learning a physically meaningful nominal trajectory to stabilize.

What carries the argument

The central object is the DSMS nonlinear program, whose transition constraint uses the discrete flow map of a differentiable simulator instead of an analytic whole-body dynamics equation. Multiple shooting splits the horizon into N intervals, treats each shooting-node state as a decision variable, and enforces continuity through defect constraints: the simulator rollout starting at one node must land exactly on the next node's state. The simulator advances fine substeps internally, so stiff contact dynamics are resolved while the NLP stays small, and its convex contact formulation provides gradients via finite differences. The cost function blends state and key-body tracking with torque and command-rate regularization, and the same formulation is reused in a receding-horizon mode for dynamic maneuvers and in a limit-cycle-closure mode for synthesizing periodic gait libraries.

What would settle it

Record the ground-reaction forces and joint torques on the Unitree G1 while it executes the DSMS-optimized 180-degree jump-turn and compare them against the values predicted by the simulator along the optimized trajectory; a disagreement beyond sensor tolerance in the landing phase, or a trajectory that fails when replayed in a second simulator with different contact parameters, would falsify the dynamic-feasibility claim.

Watch

Extended reading notes

Core claim

The central discovery is that a direct simulation-based multiple-shooting nonlinear program can serve as a general retargeting filter: it takes any kinematically feasible reference and returns a trajectory that satisfies the full-order rigid-body dynamics as represented by a differentiable simulator. Because the simulator internally resolves contact, friction, impacts, self-collision, and joint limits, the optimizer can adjust states and commands to make or break contact as needed without explicit complementarity constraints. The paper shows that these dynamically feasible references accelerate motion-imitation RL training, improve landing success and tracking accuracy over kinematic, kinodynamic, and sampling-based retargeting baselines, and support zero-shot hardware deployment of contact-rich crawling and a highly dynamic jump-turn on the Unitree G1.

Load-bearing premise

The claim that DSMS trajectories are dynamically feasible by construction rests on the assumption that the simulator's discrete contact model, advanced by fine substeps and differentiated by finite differences, is a faithful enough model of the real robot's contact-rich dynamics that feasibility in simulation implies feasibility on hardware.

Editorial extensions

If this is right

  • Policies trained on DSMS-refined references converge faster, reach higher reward, and achieve higher landing success and lower tracking error than policies trained on raw kinematic, single-rigid-body, or kinodynamic references.
  • DSMS can be applied at any stage of a reference-generation pipeline, so a user can start from a simple rigid-body model or a motion-capture clip and still obtain a full-order dynamically feasible humanoid trajectory.
  • The limit-cycle gait library converts a few short motion clips into a continuous velocity-command interface, allowing one policy to execute forward, backward, and turning crawls at commanded speeds.
  • Because the optimizer does not prescribe a contact schedule, the method naturally handles sliding contacts and unplanned body-part contacts such as hands, elbows, and knees without retuning, and it supports arbitrary equality and inequality constraints.
  • Zero-shot deployment on the Unitree G1 suggests that dynamically feasible references can remove the need for real-world fine-tuning for at least some contact-rich behaviors.

Reading between the lines

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

  • If the central claim holds, DSMS references could also serve as warm starts or rollout targets for contact-implicit model predictive control, potentially improving closed-loop robustness beyond what RL alone achieves, though the paper does not explore this combination.
  • The 'dynamically feasible by construction' claim is only as strong as the simulator's contact model; swapping the differentiable simulator for another one with a different contact discretization would reveal how much of the result is simulator-specific.
  • A sharper hardware test than task-level success would compare predicted and measured ground-reaction forces or joint torques during the jump-turn, which the paper does not report; such a comparison would directly validate the simulator's role in the feasibility claim.
  • A natural extension consistent with the method is to generate failure-prone or corner-case references, such as recovery from pushes, by encoding the disturbance as a constraint or cost, though the paper does not attempt this.
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

4 major / 5 minor

Summary. The paper proposes DSMS, a direct simulation-based multiple-shooting NLP for motion retargeting. It embeds MuJoCo's differentiable discrete transition map as the dynamics model, so contact, friction, impacts, self-collision, and actuation limits are handled internally. Tracking and task constraints are imposed at shooting nodes, and the method is applied to reduced-order-model and motion-capture references to generate dynamically feasible trajectories and command-conditioned gait libraries. The paper evaluates the resulting references as imitation targets for RL in simulation, comparing against SRB, kinodynamic, OmniRetarget, and DynaRetarget baselines, and presents qualitative zero-shot hardware demos on a Unitree G1 for crawling and a 180-degree jump-turn.

Significance. If the central claims hold, DSMS is a practically valuable contribution: it offers a general, contact-implicit retargeting pipeline that avoids prescribing a contact schedule and produces references that accelerate RL training. The paper's strengths include the clean NLP formulation, the use of multiple shooting with fine simulator substeps, the careful sim-to-sim protocol with five seeds and a separate MuJoCo instance, and the breadth of evaluated behaviors. The comparisons in Tables I and II provide concrete evidence that DSMS references lead to higher landing success and faster convergence than the tested alternatives. However, the dynamic-feasibility and zero-shot sim-to-real claims are currently stronger than the evidence: feasibility is established only with respect to MuJoCo's soft-contact model, and the hardware evidence is qualitative.

major comments (4)
  1. [Sec. III-A, Eq. (5b), Sec. III-C] The claim that DSMS trajectories are dynamically feasible 'by construction' is only with respect to the discrete forward map F in Eq. (5b), which is implemented by MuJoCo's convex soft-contact time-stepping model, not by the continuous rigid-body dynamics in Eq. (1). Since MuJoCo permits penetration and finite compliance, and since finite-difference gradients are not derivatives of a hard-contact model, a trajectory satisfying the defect constraints may rely on virtual compliance or unmodeled impulses that the physical G1 cannot reproduce. Because dynamic feasibility is the stated reason the references should transfer to RL and hardware, this gap is load-bearing. I request either a more precise claim ('MuJoCo-feasible') or additional evidence: e.g., report peak penetration depths and contact-force residuals along optimized trajectories, verify a subset in a second simulator with a different contact model, or compare optimized joint torques with hardware-recorded actuator limits.
  2. [Sec. V-B] The zero-shot sim-to-real evidence is presented qualitatively (Fig. 1, Fig. 4) for two behaviors, with no success rates, tracking error metrics, or actuator/contact-force verification on hardware. The abstract's 'zero-shot sim-to-real transfer' claim is therefore not quantitatively established. Please report the number of trials, success criteria, and representative tracking or velocity metrics for the jump-turn and the crawling courses, or soften the claim to a qualitative demonstration.
  3. [Sec. II-B and Sec. V-A] The method depends on the assumption that finite-difference gradients of MuJoCo's stiff contact dynamics are sufficiently accurate for IPOPT with L-BFGS to converge, which the text describes as working 'surprisingly well in practice.' No sensitivity study is provided for the finite-difference step size, the substep count S, or the shooting interval Δt, and no comparison is made with analytic gradients from MJX, even though the paper notes MJX provides second-order derivatives. Because NLP convergence is the core computational claim, please add a small sensitivity experiment (e.g., gradient check against MJX autodiff, or convergence under varying step size and S) and state the operating point used for the reported results.
  4. [Sec. V-D] The sim-to-sim evaluation uses the same MuJoCo physics for reference optimization, RL training, and policy evaluation, so the comparison against OR, DR, and BS may partly reflect the reference's compatibility with MuJoCo's specific contact model rather than with real dynamics. The separate evaluation instance with asynchronous control and randomized initial configurations is a reasonable mitigation, but it does not change the contact model. Please discuss this limitation explicitly in the ablation section, and if feasible evaluate at least one trajectory in a second physics engine or with model perturbations to test sensitivity to contact parameters.
minor comments (5)
  1. [Sec. III-C] The quantity Δt_ctrl is used but never defined; please define it in the discretization paragraph of Sec. III-A.
  2. [Sec. V-A] The text 'HSLma57 solver' should be formatted as 'HSL MA57 solver', and the capitalization of MuJoCo should be made consistent throughout.
  3. [Table III] The 'Solver Seq.' row contains 'Unspecified' for SPARK; please identify the solver or add a footnote explaining why it is not specified.
  4. [Sec. V-D.2] The phrase 'our method's∼12-minute optimization time' has a formatting issue; it should read 'our method's ∼12-minute optimization time'.
  5. [Fig. 6 caption] The caption uses 'actual' and 'achieved' interchangeably for the same velocity quantity; please choose one term and use it consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DSMS feasibility is defined by its own simulator dynamics constraint, and downstream comparisons are empirical rather than fitted.

full rationale

The paper's derivation chain is self-contained. DSMS transcribes the optimal control problem with defect constraints x_{k+1}=F(x_k,u_k), where F is MuJoCo's discrete transition map. The claim that optimized trajectories are dynamically feasible by construction refers to satisfying this model constraint by construction, which is a direct consequence of the NLP feasibility constraint (5b), not a circular reduction. The step that might appear circular—'Concatenating the executed segments produces a long rollout whose transitions are generated directly by the simulator and therefore satisfy (1) by construction'—is an overstatement about model fidelity (F is not Eq. (1)), but it is not a case where an output is defined in terms of an input or a fitted parameter is renamed as a prediction. The RL acceleration and tracking-error claims are empirical comparisons against SRB, KD, OR, and DR baselines with held-out evaluation in a separate MuJoCo instance plus qualitative zero-shot hardware demos; these are external evidence rather than artifacts of the optimization objective. Self-citations support algorithmic context but are not load-bearing uniqueness theorems. If the simulator-fidelity assumption fails, the feasibility claim would be weakened, but that is a validity threat, not circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The central claim depends on the simulator as a trust anchor and on a handful of hand-tuned weights and discretization choices; none of these are fitted to external data, but they are not all reported numerically.

free parameters (4)
  • Tracking cost weights Q_x, Q_y, R_tau, R_delta_u, terminal gain gamma = Not reported
    Chosen by hand in Sec. III-B; the generated trajectories and downstream RL results depend on these weights, but exact values are not stated.
  • No-slip penalty weight for gait synthesis = Not reported
    Added in Sec. III-D as a soft penalty on horizontal velocity of planted contact links; its value determines how much sliding is allowed in the crawling library.
  • PD gains (Kp, Kd) for position-control interface = Not reported
    Referenced in Sec. III-A footnote; used to convert position commands into torques in the simulator and on hardware, and not specified.
  • Control discretization (Delta_t, S, N, H, Delta_t_ctrl) = Not reported
    Shooting interval, simulator substeps, horizon length, receding-horizon window, and execution cadence are declared in Sec. III-A and Sec. III-C but exact numbers are not given.
assumptions (3)
  • domain assumption MuJoCo's convex contact model accurately captures rigid-body contacts, friction, impacts, self-collision, and joint limits for the Unitree G1.
    Invoked in Sec. II-B and Sec. III-A; the dynamic feasibility claim is defined by this model.
  • ad hoc to paper Finite-difference gradients of the simulator are sufficiently accurate for IPOPT with L-BFGS to converge on contact-rich problems.
    Stated in Sec. V-A without a sensitivity or convergence study.
  • domain assumption The reference trajectories used as inputs, such as BONES-SEED, SRB, and IK, are kinematically feasible enough that tracking costs can guide the optimizer to useful solutions.
    Needed for the cost functions in Sec. III-B to anchor the solution; not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Shooting for Contact: Contact-Implicit Multiple Shooting for Dynamic Motion Retargeting." pith.science (2026). https://pith.science/paper/5CRWRYVF

@misc{pith2026260803116,
  author       = {Pith},
  title        = {Pith review of: Shooting for Contact: Contact-Implicit Multiple Shooting for Dynamic Motion Retargeting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CRWRYVF}},
  note         = {Machine review of arXiv:2608.03116}
}
read the original abstract

Motion retargeting approaches often prioritize kinematic similarity over whole-body dynamics, contact consistency, and actuation limits, yielding references that are difficult for reinforcement learning (RL) policies to reproduce, particularly for contact-rich behaviors. We present a contact-implicit, direct simulation-based multiple shooting (DSMS) framework that transforms kinematically feasible references into dynamically feasible whole-body trajectories. By embedding a differentiable simulator within a nonlinear program, DSMS resolves contact, friction, impacts, self-collision, and joint limits internally while enforcing tracking, actuation, and task constraints without prescribing a contact schedule or introducing explicit contact constraints. Compared with existing retargeting methods, DSMS accelerates motion-imitation RL training and yields policies with high success rates and low tracking error. We further demonstrate zero-shot sim-to-real transfer on the Unitree G1 through command-conditioned contact-rich crawling and a highly dynamic 180-degree jump-turn.

Figures

Figures reproduced from arXiv: 2608.03116 by the authors.

Figure 1
Figure 1. DSMS trajectory optimization solutions in simulation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed multi-shooting trajectory optimization. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) A gait reference is closed into a command-constrained limit cycle: the planar pose advances by [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Experimental demonstrations of (a) crawling forward and backward under height-constrained space, (b) crawling [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Simulation snapshots for dynamic contact-rich whole-body maneuvers: (a) humanoid backflip, (b) humanoid super hero [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Commanded (dashed), achieved body-frame instan [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Training performance of motions generated from the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 24 canonical work pages

  1. [1]

    Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,

    Z. Gu, J. Li, W. Shen, W. Yu, Z. Xie, S. McCrory, X. Cheng, A. Shamsah, R. Griffin, C. K. Liu,et al., “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,”IEEE/ASME Transactions on Mechatronics, vol. 31, no. 2, pp. 2300–2330, 2026

    work page 2026

  2. [2]

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

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

    arXiv 2025

  3. [3]

    Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,”ACM TOG, vol. 37, no. 4, pp. 1–14, 2018

    work page 2018

  4. [4]

    Visual imitation enables contextual humanoid control,

    A. Allshire, H. Choi, J. Zhang, D. McAllister, A. Zhang, C. M. Kim, T. Darrell, P. Abbeel, J. Malik, and A. Kanazawa, “Visual imitation enables contextual humanoid control,” inRobot Learning

    work page

  5. [5]

    Omniretarget: Interaction-preserving data generation for humanoid whole-body loco-manipulation and scene interaction,

    L. Yang, X. Huang, Z. Wu, A. Kanazawa, P. Abbeel, C. Sferrazza, C. K. Liu, R. Duan, and G. Shi, “Omniretarget: Interaction-preserving data generation for humanoid whole-body loco-manipulation and scene interaction,”arXiv:2509.26633, 2025

    arXiv 2025

  6. [6]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan,et al., “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,” arXiv:2502.01143, 2025

    arXiv 2025

  7. [7]

    Smpl-ik: Learned morphology-aware inverse kinematics for ai driven artistic workflows,

    V . V oleti, B. Oreshkin, F. Bocquelet, F. Harvey, L.-S. M ´enard, and C. Pal, “Smpl-ik: Learned morphology-aware inverse kinematics for ai driven artistic workflows,” inSIGGRAPH Asia 2022 Technical Communications, pp. 1–7, 2022

    work page 2022

  8. [8]

    Motion retargeting for humanoid robots based on simultaneous morphing parameter identification and mo- tion optimization,

    K. Ayusawa and E. Yoshida, “Motion retargeting for humanoid robots based on simultaneous morphing parameter identification and mo- tion optimization,”IEEE Transactions on Robotics, vol. 33, no. 6, pp. 1343–1357, 2017

    work page 2017

Show all 43 references
  1. [9]

    Reduced- order model guided contact-implicit model predictive control for humanoid locomotion,

    S. A. Esteban, V . Kurtz, A. B. Ghansah, and A. D. Ames, “Reduced- order model guided contact-implicit model predictive control for humanoid locomotion,” in2025 IEEE ICRA, pp. 14735–14741, IEEE, 2025

    2025

  2. [10]

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

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

    2025

  3. [11]

    Spark: Skeleton-parameter aligned retargeting on humanoid robots with kinodynamic trajectory optimization,

    H. Wang, Q. Liao, B. Zhang, K. Ren, K. Sreenath, and X. Xiong, “Spark: Skeleton-parameter aligned retargeting on humanoid robots with kinodynamic trajectory optimization,”arXiv:2603.11480, 2026

    2026

  4. [12]

    Retargeting matters: General motion retargeting for humanoid motion tracking,

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

    2025

  5. [13]

    Make tracking easy: Neural motion retargeting for humanoid whole-body control,

    Q. Zhao, K. Yang, X. Wang, S. Zhao, Y . Lu, X. Zhang, Q. Shen, X.- X. Long, and X. Cao, “Make tracking easy: Neural motion retargeting for humanoid whole-body control,”arXiv:2603.22201, 2026

    2026 arXiv

  6. [14]

    Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,

    F. Liu, Z. Gu, Y . Cai, Z. Zhou, H. Jung, J. Jang, S. Zhao, S. Ha, Y . Chen, D. Xu,et al., “Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,” IEEE Robotics and Automation Letters, 2025

    2025

  7. [15]

    Bi-level motion imitation for humanoid robots,

    W. Zhao, Y . Zhao, J. Pajarinen, and M. Muehlebach, “Bi-level motion imitation for humanoid robots,” inConference on Robot Learning, pp. 963–981, PMLR, 2025

    2025

  8. [16]

    Omnitrack: General motion tracking via physics-consistent reference,

    Y . Li, P. Zhi, Y . Wang, T. Liu, S. Yan, W. Liu, X. Wang, B. Jia, and S. Huang, “Omnitrack: General motion tracking via physics-consistent reference,”arXiv:2602.23832, 2026

    2026

  9. [17]

    Dynaretarget: Dynamically-feasible retargeting using sampling-based trajectory optimization,

    V . Dhedin, I. Taouil, S. Omar, D. Yu, K. Tao, A. Dai, and M. Khadiv, “Dynaretarget: Dynamically-feasible retargeting using sampling-based trajectory optimization,”arXiv:2602.06827, 2026

    2026 arXiv

  10. [18]

    Direct dynamic retargeting for humanoid imitation learning from videos,

    C. Roux, L. De Matte ¨ıs, A. Jordana, V . Guillet, N. Mansard, O. Stasse, and P. Sou `eres, “Direct dynamic retargeting for humanoid imitation learning from videos,”arXiv:2605.23762, 2026

    2026 arXiv

  11. [19]

    3d dynamic walking with underactuated humanoid robots: A direct collocation framework for optimizing hybrid zero dynamics,

    A. Hereid, E. A. Cousineau, C. M. Hubicki, and A. D. Ames, “3d dynamic walking with underactuated humanoid robots: A direct collocation framework for optimizing hybrid zero dynamics,” in2016 IEEE ICRA, pp. 1447–1454, 2016

    2016

  12. [20]

    Chasing autonomy: Dynamic retargeting and control guided rl for performant and controllable humanoid running,

    Z. Olkin, W. D. Compton, R. M. Bena, and A. D. Ames, “Chasing autonomy: Dynamic retargeting and control guided rl for performant and controllable humanoid running,”arXiv:2603.25902, 2026

    2026

  13. [21]

    A direct method for trajectory op- timization of rigid bodies through contact,

    M. Posa, C. Cantu, and R. Tedrake, “A direct method for trajectory op- timization of rigid bodies through contact,”The International Journal of Robotics Research, vol. 33, no. 1, pp. 69–81, 2014

    2014

  14. [22]

    Variational contact-implicit tra- jectory optimization,

    Z. Manchester and S. Kuindersma, “Variational contact-implicit tra- jectory optimization,” inRobotics Research: The 18th International Symposium ISRR, pp. 985–1000, Springer, 2019

    2019

  15. [23]

    Fast contact-implicit model pre- dictive control,

    S. Le Cleac’h, T. A. Howell, S. Yang, C.-Y . Lee, J. Zhang, A. Bishop, M. Schwager, and Z. Manchester, “Fast contact-implicit model pre- dictive control,”Robotics, vol. 40, pp. 1617–1629, 2024

    2024

  16. [24]

    Inverse dynamics trajectory optimization for contact-implicit model predictive control,

    V . Kurtz, A. Castro, A. ¨O. ¨Onol, and H. Lin, “Inverse dynamics trajectory optimization for contact-implicit model predictive control,” The International Journal of Robotics Research, vol. 45, no. 1, pp. 23– 40, 2026

    2026

  17. [25]

    Contact-implicit differ- ential dynamic programming for model predictive control with relaxed complementarity constraints,

    G. Kim, D. Kang, J.-H. Kim, and H.-W. Park, “Contact-implicit differ- ential dynamic programming for model predictive control with relaxed complementarity constraints,” in2022 IEEE/RSJ IROS, pp. 11978– 11985, IEEE, 2022

    2022

  18. [26]

    Contact-implicit trajectory optimization for dynamic object manipu- lation,

    J.-P. Sleiman, J. Carius, R. Grandia, M. Wermelinger, and M. Hutter, “Contact-implicit trajectory optimization for dynamic object manipu- lation,” in2019 IEEE/RSJ IROS, pp. 6814–6821, IEEE, 2019

    2019

  19. [27]

    Contact-implicit trajectory optimization with hydroelastic contact and ilqr,

    V . Kurtz and H. Lin, “Contact-implicit trajectory optimization with hydroelastic contact and ilqr,” in2022 IEEE/RSJ IROS, pp. 8829– 8834, IEEE, 2022

    2022

  20. [28]

    Whole-body model-predictive control of legged robots with mujoco, 2025,

    J. Z. Zhang, T. A. Howell, Z. Yi, C. Pan, G. Shi, G. Qu, T. Erez, Y . Tassa, and Z. Manchester, “Whole-body model-predictive control of legged robots with mujoco, 2025,”URL https://arxiv. org/abs/2503.04613, 2025

    2025

  21. [29]

    Predictive sampling: Real-time behaviour synthesis with mujoco,

    T. Howell, N. Gileadi, S. Tunyasuvunakool, K. Zakka, T. Erez, and Y . Tassa, “Predictive sampling: Real-time behaviour synthesis with mujoco,”arXiv:2212.00541, 2022

    2022 arXiv

  22. [30]

    Mujoco: A physics engine for model-based control,

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

    2012

  23. [31]

    Dojo: A differentiable simulator for robotics,

    T. A. Howell, S. Le Cleac’h, J. Z. Kolter, M. Schwager, and Z. Manchester, “Dojo: A differentiable simulator for robotics,” arXiv:2203.00806, vol. 9, no. 2, p. 4, 2022

    2022 arXiv

  24. [32]

    Single-level differentiable contact simulation,

    S. Le Cleac’h, M. Schwager, Z. Manchester, V . Sindhwani, P. Florence, and S. Singh, “Single-level differentiable contact simulation,”IEEE Robotics and Automation Letters, vol. 8, no. 7, pp. 4012–4019, 2023

    2023

  25. [33]

    Highly-efficient differentiable simulation for robotics,

    Q. Le Lidec, L. Montaut, Y . de Mont-Marin, F. Schramm, and J. Carpentier, “Highly-efficient differentiable simulation for robotics,”

  26. [34]

    Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems,

    M. Anitescu and F. A. Potra, “Formulating dynamic multi-rigid-body contact problems with friction as solvable linear complementarity problems,”Nonlinear Dynamics, vol. 14, no. 3, pp. 231–247, 1997

    1997

  27. [35]

    Irrotational contact fields,

    A. Castro, X. Han, and J. Masterjohn, “Irrotational contact fields,” IEEE Transactions on Robotics, 2025

    2025

  28. [36]

    From compliant to rigid contact simulation: a unified and efficient approach,

    J. Carpentier, L. Montaut, and Q. L. Lidec, “From compliant to rigid contact simulation: a unified and efficient approach,” arXiv:2405.17020, 2024

    2024 arXiv

  29. [37]

    Convex and analytically-invertible dynamics with con- tacts and constraints: Theory and implementation in mujoco,

    E. Todorov, “Convex and analytically-invertible dynamics with con- tacts and constraints: Theory and implementation in mujoco,” in2014 IEEE ICRA, pp. 6054–6061, IEEE, 2014

    2014

  30. [38]

    Optimization-based control for dynamic legged robots,

    P. M. Wensing, M. Posa, Y . Hu, A. Escande, N. Mansard, and A. Del Prete, “Optimization-based control for dynamic legged robots,” IEEE Transactions on Robotics, vol. 40, pp. 43–63, 2023

    2023

  31. [39]

    On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,

    A. W ¨achter and L. T. Biegler, “On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,”Mathematical programming, vol. 106, no. 1, pp. 25–57, 2006

    2006

  32. [40]

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

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

    2026

  33. [41]

    BONES-SEED

    BONES Studio, “BONES-SEED.” [Online]. Available:https:// huggingface.co/datasets/bones-studio/seed, 2025

    2025

  34. [42]

    mjlab tracking settings

    “mjlab tracking settings.” [Online]. Available:https: //github.com/mujocolab/mjlab/blob/v1.5.3/src/ mjlab/tasks/tracking/tracking_env_cfg.py, 2026

    2026

  35. [43]

    Kinodynamic motion retargeting for humanoid locomotion via multi-contact whole- body trajectory optimization,

    X. Zhang, S. Haener, V . Madabushi, and M. Tucker, “Kinodynamic motion retargeting for humanoid locomotion via multi-contact whole- body trajectory optimization,”arXiv:2603.09956, 2026

    2026 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.