Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Sampling-based robot control can run much faster by rolling out a learned linear Koopman model instead of the true nonlinear dynamics.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-15 14:32 UTC pith:5JKMQDT5

load-bearing objection Solid engineering paper: linear deep-Koopman rollouts make MPPI faster with near-parity closed-loop performance, including clean quadruped hardware numbers. the 3 major comments →

arxiv 2603.05385 v2 pith:5JKMQDT5 submitted 2026-03-05 cs.RO cs.SYeess.SY

Accelerating Sampling-Based Control via Learned Linear Koopman Dynamics

Wenjian Hao , Yuxuan Fang , Zehui Lu , Shaoshuai Mou This is my paper
classification cs.RO cs.SYeess.SY
keywords model predictive path integraldeep Koopman operatorsampling-based controllearned dynamicsreal-time roboticsquadruped trackingGPU acceleration
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.

Classic model predictive path integral (MPPI) control is powerful for nonlinear robots because it samples many noisy trajectories and reweights them, but every sample must be rolled forward with an expensive nonlinear simulator. This paper claims that a deep Koopman operator model—learned only from interaction data—can replace that nonlinear rollout with cheap linear matrix multiplications in a lifted space, while still producing nearly the same closed-loop behavior. The resulting controller, MPPI-DK, is shown on a pendulum, a surface vehicle, and a Unitree Go1 quadruped to match true-dynamics MPPI accuracy at substantially lower cost, especially when the rollouts are parallelized on a GPU. If the claim holds, practitioners can keep the flexibility of sampling-based control without needing an analytic model or paying the full nonlinear simulation tax at every control cycle.

Core claim

Replacing the nonlinear dynamics inside MPPI trajectory sampling with a finite-dimensional linear deep Koopman model learned from data yields closed-loop performance close to classic MPPI while cutting the per-step computation enough for real-time robotic control.

What carries the argument

MPPI-DK: after the deep lifting map g is evaluated once at the current state, subsequent states of each sample are advanced solely by the linear operators A* and B* (then decoded by C*), so the expensive neural network is never re-evaluated during the horizon.

Load-bearing premise

The true nonlinear plant must be close enough to a finite-dimensional linear Koopman model that the costs of the sampled trajectories remain reliable for closed-loop decisions.

What would settle it

On the same quadruped tracking task, measure whether the final tracking error or success rate of MPPI-DK degrades sharply relative to true-dynamics MPPI when the robot is started from states far outside the training distribution used to learn the Koopman model.

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

If this is right

  • MPPI-style controllers can be deployed on platforms whose analytic dynamics are unavailable or too costly to simulate online.
  • GPU parallel sampling of linear Koopman rollouts can undercut both classic nonlinear MPPI and linear MPC that uses the same learned model.
  • Only interaction data (even random or mixed expert/random) are required; no hand-derived equations of motion are needed at control time.
  • Control smoothness and final tracking error can remain comparable to true-dynamics MPPI when the learned lift is adequate.

Where Pith is reading between the lines

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

  • The same lifted-linear rollout trick could accelerate other sampling-based planners (e.g., CEM or information-theoretic variants) that currently bottleneck on nonlinear simulation.
  • If the residual Koopman error is state-dependent, an online residual corrector or periodic re-learning of A*, B*, C* may be needed to keep long-horizon costs unbiased.
  • Because the lift is evaluated only once per sample, the method’s speedup grows with horizon length and with the complexity of g, favoring deep lifts over shallow ones.

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

3 major / 5 minor

Summary. The paper proposes MPPI-DK, an MPPI controller that replaces nonlinear dynamics rollouts with a learned finite-dimensional linear deep Koopman operator (DKO) model of the form x(t+1)=C*(A*g(x,θ*)+B*v). DKO parameters are identified from interaction data via the multi-variable loss and alternating update of Eqs. (6)–(9). Trajectory sampling then propagates the lifted state by matrix multiplications (Algorithm 1, Remark 1), avoiding repeated DNN evaluations of g. The method is evaluated on inverted-pendulum swing-up (ablations of network width, lift dimension, and expert data), surface-vehicle navigation (CPU/GPU timing vs. true-dynamics MPPI and MPC-DK), and Unitree Go1 reference tracking (GPU, 10 initial states). Reported results show closed-loop performance close to true-dynamics MPPI with lower per-step wall-clock time (e.g., 8.8 ms vs. 11.7 ms on the quadruped).

Significance. If the empirical trade-off holds, the work supplies a practical route to higher-frequency sampling-based control on platforms whose true models are expensive to evaluate, without requiring analytical dynamics. Strengths include a clear algorithmic statement (Algorithm 1), complementary evaluation across simulation ablations, multi-method timing (Table II), and hardware validation with success rate, tracking error, and smoothness (Table III, Fig. 4). The contribution is primarily engineering rather than theoretical; its value lies in demonstrating that linear Koopman surrogates can accelerate MPPI while preserving usable closed-loop performance.

major comments (3)
  1. The central claim rests on the finite-dimensional linear Koopman representation (Eq. 2 / §II-A). The manuscript validates this only through closed-loop cost/tracking comparisons. Open-loop multi-step prediction error of the learned DKO (or residual statistics over the state space used by the sampler) is not reported for any of the three systems. Without those diagnostics it is hard to judge how much of the residual tracking gap (e.g., higher average track error in Table III) is due to model bias versus sampling variance, and whether the approximation remains adequate outside the demonstrated operating regions.
  2. Data collection for DKO learning is described as uniform random inputs (§III, §IV-B, §V). For the surface vehicle and especially the 8-D task-oriented quadruped state, coverage of the relevant state–input manifold is not quantified (e.g., no coverage metric, no hold-out prediction error, no sensitivity to data volume). Because MPPI samples around the current nominal sequence, systematic under-sampling of high-cost or near-goal regions could bias the learned linear operators and undermine the “close to true dynamics” claim under distribution shift.
  3. Hardware comparison (Table III) reports a modest absolute speedup (8.8 ms vs. 11.7 ms) on GPU with N=900, T=40. The abstract and conclusion frame the result as “substantially reducing computational cost” and “enabling efficient real-time control.” The manuscript should either (i) quantify the fraction of time spent in dynamics evaluation versus cost evaluation / weighting so that the source of the gain is clear, or (ii) temper the language to match the observed ~25 % reduction, and discuss how the gain scales with larger N or more expensive true models.
minor comments (5)
  1. Algorithm 1 line 8 writes the control-cost term as γ_u T_{s-1} Σ^{-1} ε; the classic MPPI running cost in Eq. (11) uses a different quadratic form. Align the notation with Eq. (11) or state the simplification explicitly.
  2. Table I lists only five architectures; the surface-vehicle and quadruped experiments use 256-neuron layers and lift dimensions 8 and 10 that are not in the table. State the full architecture used for each experiment in one place.
  3. Typographical issues: “vehivle” (§IV-B), “EXPERIMENTSRESULTS” (section heading), missing spaces after periods in several places, and inconsistent use of δu vs. ε for the noise samples.
  4. Fig. 3 caption and Table II report means over 4 trials; Fig. 1–2 over 5 trials. State the random seeds or training variability more clearly so that the shaded regions are reproducible.
  5. Related-work discussion of Koopman-MPC (e.g., Korda & Mezić) is present, but a short comparison of when linear Koopman + MPPI is preferable to linear Koopman + quadratic MPC would help position the contribution.

Circularity Check

0 steps flagged

No significant circularity; the MPPI-DK claim is an empirical engineering trade-off validated against independent true-dynamics baselines and hardware.

full rationale

The paper's derivation chain is self-contained and non-circular. It assumes a finite-dimensional linear Koopman representation (Eq. 2) as a modeling premise, learns the DKO parameters A*,B*,C*, heta* from interaction data via the multi-variable loss Lf and alternating updates (Eqs. 6–9, citing the authors' prior DKO framework only for the learning algorithm and its convergence analysis), then substitutes the resulting linear lifted propagation into the standard MPPI Monte-Carlo rollout and weighted update (Algorithm 1 and Remark 1). The central performance claim—closed-loop costs and tracking errors close to classic MPPI with true nonlinear dynamics, at substantially lower per-step compute—is not derived from those equations or from any self-citation uniqueness theorem; it is measured directly against an independent true-dynamics MPPI baseline (Figs. 1–3, Tables II–III) and against hardware ground truth on the Unitree Go1 (Fig. 4, Table III). No quantity is fitted and then re-presented as a prediction, no result is forced by definition, and residual model error is openly visible in the slightly higher average tracking error. Self-citations supply reusable learning machinery but are not load-bearing for the control claim, which stands or falls on the reported empirical numbers.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central empirical claim rests on the finite-dimensional Koopman representation assumption, on standard full-rank conditions for the least-squares solves inside DKO training, and on a collection of hand-chosen hyper-parameters (network width, lift dimension, MPPI temperature, sample counts). No new physical entities are postulated.

free parameters (3)
  • lift dimension r and DNN hidden widths
    Chosen by ablation (Table I and §IV-A); different values change both approximation quality and compute cost.
  • MPPI horizon T, sample count N, noise covariance Σ, temperature λ
    Hand-tuned per task (T=20/40, N=600–2000, etc.); directly affect both performance and reported wall-clock times.
  • DKO learning rate α and Adam optimizer settings
    Standard but unreported exact schedule; affect the quality of the linear matrices A*,B*,C* used for all subsequent control.
axioms (3)
  • domain assumption The true nonlinear dynamics admit a finite-dimensional linear Koopman representation of the form given in Eq. (2).
    Stated as an assumption in §II-A; without it the linear rollouts are systematically biased.
  • standard math The data matrices Ḡk and [G′k U′]′ have full row rank so that the least-squares solutions (9) are well-defined.
    Invoked explicitly before Eq. (9); standard EDMD/DKO technical condition.
  • domain assumption The lifting map g(·,θ) is Lipschitz continuous.
    Stated in §II-A to guarantee well-posedness of the lifted dynamics.

pith-pipeline@v1.1.0-grok45 · 15404 in / 2473 out tokens · 30337 ms · 2026-07-15T14:32:36.523006+00:00 · methodology

0 comments
read the original abstract

This paper presents an efficient model predictive path integral (MPPI) control framework for systems with complex nonlinear dynamics. To improve the computational efficiency of classic MPPI while preserving control performance, we replace the nonlinear dynamics used for trajectory propagation with a learned linear deep Koopman operator (DKO) model, enabling faster rollout and more efficient trajectory sampling. The DKO dynamics are learned directly from interaction data, eliminating the need for analytical system models. The resulting controller, termed MPPI-DK, is evaluated in simulation on pendulum balancing and surface vehicle navigation tasks, and validated on hardware through reference-tracking experiments on a quadruped robot. Experimental results demonstrate that MPPI-DK achieves control performance close to MPPI with true dynamics while substantially reducing computational cost, enabling efficient real-time control on robotic platforms.

Figures

Figures reproduced from arXiv: 2603.05385 by Shaoshuai Mou, Wenjian Hao, Yuxuan Fang, Zehui Lu.

Figure 1
Figure 1. Figure 1: Average trajectory cost over 5 independent trials for different lifting function architectures and training datasets. 0 50 100 150 200 Time step t 0 2 4 6 8 10 12 14 State dependent cost r = 4, with expert data r = 6, with expert data r = 8, with expert data r = 4, no expert data r = 6, no expert data r = 8, no expert data MPPI-true dynamics [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average trajectory cost over 5 independent trials for different lifting dimensions and training datasets. number of neurons results in faster convergence of the MPPI￾DK controller to the goal state, with control inputs becoming [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Tracking error over time. The solid line represents [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative trajectories generated by MPPI-DK [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 1 linked inside Pith

  1. [1]

    Nonlinear model predictive control

    Lars Gr ¨une and J ¨urgen Pannek. Nonlinear model predictive control. InNonlinear model predictive control: Theory and algorithms, pages 45–69. Springer, 2016

    2016

  2. [2]

    Model predictive control: theory, computation, and design.(No Title), 2020

    James B Rawlings, David Q Mayne, and Moritz M Diehl. Model predictive control: theory, computation, and design.(No Title), 2020

    2020

  3. [3]

    Aggressive driving with model predictive path integral control

    Grady Williams, Paul Drews, Brian Goldfain, James M Rehg, and Evangelos A Theodorou. Aggressive driving with model predictive path integral control. In2016 IEEE international conference on robotics and automation (ICRA), pages 1433–1440. IEEE, 2016

    2016

  4. [4]

    Model predictive path integral control for agile unmanned aerial vehicles

    Michal Minaˇr´ık, Robert Pˇeniˇcka, V ojtˇech V on´asek, and Martin Saska. Model predictive path integral control for agile unmanned aerial vehicles. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13144–13151. IEEE, 2024

    2024

  5. [5]

    Real-time whole-body control of legged robots with model-predictive path integral control

    Juan Alvarez-Padilla, John Z Zhang, Sofia Kwok, John M Dolan, and Zachary Manchester. Real-time whole-body control of legged robots with model-predictive path integral control. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 14721–14727. IEEE, 2025

    2025

  6. [6]

    Information theoretic mpc for model-based reinforcement learning

    Grady Williams, Nolan Wagener, Brian Goldfain, Paul Drews, James M Rehg, Byron Boots, and Evangelos A Theodorou. Information theoretic mpc for model-based reinforcement learning. In2017 IEEE international conference on robotics and automation (ICRA), pages 1714–1721. IEEE, 2017

    2017

  7. [7]

    On applications of the spectral theory of the koopman operator in dynamical systems and control theory

    Igor Mezi ´c. On applications of the spectral theory of the koopman operator in dynamical systems and control theory. In2015 54th IEEE Conference on Decision and Control (CDC), pages 7034–7041. IEEE, 2015

    2015

  8. [8]

    Generalizing koopman theory to allow for inputs and control.SIAM Journal on Applied Dynamical Systems, 17(1):909–930, 2018

    Joshua L Proctor, Steven L Brunton, and J Nathan Kutz. Generalizing koopman theory to allow for inputs and control.SIAM Journal on Applied Dynamical Systems, 17(1):909–930, 2018

    2018

  9. [9]

    Linear identification of nonlinear systems: A lifting technique based on the koopman operator

    Alexandre Mauroy and Jorge Goncalves. Linear identification of nonlinear systems: A lifting technique based on the koopman operator. In2016 IEEE 55th Conference on Decision and Control (CDC), pages 6500–6505. IEEE, 2016

    2016

  10. [10]

    On convergence of extended dynamic mode decomposition to the koopman operator.Journal of Nonlinear Science, 28(2):687–710, 2018

    Milan Korda and Igor Mezi ´c. On convergence of extended dynamic mode decomposition to the koopman operator.Journal of Nonlinear Science, 28(2):687–710, 2018

    2018

  11. [11]

    Data-driven discovery of koopman eigenfunctions using deep learning.Bulletin of the American Physical Society, 2017

    Bethany Lusch, Steven L Brunton, and J Nathan Kutz. Data-driven discovery of koopman eigenfunctions using deep learning.Bulletin of the American Physical Society, 2017

    2017

  12. [12]

    Deep learning of koopman representation for control

    Yiqiang Han, Wenjian Hao, and Umesh Vaidya. Deep learning of koopman representation for control. In2020 59th IEEE Conference on Decision and Control (CDC), pages 1890–1895. IEEE, 2020

    2020

  13. [13]

    Deep koopman learning of nonlinear time-varying systems.Automatica, 159:111372, 2024

    Wenjian Hao, Bowen Huang, Wei Pan, Di Wu, and Shaoshuai Mou. Deep koopman learning of nonlinear time-varying systems.Automatica, 159:111372, 2024

    2024

  14. [14]

    Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control.Automatica, 93:149–160, 2018

    Milan Korda and Igor Mezi ´c. Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control.Automatica, 93:149–160, 2018

    2018

  15. [15]

    Heredia, and Shaoshuai Mou

    Wenjian Hao, Paulo C. Heredia, and Shaoshuai Mou. Optimal control of nonlinear systems with unknown dynamics, 2025

    2025

  16. [16]

    Model predictive path integral control: From theory to parallel computation

    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(2):344–357, 2017

    2017

  17. [17]

    Smoothing and differentiation of data by simplified least squares procedures.Analytical chemistry, 36(8):1627–1639, 1964

    Abraham Savitzky and Marcel JE Golay. Smoothing and differentiation of data by simplified least squares procedures.Analytical chemistry, 36(8):1627–1639, 1964

    1964

  18. [18]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  19. [19]

    Unitree go1

    Unitree Robotics. Unitree go1. https://shop.unitree.com,

  20. [20]

    Accessed: YYYY-MM-DD

  21. [21]

    Centroidal dynamics of a humanoid robot.Autonomous robots, 35(2):161–176, 2013

    David E Orin, Ambarish Goswami, and Sung-Hee Lee. Centroidal dynamics of a humanoid robot.Autonomous robots, 35(2):161–176, 2013

    2013

  22. [22]

    Whole-body motion planning with centroidal dynamics and full kinematics

    Hongkai Dai, Andr ´es Valenzuela, and Russ Tedrake. Whole-body motion planning with centroidal dynamics and full kinematics. In 2014 IEEE-RAS International Conference on Humanoid Robots, pages 295–302. IEEE, 2014

    2014