Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

This paper shows that a coherence-based derivative approximation called WASP, which reuses Jacobian-vector products from previous MPC iterations, can replace finite differencing in the MuJoCo MPC library to cut model-derivative computation

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 ↗

WASP derivative reuse speeds up MuJoCo MPC by 1.26–2.08x versus finite differences on selected locomotion tasks, with comparable task costs.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A clean, honest integration of WASP into MJPC with plausible speedups, but the 'drop-in' claim is undercut by per-task tuning and task selection; worth a careful referee. the 4 major comments →

arxiv 2512.21109 v3 pith:SFMVDSAH submitted 2025-12-24 cs.RO

Robust and Efficient MuJoCo-based Model Predictive Control via Web of Affine Spaces Derivatives

Chen Liang , Daniel Rakita This is my paper
classification cs.RO
keywords WASPmodel predictive controlMuJoCofinite differencingapproximate derivativesiLQGlegged locomotioncoherence-based derivative estimation
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 aims to remove the main computational bottleneck in gradient-based model predictive control built on the MuJoCo physics simulator: finite-differencing the model derivatives, whose cost grows with state and control dimension. It argues that Web of Affine Spaces (WASP) derivatives, which reuse Jacobian-vector products from earlier, similar derivative evaluations to fit a new approximate Jacobian, are a drop-in replacement for finite differencing. Across ten locomotion tasks, the paper reports 1.26–2.08× faster model-derivative computation and 1.10–1.45× faster total planning with comparable or better performance, and shows that WASP-based iLQG outperforms the stochastic sampling planners in the MJPC library on these tasks. A sympathetic reader would care because the method makes real-time derivative-based MPC more tractable for high-degree-of-freedom robots without changing the simulator code.

Core claim

The paper's central claim is that WASP derivatives, a recently developed coherence-based method for computing sequences of approximate derivatives, work as an effective drop-in replacement for finite differencing inside MJPC's iLQG planner. WASP treats each derivative as a constrained least-squares problem: one new Jacobian-vector product per update defines an affine subspace that must contain the approximate derivative, while previously computed JVPs are reused to fill the rest of the tangent matrix. Because MPC re-differentiates the same dynamics along nearby trajectories, the derivative sequence is coherent and previous computations stay informative. The paper reports model-derivative spe

What carries the argument

The central object is the Web of Affine Spaces (WASP) derivative estimator: for each new derivative, WASP performs a single finite-difference Jacobian-vector product along one direction, then solves a constrained least-squares problem that keeps the approximate derivative consistent with that one true direction while reusing previously computed Jacobian-vector products for all other directions. The reused directions are stored in a shared tangent matrix and are valid because successive MPC iterations pass through similar states, i.e., the derivative sequence is coherent. A 'frac' parameter controls how many fresh Jacobian-vector products are taken, and 'tol' controls convergence; at full acc

Load-bearing premise

The paper's 'consistently robust performance' claim rests on an evaluation set chosen only because finite-differencing iLQG could already solve those tasks, and on the WASP coherence assumption that successive MPC derivative evaluations are similar enough that reused Jacobian-vector products remain accurate.

What would settle it

Run WASP-based iLQG on a task outside the selectable set, e.g., the contact-rich manipulation tasks the paper itself identifies as failing for both FD and WASP, with default parameters; if the planner fails to stabilize the object while a sampling-based planner succeeds, the claim that WASP outperforms sampling-based planners on contact-rich tasks is false. More narrowly, on a system with rapidly changing contact geometry, measure the error between WASP's approximate Jacobians and true FD Jacobians; large errors would falsify the coherence assumption at the core of the method.

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

If this is right

  • MJPC users can swap finite differencing for WASP with no changes to simulator code and obtain 1.10–1.45× faster planning iterations at similar or better task performance.
  • Derivative-based planners like iLQG become competitive with or faster than stochastic sampling planners on contact-rich locomotion, where sampling planners typically struggle.
  • WASP combines with prior horizon-interpolation acceleration (skipping derivative evaluations at selected time points) to target both high-dimensional and long-horizon MPC costs simultaneously.
  • State-transition accuracy is the binding constraint for iLQG stability, so users should prioritize WASP's state-derivative accuracy over control-derivative accuracy when tuning parameters.
  • Because WASP defaults to orthonormal tangent matrices shared across time steps, its memory overhead is small and scales with state and control dimension rather than horizon length.

Where Pith is reading between the lines

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

  • The speedup should grow with state and action dimensionality, because finite-differencing cost scales linearly with dx+du while WASP spends only a few JVPs per update; humanoid-scale robots and dexterous manipulation may see larger gains than the 1.26–2.08× range reported here.
  • The performance ratios above 1.0 on three tasks hint that the approximation error acts as a regularizer; if confirmed, WASP could be used deliberately as a smoothing device, not just an efficiency shortcut.
  • Because the task suite was filtered to those where FD-iLQG already succeeds, the method's main untested regime is exactly where FD fails; a natural next experiment is to run WASP on the contact-rich manipulation tasks the paper names as failures for both, to see whether coherence-based derivatives extend the solvable task class or merely speed up an already-solvable one.
  • The manual tuning of frac/tol (raise until success) suggests an automation target: adapt the accuracy parameters online from cost improvement or gradient consistency, which would make WASP a truly parameter-free drop-in.
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

4 major / 5 minor

Summary. The paper proposes replacing finite-differencing (FD) model derivatives in the MuJoCo MPC (MJPC) library with WASP coherence-based approximate derivatives. WASP computes approximate Jacobians using a constrained least-squares formulation that reuses information from previous derivative evaluations and requires only a limited number of new Jacobian-vector products per step. The authors integrate WASP into MJPC's C implementation, expose two user-facing tuning parameters (frac and tol), and evaluate the backend on ten locomotion tasks against FD-based iLQG and against MJPC's stochastic sampling-based planners. The central empirical claims are that WASP achieves 1.26–2.08x model-derivative speedups and 1.10–1.45x total planning-time speedups while maintaining a Performance Ratio of at least 0.7, and that WASP-based iLQG outperforms sampling-based planners on contact-rich locomotion tasks.

Significance. If the empirical claims hold, the contribution is practically useful: a roughly 1.3–2x speedup in the dominant derivative-computation step of MJPC, combined with a maintained task performance, would make derivative-based MPC more attractive for real-time control. The paper ships an open-source implementation and integrates WASP directly into the MuJoCo C source, which is a concrete and reproducible engineering contribution. The performance claim is an empirical comparison, not a circular consequence of WASP's derivation. However, the current evaluation does not fully support the 'consistently robust performance' and 'drop-in replacement' wording, primarily because of per-task tuning, lack of error bars, and the comparison protocol against sampling-based planners. These are fixable with additional experiments and reporting, so the manuscript merits revision rather than rejection.

major comments (4)
  1. [§V-B.1 and Table I] The headline claim of 'consistently robust performance' rests on per-task tuning. The procedure states that frac_x and frac_u were 'minimally tuned ahead of time ... until the robot first succeeds', with success defined by Performance Ratio ≥ 0.7. Thus each row in Table I is evaluated at a task-specific parameter setting, and the 1.26–2.08x speedups are not measured at a single default or a principled automated choice. Without a fixed-parameter evaluation (or an automatic selector) across the suite, the data support that WASP can be tuned to be faster on these tasks, not that it is a drop-in replacement that robustly preserves performance across tasks. §VI-A's own limitation statement ('accuracy parameters still require manual tuning') reinforces that this is a load-bearing gap.
  2. [§V-B.1, §V-B.2, Table I] No error bars, seed repetitions, or statistical tests are reported. Every speedup and performance ratio is a point estimate from a single 30-s run. Combined with the per-task tuning, the PR≥0.7 threshold (e.g., Quadruped Walk at exactly 0.70) could be within run-to-run noise. The manuscript should report multiple independent runs, timing variance, and a more principled success threshold.
  3. [§V-C, Table II] The comparison against sampling-based planners conflates failure and speed. The criteria for a red X are not specified, and it is unclear whether average planning time and average cost are computed over identical numbers of iterations and simulation durations when the baseline planner fails. Ratios like 173.1 or 442.5 are not meaningful performance ratios; they are markers of task failure. The claim that WASP-based iLQG achieves up to 4.0x speedups while sampling planners struggle should be supported by separating successful and failed runs and by reporting success rates.
  4. [§V-A and §VI-A] The scoping of the robustness claim is too broad. Criterion 2 admits only tasks already solvable by FD-based derivative MPC, and the limitations section reports persistent failures of both FD and WASP on contact-rich manipulation. The abstract's 'consistently robust performance' should therefore be qualified to the FD-solvable locomotion benchmark; as written, the claim overreaches the evidence.
minor comments (5)
  1. [§V-C.3] Typo: 'Predicative Sampling' should be 'Predictive Sampling'. Also, 'Results for Experiment 1 are shown in Table II' should refer to Experiment 2.
  2. [§III-B and Eq. (4)] The dimensions of the matrices in the WASP optimization are not fully stated. A reader unfamiliar with [12] would benefit from explicit definitions of dimensions and the role of p_max, p_min, p_θ, p_n in the JVP budget.
  3. [§V-B.2] The Performance Ratio definition should be stated more concretely: PR < 1 means WASP has higher average cost; PR = 0.7 means WASP cost is roughly 1.43x FD cost. This would make the threshold choice more transparent.
  4. [§V-A and §V-B.1] The impedance ratio impratio=100 is a non-default MJPC parameter, but it is not listed among the tuned parameters. The paper should make clear that this setting is shared across tasks and justify the choice, since it directly affects contact dynamics.
  5. [§III-A and §VI] The coherence assumption is adopted from [12] without a direct diagnostic. The manuscript would be strengthened by reporting a simple measure of how much model Jacobians change between MPC iterations, or a direct comparison of WASP Jacobian error to FD, since accuracy is currently inferred only from task-level cost.

Circularity Check

0 steps flagged

No significant circularity: the speedup claim is an empirical comparison against FD, and the only self-citation (WASP from [12]) is not load-bearing; per-task parameter tuning and task selection limit scope but do not make the result reduce to its inputs.

full rationale

The paper's central speedup result is an empirical benchmark comparison (WASP vs. FD on MJPC tasks), not a quantity derived by construction from WASP's equations. Equation 4/5 describe WASP's standard approximate-derivative formulation imported from the authors' prior work [12]; the paper does not claim to re-derive or validate WASP mathematically, so the self-citation provides context rather than forcing the outcome. The measured 1.26–2.08x model-derivative speedups and Table I/II comparisons are external to the derivation chain. Two passages warrant flagging but are not circular steps: §V-B.1 tunes (frac_x,frac_u) per task until the robot first succeeds (with PR≥0.7 as the stated success criterion), so the 'consistently robust performance' claim is weaker than a fixed-default-parameter claim; and §V-A criterion 2 only admits FD-solvable tasks, while §VI-A concedes failures on contact-rich manipulation tasks. These are evaluation-scope and generalizability limitations, not a reduction of any prediction to its inputs. No fitted parameter is renamed as a prediction in the derivative derivation, and no uniqueness theorem or result is imported from the authors' own work to forbid alternatives. Overall: no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central empirical claims rest on the authors' prior WASP implementation, a coherence assumption about MPC derivative sequences, and a task-selection/tuning protocol that fixes hyperparameters until success. No new physical entities are introduced.

free parameters (4)
  • frac_x (state accuracy fraction) = task-dependent: 0.3–0.8 (e.g., Quadrotor 0.3; Quadruped Climb/Walk/Trot 0.5; Humanoid 0.8)
    Tuned per task in §V-B.1 by incrementally raising from 0.3 until the robot first succeeds (Performance Ratio >0.7). Directly controls the speed/accuracy tradeoff.
  • frac_u (control accuracy fraction) = task-dependent: 0.3–0.6 (e.g., Quadruped Gallop 0.3; Humanoid 0.6; most 0.3–0.5)
    Tuned in the same per-task search; determines the minimum JVP count for control derivatives.
  • tol (tol_x = tol_u) = 0.5
    WASP stopping thresholds p_theta = p_n fixed at 0.5 for all experiments; hand-set, not swept.
  • impratio (impedance ratio) = 100
    Set to make frictional constraints stiffer for 'high-fidelity contact dynamics'; affects simulator behavior across all tasks.
axioms (4)
  • domain assumption WASP's closed-form solution, caching, and parameter recommendations from Rakita et al. [12] are correct.
    The paper imports the entire WASP machinery from the authors' prior work; §III-B/C and §IV-B defer to [12] for derivation and tuning guidance.
  • domain assumption Temporal coherence: the sequence of model derivatives over MPC iterations changes slowly, so reusing old JVPs preserves accuracy.
    Stated in §III-A as the key insight; no direct measurement of derivative error is provided.
  • domain assumption MuJoCo simulation is a faithful proxy for the robot/contact behavior of interest.
    All experiments are simulated (§V), and §VI-A admits hardware transfer is untested.
  • domain assumption The task-selection criteria in §V-A (FD-solvable, dx+du >= 10, non-instantaneous) produce a representative evaluation of robustness.
    Selection excludes the manipulation regime where both FD and WASP fail; this scope is acknowledged, not proved representative.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust and Efficient MuJoCo-based Model Predictive Control via Web of Affine Spaces Derivatives." pith.science (2026). https://pith.science/paper/SFMVDSAH

@misc{pith2026251221109,
  author       = {Pith},
  title        = {Pith review of: Robust and Efficient MuJoCo-based Model Predictive Control via Web of Affine Spaces Derivatives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SFMVDSAH}},
  note         = {Machine review of arXiv:2512.21109}
}
Share X Bluesky LinkedIn Reddit HN
abstract

MuJoCo is a powerful and efficient physics simulator widely used in robotics. One common way it is applied in practice is through Model Predictive Control (MPC), which uses repeated rollouts of the simulator to optimize future actions and generate responsive control policies in real time. To make this process more accessible, the open source library MuJoCo MPC (MJPC) provides ready-to-use MPC algorithms and implementations built directly on top of the MuJoCo simulator. However, MJPC relies on finite differencing (FD) to compute derivatives through the underlying MuJoCo simulator, which is often a key bottleneck that can make it prohibitively costly for time-sensitive tasks, especially in high-DOF systems or complex scenes. In this paper, we introduce the use of Web of Affine Spaces (WASP) derivatives within MJPC as a drop-in replacement for FD. WASP is a recently developed approach for efficiently computing sequences of accurate derivative approximations. By reusing information from prior, related derivative calculations, WASP accelerates and stabilizes the computation of new derivatives, making it especially well suited for MPC's iterative, fine-grained updates over time. We evaluate WASP across a diverse suite of MJPC tasks spanning multiple robot embodiments. Our results suggest that WASP derivatives are particularly effective in MJPC: it integrates seamlessly across tasks, delivers consistently robust performance, and achieves up to a 2$\mathsf{x}$ speedup compared to an FD backend when used with derivative-based planners, such as iLQG. In addition, WASP-based MPC outperforms MJPC's stochastic sampling-based planners on our evaluation tasks, offering both greater efficiency and reliability. To support adoption and future research, we release an open-source implementation of MJPC with WASP derivatives fully integrated.

Figures

Figures reproduced from arXiv: 2512.21109 by Chen Liang, Daniel Rakita.

Figure 1
Figure 1. Figure 1: Visuals of nine tasks evaluated to demonstrate the efficacy of WASP-based derivatives in MJPC. Relative to finite differences, WASP accelerates the total compute time by up to 2x while maintaining robust performance. of the underlying simulator. For instance, MuJoCo offers XLA and Warp backends to use state-of-the-art AD imple￾mentations. However, in the context of MPC, these exact derivatives often prove … view at source ↗
Figure 2
Figure 2. Figure 2: A screenshot of GUI components for users to activate WASP and control WASP’s accuracy interactively experiments and present our results. All experiments were executed on a desktop computer with an Intel i7 5.4GHz 28-core processor and 32 GB of RAM. A. Evaluation Tasks We selected tasks from the MJPC benchmark suite based on three criteria: 1) The task must satisfy dx + du ≥ 10 to ensure non￾trivial computa… view at source ↗
Figure 3
Figure 3. Figure 3: Parameter sensitivity analysis of WASP on the quadruped trotting task. (a) Cost convergence across parameter settings, showing that task performance remains stable except when state accuracy is severely reduced. (b) Average computation time per derivative evaluation, consistently lower for WASP than FD. (c) Number of forward simulations required, illustrating how WASP adapts resource usage based on accurac… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Turning Stale Gradients into Stable Gradients: Coherent Coordinate Descent with Implicit Landscape Smoothing for Lightweight Zeroth-Order Optimization

    cs.LG 2026-05 unverdicted novelty 7.0

    CoCD converts stale gradients into stable descent directions for zeroth-order optimization through coherent coordinate updates and implicit landscape smoothing from larger finite-difference steps.

  2. Turning Stale Gradients into Stable Gradients: Coherent Coordinate Descent with Implicit Landscape Smoothing for Lightweight Zeroth-Order Optimization

    cs.LG 2026-05 unverdicted novelty 6.0

    CoCD converts stale gradients into stable descent directions through coherent coordinate updates equivalent to BCCD with warm starts, enabling O(1) queries per step plus implicit landscape smoothing from larger finite...

  3. MPC and System Identification with Differentiable Physics: Fluid System and Particle Beam Control

    math.OC 2026-04 unverdicted novelty 6.0

    A framework for simultaneous model predictive control and online parameter estimation is introduced by treating differentiable physics simulators as computational objects for gradient-based joint optimization.

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages · cited by 2 Pith papers

  1. [1]

    On the implementation of automatic differentiation tools.Higher-Order and Symbolic Computation, 21(3): 311–331, 2008

    Christian H Bischof, Paul D Hovland, and Boyana Nor- ris. On the implementation of automatic differentiation tools.Higher-Order and Symbolic Computation, 21(3): 311–331, 2008

    2008

  2. [2]

    Performance of automatic differentiation tools in the dynamic simulation of multibody systems.Advances in Engineering Software, 73:35–44, 2014

    Alfonso Callejo, Sri Hari Krishna Narayanan, Javier Garc´ıa de Jal´on, and Boyana Norris. Performance of automatic differentiation tools in the dynamic simulation of multibody systems.Advances in Engineering Software, 73:35–44, 2014

    2014

  3. [3]

    End-to- end differentiable physics for learning and control

    Filipe de Avila Belbute-Peres, Kevin Smith, Kelsey Allen, Josh Tenenbaum, and J Zico Kolter. End-to- end differentiable physics for learning and control. Advances in neural information processing systems, 31, 2018

    2018

  4. [4]

    A tutorial on the cross-entropy method.Annals of operations research, 134(1):19–67, 2005

    Pieter-Tjerk De Boer, Dirk P Kroese, Shie Mannor, and Reuven Y Rubinstein. A tutorial on the cross-entropy method.Annals of operations research, 134(1):19–67, 2005

    2005

  5. [5]

    Pre- dictive sampling: Real-time behaviour synthesis with mujoco, 2022

    Taylor Howell, Nimrod Gileadi, Saran Tunyasuvu- nakool, Kevin Zakka, Tom Erez, and Yuval Tassa. Pre- dictive sampling: Real-time behaviour synthesis with mujoco, 2022. URLhttps://arxiv.org/abs/ 2212.00541

    Pith/arXiv arXiv 2022

  6. [6]

    Tenenbaum, and Chuang Gan

    Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B. Tenenbaum, and Chuang Gan. Plas- ticinelab: A soft-body manipulation benchmark with differentiable physics, 2021. URLhttps://arxiv. org/abs/2104.03311

    Pith/arXiv arXiv 2021

  7. [7]

    gradsim: Differen- tiable simulation for system identification and visuo- motor control, 2021

    Krishna Murthy Jatavallabhula, Miles Macklin, Florian Golemo, Vikram V oleti, Linda Petrini, Martin Weiss, Breandan Considine, Jerome Parent-Levesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, and Sanja Fidler. gradsim: Differen- tiable simulation for system identification and visuo- motor control, 2021. URLhttps://arxiv...

    Pith/arXiv arXiv 2021

  8. [8]

    ad-trait: A fast and flexible automatic differentiation library in rust, 2025

    Chen Liang, Qian Wang, Andy Xu, and Daniel Rakita. ad-trait: A fast and flexible automatic differentiation library in rust, 2025. URLhttps://arxiv.org/ abs/2504.15976

    Pith/arXiv arXiv 2025

  9. [9]

    A primer on zeroth-order optimization in signal process- ing and machine learning: Principals, recent advances, and applications.IEEE Signal Processing Magazine, 37(5):43–54, 2020

    Sijia Liu, Pin-Yu Chen, Bhavya Kailkhura, Gaoyuan Zhang, Alfred O Hero III, and Pramod K Varshney. A primer on zeroth-order optimization in signal process- ing and machine learning: Principals, recent advances, and applications.IEEE Signal Processing Magazine, 37(5):43–54, 2020

    2020

  10. [10]

    O. L. Mangasarian. Sufficient conditions for the optimal control of nonlinear systems.SIAM Journal on Control, 4(1):139–152, 1966. doi: 10.1137/0304013

    work page doi:10.1137/0304013 1966

  11. [11]

    Margossian

    Charles C. Margossian. A review of automatic differ- entiation and its efficient implementation.WIREs Data Mining and Knowledge Discovery, 9(4), March 2019. ISSN 1942-4795. doi: 10.1002/widm.1305

    doi:10.1002/widm.1305 2019

  12. [12]

    Coherence-based approximate derivatives via web of affine spaces optimization, 2025

    Daniel Rakita, Chen Liang, and Qian Wang. Coherence-based approximate derivatives via web of affine spaces optimization, 2025. URL https://arxiv.org/abs/2504.18790

    Pith/arXiv arXiv 2025

  13. [13]

    H. J. Terry Suh, Max Simchowitz, Kaiqing Zhang, and Russ Tedrake. Do differentiable simulators give better policy gradients?, 2022. URLhttps://arxiv. org/abs/2202.00817

    Pith/arXiv arXiv 2022

  14. [14]

    Syn- thesis and stabilization of complex behaviors through online trajectory optimization

    Yuval Tassa, Tom Erez, and Emanuel Todorov. Syn- thesis and stabilization of complex behaviors through online trajectory optimization. pages 4906–4913, 10

  15. [15]

    Todorov and Weiwei Li

    E. Todorov and Weiwei Li. A generalized iterative lqg method for locally-optimal feedback control of constrained nonlinear stochastic systems. InProceed- ings of the 2005, American Control Conference, 2005., pages 300–306 vol. 1, 2005. doi: 10.1109/ACC.2005. 1469949

    work page doi:10.1109/acc.2005 2005

  16. [16]

    Mu- joco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mu- joco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033, 2012. doi: 10.1109/IROS.2012.6386109

    arXiv 2012

  17. [17]

    Model predictive path integral control: From theory to parallel computation.Journal of Guid- ance, Control, and Dynamics, 40(2):344–357, 2017

    Grady Williams, Andrew Aldrich, and Evangelos A Theodorou. Model predictive path integral control: From theory to parallel computation.Journal of Guid- ance, Control, and Dynamics, 40(2):344–357, 2017

    2017

  18. [18]

    Zhang, Taylor A

    John Z. Zhang, Taylor A. Howell, Zeji Yi, Chaoyi Pan, Guanya Shi, Guannan Qu, Tom Erez, Yuval Tassa, and Zachary Manchester. Whole-body model-predictive control of legged robots with mujoco, 2025. URL https://arxiv.org/abs/2503.04613

    arXiv 2025

  19. [2012]

    doi: 10.1109/IROS

    ISBN 978-1-4673-1737-5. doi: 10.1109/IROS. 2012.6386025

    arXiv 2012

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.