Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Variance-Reduced Model Predictive Path Integral via Quadratic Model Approximation

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that decomposing the MPPI objective into a quadratic model and a low-variance residual, then sampling from the induced Gaussian prior, makes sampling-based control converge faster and with far fewer samples.

desk verdict A clean derivation of a model-guided MPPI, but the main sample-efficiency claim rests on an unequal comparison where the proposed method gets 3x the function evaluation budget. read the letter →

arxiv 2602.03639 v2 pith:V5YZ47ZE submitted 2026-02-03 cs.RO

Fabian Schramm , Franki Nguimatsia Tiofack , Nicolas Perrin-Gilbert , Marc Toussaint , Justin Carpentier This is my paper
classification cs.RO
keywords modelpredictivepathintegralvariancereductionquadraticapproximationimportancesamplingeffectivesamplesizerandomizedsmoothingefficiencytrajectoryoptimization
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

The paper argues that the high variance of Model Predictive Path Integral (MPPI) control comes from weighting samples by the full cost function. Its fix is to split the cost into a quadratic model that captures the local landscape and a residual that captures what the model misses. Sampling is then done from the closed-form Gaussian prior induced by the quadratic model, and only the residual is used for the importance weights. The paper shows this raises the effective sample size, accelerates convergence, and makes performance nearly independent of the number of samples, even down to two, across benchmarks, cart-pole swing-up, and contact-rich manipulation tasks.

What carries the argument

The model–residual decomposition f = m + r with m a quadratic expansion of the cost about the current mean (gradient g and Hessian H). The product of the Gaussian prior with exp(-m/λ) yields a closed-form Gaussian guided prior with covariance (Σ^{-1} + H/λ)^{-1} and a Newton-like mean shift; the residual r is then used for self-normalized importance weights. The Hessian can come from exact derivatives, Gauss-Newton, BFGS, or randomized smoothing, and a variance-control step enforces a lower bound on the guided covariance to prevent collapse.

What would settle it

Run the method on a deliberately difficult objective—say a highly multimodal function with a narrow global basin and broad local basins, initialized far from the global minimum—and record the effective sample size and final distance to the true optimum. If the ESS collapses to near 1 and convergence matches vanilla MPPI, the claimed low-sample robustness fails in that regime.

Watch

Extended reading notes

Core claim

By factoring the Boltzmann update into a model-guided prior and a residual correction, the method injects geometric information into the sampling process: the product of a Gaussian prior and the exponential of a quadratic model is again Gaussian, so the guided prior has closed form. Sampling is then done from this guided Gaussian and only the residual is used in the importance weights. The paper demonstrates that this raises effective sample size (e.g., from 1.1 to 23.4 on a narrow-valley example) and that with accurate curvature the convergence curve collapses onto nearly a single trajectory for sample counts from 2 to 1024, whereas vanilla MPPI and CMA-ES degrade sharply in low-sample regi

Load-bearing premise

The residual between the true objective and the quadratic model must have low variance relative to the full cost; if the quadratic model is not locally accurate (far from a basin, poor smoothing scale, or indefinite Hessian), the guided sampler degenerates exactly like vanilla MPPI.

Editorial extensions

If this is right

  • With reliable second-order information, MPPI can operate effectively in extremely low-sample regimes (N as low as 2), decoupling performance from the sampling budget.
  • Gauss-Newton and BFGS curvature approximations achieve similar high-precision convergence to an exact analytical Hessian, so the method does not require true Hessians.
  • Randomized smoothing extends the framework to non-smooth, black-box dynamics such as contact-rich manipulation, where analytical gradients are unavailable.
  • Higher effective sample size translates to less weight degeneracy, meaning each retained sample carries more information and the optimizer is more stable across seeds.
  • The model construction adds a per-iteration computational overhead, most notably for randomized smoothing, though those evaluations are parallelizable.

Reading between the lines

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

  • If the residual variance is bounded and small relative to the full cost, importance-sampling theory implies the estimator variance scales with residual variance, so any model that captures a large fraction of cost variance should yield gains—the quadratic is one tractable instance, not the only one.
  • The coarse-to-fine smoothing schedule (large sigma first, then small sigma) resembles an annealing strategy; an editor's testable extension is to adapt sigma and temperature online by monitoring residual effective sample size instead of hand-tuning.
  • The same model-residual factorization could be applied to other Boltzmann-based samplers, such as the cross-entropy method, or to control formulations with terminal costs, potentially giving similar low-sample robustness.
  • A concrete risky case: on a heavily multimodal landscape where the quadratic model is initialized far from the global basin, the method may converge to a local minimum just like Newton's method; failure there would not refute the variance-reduction claim but would show the model guidance is only as good as the local model.
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 a variance-reduced Model Predictive Path Integral (MPPI) framework by decomposing the objective into a quadratic model and a residual. It derives a closed-form model-guided Gaussian prior (Eqs. 8–9) and an incremental update (Eqs. 16–17), and supports several ways to obtain curvature information (analytical, Gauss-Newton, quasi-Newton, randomized smoothing). The method is evaluated on static benchmarks, a cart-pole swing-up, and a contact-rich sphere manipulation task, reporting faster convergence, higher effective sample size, and better robustness than vanilla MPPI and CMA-ES.

Significance. The core idea of factoring the Boltzmann update into a model-guided prior and a residual correction is elegant and could make sampling-based control more practical in low-sample regimes. The algebraic derivation is sound, and the framework's agnosticism to the source of curvature information is a practical strength. However, the central variance-reduction claim is not supported by a bound or careful empirical validation of the residual's variance, and the main empirical comparison in the contact task uses an unequal total evaluation budget, which undermines the headline claim of sample efficiency. If these issues are addressed, the method could be a valuable contribution to sample-based control.

major comments (4)
  1. [Section IV-D and V] The contact-task experiment is not run on equal footing: Model-Guided MPPI uses 128 randomized-smoothing evaluations for the quadratic model plus 64 planning samples per iteration, while vanilla MPPI and CMA-ES use only 64 total evaluations. This is a 3× larger total budget. The paper acknowledges the overhead in Section V but still concludes 'greater robustness in low-sample regimes' without qualification. The claim as stated is misleading. The authors should either rerun the comparison with matched total evaluations or clearly separate 'planning samples' from 'model-construction evaluations' and adjust the conclusions accordingly.
  2. [Section III, after Eq. (11)] The central variance-reduction claim is asserted without a bound. The paper states that the residual r_k = f - m_k 'typically' has smaller magnitude and variance than f, but no conditions, quantitative analysis, or empirical verification is provided. If the quadratic model is not locally accurate (far from a basin, wrong σ, ill-regularized indefinite H), the residual is as difficult as f and the guided sampler degenerates exactly like vanilla MPPI. The paper must either provide a formal result linking model quality to the effective sample size or importance-weight variance, or at least an empirical study of the residual variance. This is load-bearing for the 'variance-reduced' label.
  3. [Equations (8), (16), (28)] The Hessian regularization/convexification step is left unspecified. For indefinite H_k, the paper says 'appropriate regularization or convexification is applied' but never defines the procedure. This is a free parameter that changes both the guided mean and covariance, so all experimental results depend on it. Moreover, the variance-control bound in Eq. (28) can be infeasible when κ_max ≥ λ/σ²_target (the denominator becomes nonpositive); the paper does not discuss this case or how it is handled. The method is not fully reproducible without specifying the regularization rule and its parameters.
  4. [Section IV-B] The cart-pole experiments treat analytical derivatives (gradient and Hessian from automatic differentiation) as free. The comparison counts only trajectory samples; the computational cost of computing g_k and H_k is not measured or accounted for. If derivative evaluation has non-negligible cost, the observed 'sample invariance' is not a fair sample-efficiency comparison. The authors should either quantify the wall-clock time or the equivalent function-evaluation cost of the AD oracle, or scope the claim to 'samples of the trajectory cost function' while acknowledging that the derivative oracle has its own computational price.
minor comments (5)
  1. [Section II] The feasible set is written as 'Ω ⊆ R' but the problems are multi-dimensional; it should be 'Ω ⊆ R^n'.
  2. [Equations (12)–(13)] The randomized-smoothing estimators for gradient and Hessian are stated without derivation. A short proof or reference to the standard control-variate form would help the reader verify the scaling with σ² and σ⁴.
  3. [Figure 1 caption] The caption says 'The cross indicates the vanilla MPPI update' but the cross is not labeled in the figure; please add a legend or text marker for clarity.
  4. [Section V] The conclusion repeats the low-sample-regime claim without referencing the computational overhead of randomized smoothing in the contact task, which is partly acknowledged two paragraphs earlier. Please harmonize the conclusion with the stated limitations.
  5. [References] Some references are incomplete (e.g., [30] has only a URL; [7] has a DOI but no page numbers). Please format all references consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: Eq. (6) is an exact factorization of the Boltzmann update, the closed-form guided prior follows from standard Gaussian algebra, and the only self-citations are software/infrastructure references.

full rationale

The derivation chain is self-contained. Starting from the KL-regularized update (Eq. 3) and its Boltzmann solution (Eq. 5), the paper inserts a decomposition f = m_k + r_k and obtains Eq. (6): p*(x) ∝ p_θk(x) exp(-m_k/λ) exp(-r_k/λ). This is an algebraic identity, since m_k + r_k = f; no target result is being assumed. The quadratic instantiation (Eq. 7) is a second-order model chosen for tractability, and the guided parameters (8)-(9) follow by completing the square for the product of a Gaussian and the exponential of a quadratic. These formulas are standard and parameter-free once g_k and H_k are supplied. The final weighting (Eqs. 11 and 21) is ordinary self-normalized importance sampling with the guided Gaussian as proposal and the residual as log-weight. No fitted parameter is renamed as a prediction: g_k and H_k are constructed from analytical derivatives, structural approximations, or randomized smoothing estimates (12)-(13), and the convergence claims are empirical benchmark results rather than algebraic consequences of the construction. The paper's only self-references are to the authors' software libraries Pinocchio [5,6] and robotic [30], and to prior NLP-sampling [31] for task generation; none is invoked to justify the central derivation, to forbid alternatives, or to supply a uniqueness theorem. The statement that the residual 'typically' has smaller variance is a modeling premise, not a derived result; the paper itself flags the reliance on local model validity as a limitation in Sec. V. The unequal evaluation budget in Sec. IV-D (128 randomized-smoothing evaluations in addition to 64 planning samples) is a legitimate benchmarking fairness concern and is acknowledged in Sec. V, but it affects the interpretation of the empirical comparison, not the circularity of the derivation. I find no circular step.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The math is standard, but the practical method relies on several hand-chosen numbers (lambda, sigma, EMA factors, variance floor, Hessian regularization) and on an unproven low-variance residual. These are the real cost of the method.

free parameters (6)
  • lambda (Boltzmann temperature) = 0.1 in Fig. 1; per-task otherwise
    Controls the softness of Boltzmann weights and the guided-covariance scale in Eqs. (8), (16); hand-set, not adapted.
  • sigma (randomized smoothing scale) = 0.2 (Fig. 1); 1.5 then 0.1 (Rastrigin)
    Kernel width in RS estimators Eqs. (12)-(13); determines coarse vs fine quadratic model; hand-scheduled.
  • sigma_target^2 (minimum guided variance)
    Variance floor in Eq. (27); not reported; prevents covariance collapse.
  • alpha_delta, alpha_Sigma (EMA smoothing factors)
    Exponential moving averages in Eqs. (23)-(24); values not reported.
  • Hessian regularization/convexification rule
    Unspecified modification of indefinite H to ensure Sigma_tilde is positive definite; directly shapes the guided Gaussian.
  • Initial covariance Sigma_0/sigma_0^2
    Prior exploration width per benchmark/task; values not systematically reported.
assumptions (6)
  • domain assumption The residual r_k = f - m_k has smaller magnitude and variance than f, so residual-only weighting avoids weight degeneracy.
    Invoked in Sec. III immediately before Eq. (6); no bound is derived; it is the load-bearing premise of the variance-reduction claim.
  • standard math The product of a Gaussian density and the exponential of a quadratic function is a normalizable Gaussian.
    Used to derive closed-form guided prior Eqs. (8)-(9) and (16)-(17).
  • standard math Stein's identity gives unbiased gradient and Hessian estimators for the Gaussian-smoothed objective.
    Underlies the randomized-smoothing estimators Eqs. (12)-(13).
  • domain assumption Finite-M randomized-smoothing estimates remain informative after exponential averaging and regularization.
    The paper uses M=128 auxiliary samples in the contact task; no concentration or error analysis is given.
  • ad hoc to paper An unspecified regularization/convexification of an indefinite H preserves enough curvature information.
    Sec. III-B says 'appropriate regularization or convexification is applied' without specifying the rule; this changes the guided distribution.
  • ad hoc to paper A hand-scheduled coarse-to-fine smoothing scale (sigma=1.5 -> 0.1) is sufficient to escape local minima.
    Used in the Rastrigin example (Sec. IV-C); no adaptation rule or principled schedule is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Variance-Reduced Model Predictive Path Integral via Quadratic Model Approximation." pith.science (2026). https://pith.science/paper/V5YZ47ZE

@misc{pith2026260203639,
  author       = {Pith},
  title        = {Pith review of: Variance-Reduced Model Predictive Path Integral via Quadratic Model Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5YZ47ZE}},
  note         = {Machine review of arXiv:2602.03639}
}
read the original abstract

Sampling-based controllers, such as Model Predictive Path Integral (MPPI) methods, offer substantial flexibility but often suffer from high variance and low sample efficiency. To address these challenges, we introduce a hybrid variance-reduced MPPI framework that integrates a prior model into the sampling process. Our key insight is to decompose the objective function into a known approximate model and a residual term. Since the residual captures only the discrepancy between the model and the objective, it typically exhibits a smaller magnitude and lower variance than the original objective. Although this principle applies to general modeling choices, we demonstrate that adopting a quadratic approximation enables the derivation of a closed-form, model-guided prior that effectively concentrates samples in informative regions. Crucially, the framework is agnostic to the source of geometric information, allowing the quadratic model to be constructed from exact derivatives, structural approximations (e.g., Gauss- or Quasi-Newton), or gradient-free randomized smoothing. We validate the approach on standard optimization benchmarks, a nonlinear, underactuated cart-pole control task, and a contact-rich manipulation problem with non-smooth dynamics. Across these domains, we achieve faster convergence and superior performance in low-sample regimes compared to standard MPPI. These results suggest that the method can make sample-based control strategies more practical in scenarios where obtaining samples is expensive or limited.

Figures

Figures reproduced from arXiv: 2602.03639 by the authors.

Figure 1
Figure 1. Illustration of covariance adaptation via Newton￾like approximation. Model-guided MPPI and vanilla MPPI start at the same state (circle) and use 100 samples from an isotropic Gaussian prior. The cross indicates the vanilla MPPI update. Our method exploits the gradient and Hessian at x¯ 0 to construct a quadratic model. This leads to a guided prior centered at x˜ 0 (rectangle), concentrating the sampling distribution… view at source ↗
Figure 2
Figure 2. Comparison coarse vs fine model. (a) A large smoothing kernel σ = 1.0 suppresses sinusoidal disturbances (grey), yielding a coarse approximation (blue) capturing global geometry. The resulting quadratic model (red) generates a proposal x˜ 1 moving toward the global minimum despite local nonconvexities. (b) Near the optimum, reducing the scale to σ = 0.1 captures the local curvature for final convergence. trajectorie… view at source ↗
Figure 3
Figure 3. Cart-pole swing-up results. Comparison of con￾vergence behavior and variance across sample budgets N ∈ {2, . . . , 1024} and 20 random seeds. B. Performance with analytical gradients We first validate the method in a setting where exact gradient and Hessian information are available. Here, the quadratic model mk accurately captures local curvature for smooth functions, allowing us to isolate and analyze the benefits… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Performance profile on Single-Finger Sphere Ma￾nipulation. Comparison across 200 randomized tasks. vergence. Notably, the GN approximation exploits the least￾squares structure of the cost and closely matches the analytical Hessian. The BFGS variant, despite relying sol…
Figure 6
Figure 6. Figure 6: Single-Finger Sphere Manipulation. Three different starting states (left) and corresponding goal states (right) obtained via NLP sampling illustrate varying degrees of task difficulty. The task requires the actuated finger (blue) to navigate from its start to a target …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Information-Theoretic Adaptive Cooling for Deterministic MPPI via Entropy Feedback

    eess.SY 2026-07 conditional novelty 4.0 of 10

    Entropy of MPPI importance weights as online feedback yields an adaptive cooling schedule that provably drives temperature to zero and, on tested STL planning tasks, converges in fewer iterations than a fixed geometric decay.

Reference graph

Works this paper leans on

44 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    Jacob Abernethy, Chansoo Lee, and Ambuj Tewari.Per- turbation Techniques in Online Learning and Optimiza- tion, pages 233–264. 12 2016. ISBN 9780262337939. doi: 10.7551/mitpress/10761.003.0009

    work page doi:10.7551/mitpress/10761.003.0009 2016

  2. [2]

    CasADi – A software framework for nonlinear optimization and optimal con- trol.Mathematical Programming Computation, 11(1): 1–36, 2019

    Joel A E Andersson, Joris Gillis, Greg Horn, James B Rawlings, and Moritz Diehl. CasADi – A software framework for nonlinear optimization and optimal con- trol.Mathematical Programming Computation, 11(1): 1–36, 2019. doi: 10.1007/s12532-018-0139-4

    doi:10.1007/s12532-018-0139-4 2019

  3. [3]

    Learning with differentiable perturbed optimizers

    Quentin Berthet, Mathieu Blondel, Olivier Teboul, Marco Cuturi, Jean-Philippe Vert, and Francis Bach. Learning with differentiable perturbed optimizers. InAdvances in Neural Information Processing Systems (NeurIPS), pages 9508–9519, 2020

    2020

  4. [4]

    Curtis, and Jorge Nocedal

    L ´eon Bottou, Frank E. Curtis, and Jorge Nocedal. Op- timization methods for large-scale machine learning. ArXiv, abs/1606.04838, 2016

    arXiv 2016

  5. [5]

    Pinocchio: fast forward and inverse dy- namics for poly-articulated systems

    Justin Carpentier, Florian Valenza, Nicolas Mansard, et al. Pinocchio: fast forward and inverse dy- namics for poly-articulated systems. https://stack-of- tasks.github.io/pinocchio, 2015–2021

    2015

  6. [6]

    The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives

    Justin Carpentier, Guilhem Saurel, Gabriele Buondonno, Joseph Mirabel, Florent Lamiraux, Olivier Stasse, and Nicolas Mansard. The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives. InIEEE International Symposium on System Integrations (SII), 2019

    2019

  7. [7]

    Td-cd- mppi: Temporal-difference constraint-discounted model predictive path integral control.IEEE Robotics and Automation Letters, 11(1):498–505, 2026

    Pietro Noah Crestaz, Ludovic De Matteis, Elliot Chane- Sane, Nicolas Mansard, and Andrea Del Prete. Td-cd- mppi: Temporal-difference constraint-discounted model predictive path integral control.IEEE Robotics and Automation Letters, 11(1):498–505, 2026. doi: 10.1109/ LRA.2025.3632612

    arXiv 2026

  8. [8]

    Duchi, Peter L

    John C. Duchi, Peter L. Bartlett, and Martin J. Wain- wright. Randomized smoothing for stochastic optimiza- tion.SIAM Journal on Optimization, 22(2):674–701,

Show all 44 references
  1. [9]

    Brax - a differentiable physics engine for large scale rigid body simulation

    Daniel Freeman, Erik Frey, Anton Raichuk, Sertan Girgin, Igor Mordatch, and Olivier Bachem. Brax - a differentiable physics engine for large scale rigid body simulation. In J. Vanschoren and S. Yeung, editors, Proceedings of the Neural Information Processing Systems Track on D...

  2. [10]

    Monte carlo methods in financial engineering

    Paul Glasserman. Monte carlo methods in financial engineering. 2003

    2003

  3. [11]

    Springer Berlin Heidelberg, Berlin, Heidelberg, 2006

    Nikolaus Hansen.The CMA Evolution Strategy: A Comparing Review, pages 75–102. Springer Berlin Heidelberg, Berlin, Heidelberg, 2006. ISBN 978-3-540- 32494-2. doi: 10.1007/3-540-32494-1 4

    2006 doi

  4. [12]

    Completely derandomized self-adaptation in evolution strategies

    Nikolaus Hansen and Andreas Ostermeier. Completely derandomized self-adaptation in evolution strategies. Evolutionary Computation, 9(2):159–195, 2001

    2001

  5. [13]

    CMA-ES/pycma on Github

    Nikolaus Hansen, Youhei Akimoto, and Petr Baudis. CMA-ES/pycma on Github. Zenodo, DOI:10.5281/zenodo.2559634, February 2019

    2019 doi

  6. [14]

    Gauss-newton accelerated mppi control, 2026

    Hannes Homburger, Katrin Baumg ¨artner, Moritz Diehl, and Johannes Reuter. Gauss-newton accelerated mppi control, 2026

    2026

  7. [15]

    Predictive sampling: Real-time behaviour synthesis with mujoco,

    Taylor Howell, Nimrod Gileadi, Saran Tunyasuvunakool, Kevin Zakka, Tom Erez, and Yuval Tassa. Predictive sampling: Real-time behaviour synthesis with mujoco,

  8. [16]

    Hilbert J. Kappen. Linear theory for control of non- linear stochastic systems.Physical review letters, 95 20:200201, 2004. URL https://api.semanticscholar.org/ CorpusID:5774276

    2004

  9. [17]

    Muhammad Kazim, JunGee Hong, Min-Gyeom Kim, and Kwang-Ki K. Kim. Recent advances in path in- tegral control for trajectory optimization: An overview in theoretical and algorithmic perspectives.An- nual Reviews in Control, 57:100931, 2024. ISSN 1367-5788. doi: https://doi.org/1...

    2024 doi

  10. [18]

    Lasserre

    Jean B. Lasserre. Global Optimization with Polynomials and the Problem of Moments.SIAM Journal on Opti- mization, 11(3):796–817, 2001

    2001

  11. [19]

    Differentiable simula- tion for physical system identification.IEEE Robotics Autom

    Quentin Le Lidec, Igor Kalevatykh, Ivan Laptev, Cordelia Schmid, and Justin Carpentier. Differentiable simula- tion for physical system identification.IEEE Robotics Autom. Lett., 6(2):3413–3420, 2021. doi: 10.1109/LRA. 2021.3062323. URL https://doi.org/10.1109/LRA.2021. 3062323

    2021

  12. [20]

    Effective sample size for importance sampling based on discrepancy measures.Signal Processing, 131:386–401,

    Luca Martino, V ´ıctor Elvira, and Francisco Louzada. Effective sample size for importance sampling based on discrepancy measures.Signal Processing, 131:386–401,

  13. [21]

    Jorge Nocedal and Stephen J. Wright. Numerical optimization. InFundamental Statistical Inference,

  14. [22]

    Pablo A. Parrilo. Semidefinite programming relaxations for semialgebraic problems.Mathematical Programming, 96(2):293–320, 2003

    2003

  15. [23]

    Policy gradient methods for robotics.2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2219–2225, 2006

    Jan Peters and Stefan Schaal. Policy gradient methods for robotics.2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2219–2225, 2006. URL https://api.semanticscholar.org/CorpusID:613022

    2006

  16. [24]

    Rubinstein

    Reuven Y . Rubinstein. Simulation and the monte carlo method. InWiley series in probability and mathematical statistics, 1981. URL https://api.semanticscholar.org/ CorpusID:39230485

    1981

  17. [25]

    Rubinstein

    Reuven Y . Rubinstein. The cross-entropy method for combinatorial and continuous optimization.Methodol- ogy And Computing In Applied Probability, 1:127–190,

  18. [26]

    Finding global minima via kernel approxima- tions: Finding global minima via kernel approximations

    Alessandro Rudi, Ulysse Marteau-Ferey, and Francis Bach. Finding global minima via kernel approxima- tions: Finding global minima via kernel approximations. Math. Program., 209(1):703–784, April 2024. ISSN 0025-5610. doi: 10.1007/s10107-024-02081-4. URL https://doi.org/10.1007...

    2024 doi

  19. [27]

    Theodorou and Emanuel Todorov

    Evangelos A. Theodorou and Emanuel Todorov. Relative entropy and free energy dualities: Connections to path integral and kl control.2012 IEEE 51st IEEE Conference on Decision and Control (CDC), pages 1466–1473, 2012

    2012

  20. [28]

    Linearly-solvable markov decision problems

    Emanuel Todorov. Linearly-solvable markov decision problems. InNeural Information Processing Systems, 2006

    2006

  21. [29]

    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. IEEE, 2012. doi: 10.1109/IROS.2012.6386109

    2012

  22. [30]

    Robotic: Robotic control interface & manipulation planning library

    Marc Toussaint. Robotic: Robotic control interface & manipulation planning library. https://github.com/ MarcToussaint/robotic

  23. [31]

    Braun, and Joaquim Ortiz de Haro

    Marc Toussaint, Cornelius V . Braun, and Joaquim Ortiz de Haro. NLP Sampling: Combining MCMC and NLP Methods for Diverse Constrained Sampling.ArXiv, abs/2407.03035, 2024

    2024

  24. [32]

    Giulio Turrisi, Valerio Modugno, Lorenzo Amatucci, Dimitrios Kanoulas, and Claudio Semini. On the Ben- efits of GPU Sample-Based Stochastic Predictive Con- trollers for Legged Locomotion.2024 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), pages 1...

    2024

  25. [33]

    Karen Liu

    Keenon Werling, Dalton Omens, Jeongseok Lee, Ioan- nis Exarchos, and C. Karen Liu. Fast and feature- complete differentiable physics engine for articulated rigid bodies with contact constraints. In Dylan A. Shell, Marc Toussaint, and M. Ani Hsieh, editors,Robotics: Science and...

    2021 doi

  26. [34]

    Theodorou

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

    2017

  27. [35]

    Rehg, and Evangelos A

    Grady Williams, Paul Drews, Brian Goldfain, James M. Rehg, and Evangelos A. Theodorou. Information- theoretic model predictive control: Theory and appli- cations to autonomous driving.IEEE Transactions on Robotics, 34:1603–1622, 2017. URL https://api. semanticscholar.org/Corpu...

    2017

  28. [36]

    Williams

    Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning.Ma- chine Learning, 8:229–256, 2004

    2004

  29. [37]

    Full-order sampling-based mpc for torque- level locomotion control via diffusion-style annealing

    Haoru Xue, Chaoyi Pan, Zeji Yi, Guannan Qu, and Guanya Shi. Full-order sampling-based mpc for torque- level locomotion control via diffusion-style annealing. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4974–4981, 2025. doi: 10.1109/ICRA55743.2...

    2025

  30. [38]

    Pa- mppi: Perception-aware model predictive path integral control for quadrotor navigation in unknown environ- ments.ArXiv, abs/2509.14978, 2025

    Yifan Zhai, Rudolf Reiter, and Davide Scaramuzza. Pa- mppi: Perception-aware model predictive path integral control for quadrotor navigation in unknown environ- ments.ArXiv, abs/2509.14978, 2025. URL https://api. semanticscholar.org/CorpusID:281394267

    2025

  31. [1999]

    URL https://api.semanticscholar.org/CorpusID: 18838459

  32. [2012]

    doi: 10.1137/110831659

    doi

  33. [2017]

    doi: https://doi.org/10.1016/j

    ISSN 0165-1684. doi: https://doi.org/10.1016/j. sigpro.2016.08.025

    2016 doi

  34. [2018]

    URL https://api.semanticscholar.org/CorpusID: 189864167

  35. [2021]

    neurips.cc/paper files/paper/2021/file/ d1f491a404d6854880943e5c3cd9ca25-Paper-round1

    URL https://datasets-benchmarks-proceedings. neurips.cc/paper files/paper/2021/file/ d1f491a404d6854880943e5c3cd9ca25-Paper-round1. pdf

    2021

  36. [2022]

    URL https://arxiv.org/abs/2212.00541

    arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.