Pith. sign in

REVIEW 3 major objections 5 minor 47 references

MetaKoopman turns nonlinear dynamics into linear Bayesian updates that adapt online to new driving conditions.

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 · deepseek-v4-flash

2026-08-01 00:04 UTC pith:AO3RN525

load-bearing objection The meta-learned conjugate prior over Koopman operators is a genuinely new idea with a strong empirical story, but the paper's core Bayesian update equations are internally inconsistent and the 'closed-form multi-step' claim is overstated. the 3 major comments →

arxiv 2607.26345 v1 pith:AO3RN525 submitted 2026-07-28 cs.LG cs.ROcs.SYeess.SY

MetaKoopman: Bayesian Meta-Learning of Koopman Operators for Modeling Structured Dynamics under Distribution Shifts

Mahmoud Selim , Sriharsha Bhat , Karl H. Johansson This is my paper
classification cs.LG cs.ROcs.SYeess.SY
keywords Koopman operatorBayesian meta-learningMatrix Normal-Inverse Wishartdistribution shiftuncertainty quantificationmotion planningautonomous truckclosed-form adaptation
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.

MetaKoopman claims that nonlinear dynamics under distribution shift can be modeled by learning a probability distribution over a linear Koopman operator, updated in closed form from short recent trajectory segments. It argues that this yields fast, uncertainty-aware adaptation that beats prior methods in multi-step prediction accuracy and calibration, and that it works on a full-scale 37.5-ton truck in snow, ice, and mixed-friction conditions. If right, it means a lightweight, analytic Bayesian update—not gradient inner loops—can keep a learned dynamics model reliable when the environment changes.

Core claim

MetaKoopman places a Matrix Normal-Inverse Wishart prior over the finite-dimensional Koopman operator K and meta-learns its parameters across a distribution of environment-specific trajectory tasks. At test time, a few recent state-action pairs are sufficient to update the posterior over K in closed form, and the posterior predictive distribution over future states is also closed-form (a Student-t approximated by a Gaussian). A learned scalar tempering factor β rescales prior precision to balance responsiveness and calibration. The paper reports consistently lower multi-step prediction error than adaptive, probabilistic, and structured baselines across simulated shifts (terrain slope, gravit

What carries the argument

The Matrix Normal-Inverse Wishart (MNIW) prior over the Koopman operator is the central object: it is the conjugate prior for the Gaussian latent-transition likelihood, so the posterior update is a closed-form rank-1-ish matrix update and the posterior predictive over future trajectories can be computed by recursively matching the first two moments. This lets adaptation happen without any gradient inner loop, making online forecasting and planning fast enough for real-time control.

Load-bearing premise

The encoder Gθ is assumed to stay valid under distribution shift, so that all adaptation can be captured by updating the linear Koopman operator K; if the shift changes the latent representation itself, updating K alone cannot compensate.

What would settle it

Take the truck dataset or the HalfCheetah-Slope environment and apply a shift that changes the latent representation—for example, adding severe sensor noise or changing the state observation space at test time—and measure whether prediction error grows monotonically with the shift magnitude; if the fixed-encoder assumption holds, error should stay bounded, otherwise the closed-form Bayesian update will not recover.

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

If this is right

  • If MetaKoopman's central claim holds, closed-form Bayesian adaptation can replace iterative gradient-based inner loops in online dynamics modeling, reducing latency by an order of magnitude in trajectory rollouts.
  • The framework offers a template for combining structured linear latent models with conjugate Bayesian priors, potentially extending beyond vehicle dynamics to any control task where the encoder remains stable but the operator drifts.
  • The reported real-world results suggest that uncertainty-aware online adaptation can change a planner's safety-critical decisions (e.g., choosing a lane change over braking on ice), which is a direct consequence of better calibrated predictive uncertainty.
  • The meta-learned tempering factor β provides a principled, data-driven way to control how strongly prior knowledge constrains online adaptation, without manual tuning per environment.

Where Pith is reading between the lines

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

  • The paper's central assumption—that distribution shift affects the dynamics operator but not the learned latent representation—implies that shifts which change the appropriate observables (e.g., a completely new sensor modality or qualitatively different physics) would not be handled by updating K alone; the closed-form update would then be updating the wrong object.
  • A testable extension would be to run MetaKoopman on an environment shift that is representation-level (e.g., changing the observation noise model or the state definition) to see where the fixed-encoder assumption breaks.
  • The CLAS variational action encoder, by letting the planner sample directly in latent action space, could be reused for other Koopman-based motion planners, decoupling action sampling from the cost of the encoder network.
  • The moment-matched multi-step predictive rule suggests a direct way to add uncertainty-aware cost terms in planning: the planner can use the posterior predictive covariance as an explicit risk penalty, which the paper uses implicitly but does not formalize as a general cost-shaping principle.

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 MetaKoopman, a Bayesian meta-learning framework for forecasting nonlinear dynamical systems under distribution shift. The core idea is to learn a finite-dimensional Koopman operator in a latent space and place a Matrix Normal-Inverse Wishart (MNIW) prior over the operator and process noise covariance. The prior is meta-learned across tasks, and at test time a conjugate Bayesian update from a short trajectory segment yields a posterior over the Koopman operator. The paper derives a closed-form one-step posterior predictive and a moment-matched Gaussian recursion for multi-step forecasts, and integrates this into a sampling-based planner via a variational action encoder. Experiments include MuJoCo benchmarks with varying slope, gravity, friction, joint failures, and damping, plus full-scale truck field tests in winter conditions. The paper claims consistent improvements in multi-step prediction accuracy, uncertainty calibration, robustness, and real-time planning feasibility over several baselines.

Significance. If correct, the contribution is valuable: it combines meta-learning with conjugate Bayesian inference over Koopman operators, giving fast, uncertainty-aware adaptation without inner-loop optimization, and it evaluates the method on an unusually demanding real-world platform. The empirical evaluation is broad and includes meaningful ablations and a real truck-and-trailer system. However, the central technical derivation contains an internal inconsistency in the definition of the posterior scale matrix that affects every predictive equation and the meta-learning objective. The abstract also overstates the exactness of the multi-step posterior predictive. These issues are fixable but must be resolved before the main claims can be accepted.

major comments (3)
  1. [Sec. 4.2, Eq. (11), and Appendix A, Eqs. (19)-(26)] Eq. (11) states that the posterior column parameter is \hat V = S_zz = \tilde Z \tilde Z^T + \tilde V. Under the stated prior K|Σ ~ MN(M̆, Σ, V̆), the standard conjugate posterior column covariance is (\tilde Z \tilde Z^T + V̆^{-1})^{-1}; if V̆ is instead interpreted as prior precision, then S_zz is the posterior precision, not the covariance. In neither case is \hat V = S_zz the posterior covariance required by Eq. (13). Appendix A is internally consistent only if \hat V is the covariance: it uses S_aa = z z^T + \hat V^{-1} and obtains the predictive scale 1 + z^T \hat V z. With Eq. (11) as written, the predictive covariance in Eq. (13) grows with the data sum S_zz rather than contracting after adaptation, and the meta-learning NLL in Eq. (15) is numerically incorrect if implemented literally. Please correct the definitions (e.g., \hat V = S_zz^{-1} with S_zz = \tilde Z \tilde Z^T + V̆^
  2. [Abstract and Sec. 4.2, Lemma 4.2] The abstract and contribution list claim a 'closed-form posterior predictive distribution over future state trajectories.' This is not what is derived. Lemma 4.2 gives a one-step Gaussian approximation, and multi-step forecasts are obtained by recursively moment-matching and re-approximating by a Gaussian at every step (Eq. (14)). This is an approximation, not an exact closed-form trajectory predictive. Since the reported NLL and calibration numbers are based on this Gaussian/moment-matched surrogate, the paper should state this limitation in the abstract and contributions, and clarify that multi-step predictive uncertainty is approximate.
  3. [Sec. 6 and Sec. 5.1] The fixed context encoder is an explicit and load-bearing assumption: all adaptation is confined to the Koopman operator K while G_θ is held fixed. Section 6 acknowledges this, but the experiments never stress-test it. Every simulated shift is in dynamics parameters (gravity, slope, friction, joint failures, damping); there is no test where the observation distribution or latent representation itself changes (e.g., sensor noise level, partial observability, sensor bias). Given the abstract's broad claim of robustness to distribution shifts, please either add such a stress test or explicitly scope the claim to shifts in dynamics parameters that leave the embedded representation valid.
minor comments (5)
  1. [Eq. (17)] The total loss is written with an extra division-like symbol in the KL term. Please clean up the notation: L_total = L_NLL + Σ_i D_KL(q_φ(ũ_i|u_i) || N(0,I)).
  2. [References] Reference [2] (Braun et al., Behavioural Brain Research) is not a standard meta-learning reference; the sentence in Section 1 citing [2] for meta-learning appears to be a citation error. Please cite an appropriate meta-learning survey or foundational paper.
  3. [Table 1] The NeuralODE entry reports 0.1683±0.0000. A zero standard deviation across runs is implausible; please report the actual spread or explain the rounding.
  4. [Sec. 5.1 and Appendix F.2] The main text says the ensemble baseline ('EMLP') uses five networks, while Appendix F.2 says ten. Please make these consistent.
  5. [Appendix D.2, Table 5] The column headers '1, 4, 8, 16' are not labelled as the history length in time steps. Please clarify the unit and mention the trade-off explicitly.

Circularity Check

0 steps flagged

No significant circularity: the derivation is standard conjugate Bayesian meta-learning, and no prediction reduces to a fitted input or to a self-citation chain.

full rationale

MetaKoopman's claimed derivation chain is standard MNIW conjugate Bayesian updating. Lemma 4.1 cites an external textbook (Murphy [23]) for the MNIW posterior form, and Lemma 4.2's posterior predictive derivation is self-contained in Appendix A, reducing to a Student-t whose Gaussian approximation is used for multi-step moment matching. The meta-learning objective (Eq. 15) optimizes the prior on held-out query sets per task, which is the standard meta-learning protocol, not a fitted prediction masquerading as a result. The prior parameters and beta are meta-trained on a task distribution and then fixed at test time; no equation defines a claimed output in terms of the same output. Section 5.3's pre/post adaptation comparison is a sanity check rather than an independent benchmark, but the paper frames it as such and it does not create a logical circle. Section 6 explicitly states the fixed-encoder assumption, which is a stated limitation, not a hidden circular input. No load-bearing self-citations or author-imported uniqueness theorems appear in the reference list. Note for the correctness review, not circularity: Eq. (11) sets Vhat = S_zz while the Appendix and Eq. (13) treat Vhat as the posterior column covariance; the conjugate posterior covariance should be (ZZ^T + V0^{-1})^{-1} (or an inverse under the paper's convention), so the closed-form update and predictive equations are internally inconsistent. That is a mathematical correctness risk, not a circularity, because the incorrect formula is not equivalent to its own input; it simply does not follow from the stated prior and likelihood.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

No new physical entities are introduced. The central assumptions are that a fixed learned embedding linearizes the dynamics and remains valid under shift, and that the Student-t predictive can be treated as Gaussian. Both are stated in the paper but not independently verified. The free parameters β, the MNIW prior parameters, the latent dimension η, and the history length q are all fitted or chosen and directly affect the reported performance.

free parameters (4)
  • Tempering factor β = not reported (meta-learned, in (0,1])
    Eq (16) scales prior precision at test time; ablation (Table 4) shows it changes results, so it is a fitted knob rather than a derived constant.
  • MNIW prior parameters (M̆, V̆, ν̆, Ψ̆) = not reported; learned by minimizing Eq (15)
    These define the prior over the Koopman operator and are fitted to the meta-training task distribution; they are the main learned quantities on which the method depends.
  • Latent state dimension η = not stated exactly; 'typically at least 32' (Appendix A)
    This is the dimension of the finite-dimensional Koopman invariant subspace. It is a model-capacity hyperparameter chosen by hand or by search, and the central claims depend on it.
  • Context history length q = 16 in implementation (Appendix F.2)
    The window of past states/actions fed to the context encoder. The ablation (Table 5) shows it affects accuracy, so it is a chosen hyperparameter that matters for the results.
axioms (7)
  • domain assumption A finite-dimensional Koopman-invariant embedding exists: the learned map g gives approximately linear latent dynamics x̃_{t+1} = K z̃_t.
    Section 4.1, Eq (4). This is the core modeling assumption; no proof is given that such a subspace exists for the truck or MuJoCo systems.
  • domain assumption The fixed context encoder Gθ remains valid under distribution shift; only K changes.
    Section 6 explicitly states the encoder is kept deterministic and only uncertainty over K is modeled. If the representation itself shifts, the method cannot adapt.
  • ad hoc to paper The posterior predictive can be approximated by a Gaussian / moment-matched Gaussian.
    Lemma 4.2 and Appendix A: the Student-t predictive is approximated by a Gaussian for ν'>30, and multi-step forecasts use recursive moment matching. This is an approximation, not an exact closed form for multi-step trajectories.
  • standard math MNIW conjugacy of Bayesian linear regression (Lemma 4.1).
    The posterior remains MNIW; the proof is cited to [23]. This is standard conjugate Bayesian linear algebra.
  • domain assumption The latent action prior N(0,I) is accurate enough for planning rollouts.
    Section 4.4, Eq (17): the variational action encoder is regularized toward a standard Gaussian, and at test time actions are sampled from N(0,I), assuming the prior covers the useful action manifold.
  • standard math Takens' time-delay embedding justifies the use of a history window as context.
    Section 4.1 cites [38]. Takens' theorem applies to deterministic systems; here it is used for stochastic controlled systems, so it is a borrowed justification rather than a proven sufficient condition.
  • domain assumption Transition noise is i.i.d. Gaussian with covariance Σ.
    Eqs (7)-(8). The likelihood and the MNIW conjugacy rely on Gaussian i.i.d. noise, which is a standard but unverified assumption for truck dynamics.

pith-pipeline@v1.3.0-alltime-deepseek · 24798 in / 15344 out tokens · 143981 ms · 2026-08-01T00:04:59.142837+00:00 · methodology

0 comments
read the original abstract

Modeling and forecasting nonlinear dynamics under distribution shifts is essential for robust decision-making in real-world systems. In this work, we propose MetaKoopman, a Bayesian meta-learning framework for modeling nonlinear dynamics through linear latent representations. MetaKoopman learns a Matrix Normal-Inverse Wishart (MNIW) prior over the Koopman operator, enabling closed-form Bayesian updates conditioned on recent trajectory segments. Moreover, it provides a closed-form posterior predictive distribution over future state trajectories, capturing both epistemic and aleatoric uncertainty in the learned dynamics. We evaluate MetaKoopman on a full-scale autonomous truck and trailer system across a wide range of adverse winter scenarios, including snow, ice, and mixed-friction conditions, as well as in simulated control tasks with diverse distribution shifts. MetaKoopman consistently outperforms prior approaches in multi-step prediction accuracy, uncertainty calibration, and robustness to distributional shifts. Field experiments further demonstrate its effectiveness in dynamically feasible motion planning, particularly during evasive maneuvers and operation at the limits of traction. Project website: https://mahmoud-selim.github.io/MetaKoopman/

Figures

Figures reproduced from arXiv: 2607.26345 by Karl H. Johansson, Mahmoud Selim, Sriharsha Bhat.

Figure 1
Figure 1. Figure 1: An Overview of the MetaKoopman framework. (1) Past states and actions are encoded into [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Validation MSE over time across selected environments. MetaKoopman consistently [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: As the truck approaches a previously unobserved ice patch on a slippery road, MetaKoopman [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: MSE distribution before and after Koopman operator adaptation on the truck dataset. The post-update distribution (blue) shows a shift toward lower errors compared to the pre-update (orange) To evaluate whether adaptation meaningfully al￾ters the underlying dynamics model, we assess MetaKoopman’s capability to refine its predictions after incorporating recent trajectory history. Using the real-world truck d… view at source ↗
Figure 5
Figure 5. Figure 5: Evasive lane change on snow. (a–d): A virtual obstacle appears mid-lane on a low￾traction road. Non-adaptive models underestimate stopping distance and attempt braking, resulting in failure. MetaKoopman adapts online and executes a safe lane change to avoid collision. (e): Sensor visualization showing obstacle detection and the feasible trajectory selected by the planner. evasive maneuver. The result is a … view at source ↗
Figure 6
Figure 6. Figure 6: Truck and semi-trailer autonomous vehicle used for testing. One of the key contributions of this work is the application of motion planning techniques for au￾tonomous truck and trailer systems. Autonomous driving datasets are typically expensive to acquire and maintain, with access often restricted to select (OEMs) and their suppliers. This study involved ex￾tensive data collection from a real autonomous t… view at source ↗
Figure 7
Figure 7. Figure 7: Mu-split test scenario captured from within the vehicle cab, illustrating asymmetric surface conditions—one side of the vehicle traveling on high-friction asphalt while the other traverses low-friction polished ice. To evaluate autonomous driving performance under extreme winter conditions, a series of targeted field tests were conducted during February and March 2025 on specialized proving grounds located… view at source ↗
Figure 8
Figure 8. Figure 8: Environments: Panda-Lift, Ant, HalfCheetah, Hopper, and Walker. [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Validation MSE over time for (a) Ant, (b) HalfCheetah-Stationary, (c) Walker, and (d) [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Variational action encoder used in the CLAS planner. The encoder learns a Gaussian distribution over action embeddings, enabling efficient sampling for trajectory roll￾out. Instead of sampling actions from the original control space and encoding them at test time, CLAS samples directly from the latent Gaussian space learned by the variational encoder ( [PITH_FULL_IMAGE:figures/full_fig_p030_10.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

47 extracted references · 10 linked inside Pith

  1. [1]

    Model-based control using koopman operators.arXiv preprint arXiv:1709.01568, 2017

    Ian Abraham, Gerardo De La Torre, and Todd D Murphey. Model-based control using koopman operators.arXiv preprint arXiv:1709.01568, 2017

    Pith/arXiv arXiv 2017

  2. [2]

    Structure learning in action.Be- havioural brain research, 206(2):157–165, 2010

    Daniel A Braun, Carsten Mehring, and Daniel M Wolpert. Structure learning in action.Be- havioural brain research, 206(2):157–165, 2010

    2010

  3. [3]

    Data-driven control of soft robots using koopman operator theory.IEEE Transactions on Robotics, 37(3):948– 961, 2020

    Daniel Bruder, Xun Fu, R Brent Gillespie, C David Remy, and Ram Vasudevan. Data-driven control of soft robots using koopman operator theory.IEEE Transactions on Robotics, 37(3):948– 961, 2020

    2020

  4. [4]

    Korol: Learning visualizable object feature with koopman operator rollout for manipulation.arXiv preprint arXiv:2407.00548, 2024

    Hongyi Chen, Abulikemu Abuduweili, Aviral Agrawal, Yunhai Han, Harish Ravichandar, Changliu Liu, and Jeffrey Ichnowski. Korol: Learning visualizable object feature with koopman operator rollout for manipulation.arXiv preprint arXiv:2407.00548, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    2018

  6. [6]

    Model predictive control of a vehicle using koopman operator.IFAC-PapersOnLine, 53(2):4228–4233, 2020

    Vít Cibulka, Tomáš Haniš, Milan Korda, and Martin Hromˇcík. Model predictive control of a vehicle using koopman operator.IFAC-PapersOnLine, 53(2):4228–4233, 2020

    2020

  7. [7]

    Model-based reinforcement learning via meta-policy optimization

    Ignasi Clavera, Jonas Rothfuss, John Schulman, Yasuhiro Fujita, Tamim Asfour, and Pieter Abbeel. Model-based reinforcement learning via meta-policy optimization. InConference on Robot Learning, pages 617–629. PMLR, 2018

    2018

  8. [8]

    Offline meta reinforcement learning– identifiability challenges and effective data collection strategies.Advances in Neural Information Processing Systems, 34:4607–4618, 2021

    Ron Dorfman, Idan Shenfeld, and Aviv Tamar. Offline meta reinforcement learning– identifiability challenges and effective data collection strategies.Advances in Neural Information Processing Systems, 34:4607–4618, 2021

    2021

  9. [9]

    Rl ˆ 2: Fast reinforcement learning via slow reinforcement learning.arXiv preprint arXiv:1611.02779, 2016

    Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl ˆ 2: Fast reinforcement learning via slow reinforcement learning.arXiv preprint arXiv:1611.02779, 2016

    Pith/arXiv arXiv 2016

  10. [10]

    Model-agnostic meta-learning for fast adap- tation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adap- tation of deep networks. InInternational conference on machine learning, pages 1126–1135. PMLR, 2017

    2017

  11. [11]

    Probabilistic model-agnostic meta-learning

    Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018

    2018

  12. [12]

    A disentangled recognition and nonlinear dynamics model for unsupervised learning.Advances in neural information processing systems, 30, 2017

    Marco Fraccaro, Simon Kamronn, Ulrich Paquet, and Ole Winther. A disentangled recognition and nonlinear dynamics model for unsupervised learning.Advances in neural information processing systems, 30, 2017

    2017

  13. [13]

    Desko: Stability-assured robust control with a deep stochastic koopman operator

    Minghao Han, Jacob Euler-Rolle, and Robert K Katzschmann. Desko: Stability-assured robust control with a deep stochastic koopman operator. InInternational conference on learning representations (ICLR), 2022

    2022

  14. [14]

    Deep koopman learning of nonlinear time-varying systems.arXiv preprint arXiv:2210.06272, 2022

    Wenjian Hao, Bowen Huang, Wei Pan, Di Wu, and Shaoshuai Mou. Deep koopman learning of nonlinear time-varying systems.arXiv preprint arXiv:2210.06272, 2022

    arXiv 2022

  15. [15]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  16. [16]

    Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931

    Bernard O Koopman. Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931. 11

    1931

  17. [17]

    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

  18. [18]

    Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    2017

  19. [19]

    Qianxiao Li, Felix Dietrich, Erik M Bollt, and Ioannis G Kevrekidis. Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the koopman operator.Chaos: An Interdisciplinary Journal of Nonlinear Science, 27(10), 2017

    2017

  20. [20]

    Deep learning for universal linear embeddings of nonlinear dynamics.Nature communications, 9(1):4950, 2018

    Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Deep learning for universal linear embeddings of nonlinear dynamics.Nature communications, 9(1):4950, 2018

    2018

  21. [21]

    Derivative-based koopman operators for real-time control of robotic systems.IEEE Transactions on Robotics, 37(6):2173–2192, 2021

    Giorgos Mamakoukas, Maria L Castano, Xiaobo Tan, and Todd D Murphey. Derivative-based koopman operators for real-time control of robotic systems.IEEE Transactions on Robotics, 37(6):2173–2192, 2021

    2021

  22. [22]

    Koopman- lqr controller for quadrotor uavs from data

    Zeyad M Manaa, Ayman M Abdallah, Mohammad A Abido, and Syed S Azhar Ali. Koopman- lqr controller for quadrotor uavs from data. In2024 IEEE International Conference on Smart Mobility (SM), pages 153–158. IEEE, 2024

    2024

  23. [23]

    MIT press, 2023

    Kevin P Murphy.Probabilistic machine learning: Advanced topics. MIT press, 2023

    2023

  24. [24]

    Learning to adapt in dynamic, real-world environments through meta- reinforcement learning.arXiv preprint arXiv:1803.11347, 2018

    Anusha Nagabandi, Ignasi Clavera, Simin Liu, Ronald S Fearing, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Learning to adapt in dynamic, real-world environments through meta- reinforcement learning.arXiv preprint arXiv:1803.11347, 2018

    Pith/arXiv arXiv 2018

  25. [25]

    Reptile: a scalable metalearning algorithm.arXiv preprint arXiv:1803.02999, 2(3):4, 2018

    Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm.arXiv preprint arXiv:1803.02999, 2(3):4, 2018

    Pith/arXiv arXiv 2018

  26. [26]

    Linearly recurrent autoencoder networks for learning dynamics.SIAM Journal on Applied Dynamical Systems, 18(1):558–593, 2019

    Samuel E Otto and Clarence W Rowley. Linearly recurrent autoencoder networks for learning dynamics.SIAM Journal on Applied Dynamical Systems, 18(1):558–593, 2019

    2019

  27. [27]

    Dynamic mode decomposition with control.SIAM Journal on Applied Dynamical Systems, 15(1):142–161, 2016

    Joshua L Proctor, Steven L Brunton, and J Nathan Kutz. Dynamic mode decomposition with control.SIAM Journal on Applied Dynamical Systems, 15(1):142–161, 2016

    2016

  28. [28]

    Efficient off-policy meta-reinforcement learning via probabilistic context variables

    Kate Rakelly, Aurick Zhou, Chelsea Finn, Sergey Levine, and Deirdre Quillen. Efficient off-policy meta-reinforcement learning via probabilistic context variables. InInternational conference on machine learning, pages 5331–5340. PMLR, 2019

    2019

  29. [29]

    Meta-learning with latent embedding optimization.arXiv preprint arXiv:1807.05960, 2018

    Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization.arXiv preprint arXiv:1807.05960, 2018

    Pith/arXiv arXiv 2018

  30. [30]

    Meta reinforcement learning with latent variable gaussian processes.arXiv preprint arXiv:1803.07551, 2018

    Steindór Sæmundsson, Katja Hofmann, and Marc Peter Deisenroth. Meta reinforcement learning with latent variable gaussian processes.arXiv preprint arXiv:1803.07551, 2018

    Pith/arXiv arXiv 2018

  31. [31]

    Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    Peter J Schmid. Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    2010

  32. [32]

    Applications of the dynamic mode decomposition.Theoretical and computational fluid dynamics, 25:249–259, 2011

    Peter J Schmid, Larry Li, Matthew P Juniper, and Oliver Pust. Applications of the dynamic mode decomposition.Theoretical and computational fluid dynamics, 25:249–259, 2011

    2011

  33. [33]

    Deep koopman operator with control for nonlinear systems

    Haojie Shi and Max Q-H Meng. Deep koopman operator with control for nonlinear systems. IEEE Robotics and Automation Letters, 7(3):7700–7707, 2022

    2022

  34. [34]

    Koopman operators in robot learning.arXiv preprint arXiv:2408.04200, 2024

    Lu Shi, Masih Haseli, Giorgos Mamakoukas, Daniel Bruder, Ian Abraham, Todd Murphey, Jorge Cortés, and Konstantinos Karydis. Koopman operators in robot learning.arXiv preprint arXiv:2408.04200, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Adaptive koopman embedding for robust control of complex nonlinear dynamical systems.arXiv preprint arXiv:2405.09101, 2024

    Rajpal Singh, Chandan Kumar Sah, and Jishnu Keshavan. Adaptive koopman embedding for robust control of complex nonlinear dynamical systems.arXiv preprint arXiv:2405.09101, 2024. 12

    arXiv 2024

  36. [36]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  37. [37]

    Learning koopman invariant sub- spaces for dynamic mode decomposition.Advances in neural information processing systems, 30, 2017

    Naoya Takeishi, Yoshinobu Kawahara, and Takehisa Yairi. Learning koopman invariant sub- spaces for dynamic mode decomposition.Advances in neural information processing systems, 30, 2017

    2017

  38. [38]

    Detecting strange attractors in turbulence

    Floris Takens. Detecting strange attractors in turbulence. InDynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80, pages 366–381. Springer, 2006

    1980

  39. [39]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026–5033. IEEE, 2012

    2012

  40. [40]

    Deep koopman data-driven control framework for autonomous racing

    Rongyao Wang, Yiqiang Han, and Umesh Vaidya. Deep koopman data-driven control framework for autonomous racing. InProc. Int. Conf. Robot. Autom.(ICRA) Workshop Opportunities Challenges Auton. Racing, pages 1–6, 2021

    2021

  41. [41]

    Koopman neural forecaster for time series with temporal distribution shifts.arXiv preprint arXiv:2210.03675, 2022

    Rui Wang, Yihe Dong, Sercan Ö Arik, and Rose Yu. Koopman neural forecaster for time series with temporal distribution shifts.arXiv preprint arXiv:2210.03675, 2022

    Pith/arXiv arXiv 2022

  42. [42]

    A data–driven approxima- tion of the koopman operator: Extending dynamic mode decomposition.Journal of Nonlinear Science, 25:1307–1346, 2015

    Matthew O Williams, Ioannis G Kevrekidis, and Clarence W Rowley. A data–driven approxima- tion of the koopman operator: Extending dynamic mode decomposition.Journal of Nonlinear Science, 25:1307–1346, 2015

    2015

  43. [43]

    Learning deep neural network representations for koopman operators of nonlinear dynamical systems

    Enoch Yeung, Soumya Kundu, and Nathan Hodas. Learning deep neural network representations for koopman operators of nonlinear dynamical systems. In2019 American Control Conference (ACC), pages 4832–4839. IEEE, 2019

    2019

  44. [44]

    Ode2vae: Deep generative second order odes with bayesian neural networks.Advances in Neural Information Processing Systems, 32, 2019

    Cagatay Yildiz, Markus Heinonen, and Harri Lahdesmaki. Ode2vae: Deep generative second order odes with bayesian neural networks.Advances in Neural Information Processing Systems, 32, 2019

    2019

  45. [45]

    Bayesian model-agnostic meta-learning.Advances in neural information processing systems, 31, 2018

    Jaesik Yoon, Taesup Kim, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, and Sungjin Ahn. Bayesian model-agnostic meta-learning.Advances in neural information processing systems, 31, 2018

    2018

  46. [46]

    Limitations

    Hao Zhang, Clarence W Rowley, Eric A Deem, and Louis N Cattafesta. Online dynamic mode decomposition for time-varying systems.SIAM Journal on Applied Dynamical Systems, 18(3):1586–1609, 2019. NeurIPS Paper Checklist 1.Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [...

    2019

  47. [47]

    V| η 2 exp −1 2 tr Σ−1 (˜xt+1 − K˜zt)(˜xt+1 − K˜zt)⊤ + K −“M “V −1 K −“M ⊤ = Z 1 (2π) η+ηk 2 |Σ| 1+k 2 | “V| η 2 exp

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

    2025