Pith. sign in

REVIEW 3 major objections 5 minor 40 references

For controlling PDEs, planning against a learned kinetic-energy readout on a latent world model outperforms minimizing raw latent L2 distance, provided the observable is injective on reachable states.

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 12:20 UTC pith:O2KOSKHL

load-bearing objection First JEPA-style latent MPC for PDEs, but the headline aperiodic result is confounded: changed cost scale and temperature together, so the 53% RMSE claim is not clean. the 3 major comments →

arxiv 2607.21644 v1 pith:O2KOSKHL submitted 2026-07-21 cs.LG cs.SYeess.SY

Toward Goal-Agnostic Joint-Embedding Predictive Control of Partial Differential Equations

Jonathan Gallagher , Roberto Guglielmi This is my paper
classification cs.LG cs.SYeess.SY
keywords joint-embedding predictive architecturelatent world modelPDE controlmodel-predictive path integralkinetic-energy probeobservable-aligned controlNavier-Stokesgoal-agnostic learning
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.

This paper tries to establish that a joint-embedding predictive world model, trained offline with no reward and no field reconstruction, can serve as a goal-agnostic controller for partial differential equations. The central claim is that when a physical observable such as kinetic energy is available and injective on the reachable state family, applying the control objective to that observable beats applying it to raw Euclidean distance in the latent space. On a 2D Navier-Stokes boundary-control benchmark, the kinetic-energy probe improves native reward from -12.08 to -10.90, lowers late velocity-field RMSE, and cuts late field error by 53% on three withheld aperiodic targets. The gain comes from removing phase lag, not from higher cumulative reward, and the probe is explicitly fragile under measurement noise and missing pixels. The practical pay-off would be one frozen world model supporting tracking, stabilization, and regulation by swapping only the deployment-time objective.

Core claim

The authors demonstrate that a frozen joint-embedding predictive architecture, trained entirely offline without task rewards, can control a 2D Navier-Stokes system through a model-predictive path integral planner. Their main finding is that the planner's cost should be computed on a calibrated physical observable, here a linear kinetic-energy probe fitted to latent rollouts, rather than on raw latent L2 distance. With this observable-aligned cost, the same frozen checkpoint and five-step latent horizon improves the matched 50-episode native reward from -12.08 to -10.90 and lowers late field RMSE from 0.0765 to 0.0692; across three deliberately held-out aperiodic targets it cuts late field RM

What carries the argument

The load-bearing component is a masked joint-embedding predictive world model: a small 2D vision transformer encoder and an action-conditioned latent predictor trained with a residual rollout loss, anti-collapse penalties, and temporal straightening, with no decoder from latent to field. On top of this frozen latent dynamics, a lightweight linear probe reads out kinetic energy from predicted latents, and the MPPI controller scores candidate action sequences by squared error in that physical scalar instead of latent Euclidean distance. The probe supplies a calibrated physical cost while the frozen latent dynamics remain goal-agnostic, and the injectivity of the kinetic-energy map on reachable

Load-bearing premise

Kinetic energy is assumed to be effectively one-to-one on the set of velocity fields reachable by the single positive scalar boundary actuator, so that matching kinetic energy forces the velocity field to match; the paper does not verify this injectivity directly and explicitly warns against extrapolating beyond this setting.

What would settle it

Run the same kinetic-energy-probe planner against latent-L2 planning on a Navier-Stokes benchmark with two or more independently controlled boundary actuators, or on a flow with vortex shedding: if field RMSE no longer improves, the injectivity premise is the failing link. More directly, search the reachable set for two distinct velocity fields with equal kinetic energy whose subsequent evolution diverges; if such a pair exists, kinetic-energy matching cannot guarantee field matching.

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

If this is right

  • A single frozen world model, trained without rewards or state reconstruction, can natively support multiple control objectives—time-dependent tracking, fixed-target stabilization, and scalar energy regulation—by changing only the deployment-time cost.
  • Using a calibrated physical observable as the control objective can systematically beat raw latent-L2 planning, improving both native reward and state-space field RMSE on the benchmark.
  • Across three withheld aperiodic targets, kinetic-energy planning cuts late field RMSE by 53% and removes phase lag in all 30 matched episodes, even though cumulative native reward is essentially unchanged.
  • The kinetic-energy probe is brittle under corrupted observations: clean 2.7% relative error rises to 46–64% under 5–15% Gaussian sensor noise and 60–85% under 5–25% missing pixels, so the reported tracking gains are limited to clean, full-state measurements.
  • The injectivity assumption, not the world model itself, is the main barrier to generalization: the paper explicitly warns against extrapolating to multi-actuator or topologically different flows without additional observables.

Where Pith is reading between the lines

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

  • One extension beyond the paper: if injectivity of a scalar observable is the real driver, then a practical design rule would be to numerically verify unique continuation of candidate observables before planning—turning the paper's post-hoc explanation into a constructive selection criterion for other PDE settings.
  • The fact that a linear probe on frozen latents suffices suggests the benefit comes primarily from cost geometry, not probe expressiveness; adding a second scalar (e.g., enstrophy or a coarse field projection) could restore injectivity in multi-actuator systems where kinetic energy alone fails.
  • The paper's observed fragility under noise points to a concrete fix: train the probe on corrupted or masked latents, or equip it with uncertainty calibration, so the observable readout remains on-manifold under imperfect measurements; this would extend the framework to realistic sensor conditions without retraining the dynamics.
  • The same goal-agnostic recipe could transfer to other PDE families that admit an injective scalar observable (energy for wave equations, entropy for compressible flows), making latent world-model control viable wherever such a quantity is known.

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 a goal-agnostic joint-embedding predictive architecture (JEPA) for PDE control. A masked ViT encoder and action-conditioned latent predictor are trained offline without reward or field reconstruction, then frozen and used by an MPPI planner. The central claim is that, when a physical observable such as kinetic energy can be read out from the latent space, planning against that observable (provided injectivity) outperforms planning against raw Euclidean distance in the latent space. Experiments on the PDE Control Gym 2D Navier–Stokes benchmark show a clean 50-episode native-reward improvement from -12.08±0.86 to -10.90±0.91 and a reduction in late-field RMSE from 0.0765 to 0.0692. On three aperiodic targets, the paper reports a 53% reduction in late-field RMSE with 30/30 paired wins.

Significance. If the central claim holds, the paper makes a useful contribution: it demonstrates that a reconstruction-free latent world model can serve as a goal-agnostic substrate for PDE control, and that a cheaply fitted observable readout can improve planning without retraining the dynamics. The clean 50-episode reward comparison, the explicit out-of-distribution audit for the aperiodic targets, the reporting of 95% CIs, and the detailed hyperparameter table are commendable. However, the headline aperiodic result (53% RMSE reduction, 30/30 wins) is confounded by a change in MPPI temperature, and the paper's own framing that 'only the deployment-time cost interface changes' is violated. This weakens the strongest quantitative claim in the abstract and Table 3, though it does not invalidate the matched-temperature 50-episode evidence.

major comments (3)
  1. [§3.2, Table 3, Figure 7, Abstract] The aperiodic comparison changes two variables simultaneously: the cost function and the MPPI temperature. The text states that the KE cost uses (β,λa)=(1,0) and T=0.004, which is equivalent to (β,T)=(25,0.1), while the latent-L2 baseline retains T=0.1. MPPI weights sampled sequences by exp(−cost/T); lowering T from 0.1 to 0.004 sharpens the distribution around low-cost plans and effectively multiplies the cost weight by 25. The reported failure mode of latent-L2 is phase lag, exactly what a higher-gain, lower-temperature planner would reduce. Thus the 53% late-field-RMSE reduction and 30/30 paired wins may be attributable to the temperature change rather than to the kinetic-energy observable. This violates the paper's claim (§6) that only the deployment-time cost interface changes. The authors should rerun the latent-L2 baseline at a matched temperature (or equivalently scale α so that
  2. [§5.1.1, Figure 5] The KE scale β is selected by sweeping over ten matched clean episodes, and the same ten-episode set appears to inform the choice β=1 that is then used for the 50-episode confirmation. If the ten-episode sweep is used to pick the operating point, the 50-episode result is partially conditioned on that small evaluation set. The paper does not report whether the ten episodes are disjoint from the 50 or whether a separate validation set was used. This is not fatal because the 50-episode confirmation is an independent check, but the tuning protocol should be described precisely (e.g., whether any hyperparameter was selected on the ten-episode set and whether the 50 seeds include those ten).
  3. [§6, Discussion] A load-bearing assumption is that kinetic energy is injective on the reachable state/action family, so that minimizing KE error implies minimizing field error. The paper asserts this based on the single positive scalar boundary actuator, but no direct verification is provided. The authors themselves caution against extrapolation to multi-actuator or topologically distinct flows. Since the central recommendation is conditioned on injectivity, the paper should either provide a concrete test (e.g., offline analysis of the map from latent states to KE, checking whether distinct reachable states with similar KE have similar fields) or explicitly downgrade the claim to an empirical observation for this benchmark. As written, the injectivity argument is plausible but unverified.
minor comments (5)
  1. [§1, Contributions] Typo: 'ever working in state space' should be 'never working in state space' (or 'without ever working in state space').
  2. [§6, Discussion] Typo: 'approacah' should be 'approach'; 'ant' should be 'and'.
  3. [§1, Introduction] Typo: 'joing-embedding' should be 'joint-embedding'.
  4. [Table 4, Appendix A] The table lists MPPI noise as 0.6, but the text in §3.2 does not define the noise scale; consider clarifying whether this is the standard deviation of the action perturbation and how it relates to the temperature scaling.
  5. [§5.1.1, Table 1] The row 'target replay (oracle)' reports reward -18.192 with late RMSE 0.00390, while constant 2.5 has reward -10.877. This contrast is surprising and worth a sentence of explanation, since 'target replay' presumably applies the true control signal and would naively be expected to achieve the best reward.

Circularity Check

0 steps flagged

No significant circularity: the central comparison is validated by external reward and field-RMSE metrics, and the KE probe is checked against true kinetic energy; the aperiodic temperature mismatch is a confound, not a circularity.

full rationale

The derivation chain is self-contained and not circular. The JEPA world model is trained without reward or field reconstruction, and the linear KE probe is an independent ridge regressor q_omega fit on frozen predictor-rollout latents; its fidelity is measured against true kinetic energy (R^2=0.989), not against the control outcome. The headline claims are reported in native benchmark reward and velocity-field RMSE, both external to the probe's coordinates, so minimizing the probe cost cannot by construction manufacture the state-space gain; the gain depends on an empirical injectivity relation that the paper explicitly treats as an assumption and flags as non-generalizable (Discussion: 'we infer that our approach succeeds here because a single scalar boundary actuator ... restricts the evaluated trajectories to a narrow response family where the map from state to kinetic energy is actually injective'). There are no load-bearing self-citations: the reference list contains no work by Gallagher or Guglielmi, and cited methods (MPPI, VICReg, Delta-JEPA, temporal straightening) are used as off-the-shelf components, not as prior justification of the central claim. The main legitimate concern is the aperiodic comparison in Section 3.2/Table 3, where KE planning uses T=0.004 while the latent-L2 baseline retains T=0.1; the paper discloses the equivalence to (beta,T)=(25,0.1), so this is a confounded experimental comparison (temperature and effective cost weight changed together) rather than a circular reduction. The unproved injectivity assertion is missing support, not circularity, and the choice of beta on ten matched episodes is hyperparameter selection, not a fitted result renamed as a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on the KE-probe planning setup rather than on any newly postulated physical entity. The main free parameters are the planning cost scale (β, temperature, action penalty) and training-loss weights; these are tuned or chosen and vary across the headline comparisons. The key conceptual axiom is injectivity of kinetic energy on the reachable response family, which is asserted but not demonstrated. The clean-probe validity and distribution shift assumptions are also load-bearing and are partially contradicted by the paper's own corruption experiments.

free parameters (5)
  • KE tracking scale β = β=1 for reward comparison; β=1 with T=0.004 (equivalent to β=25 at T=0.1) for aperiodic suite
    Swept over [0.25,200] on ten matched clean episodes (Fig. 5); the operating point for the 50-episode benchmark and for the aperiodic suite is chosen from this sweep, i.e., tuned on the evaluation benchmark.
  • MPPI temperature T = 0.1 (reward comparison) and 0.004 (aperiodic suite)
    Changes the effective cost scale by 25× between the two headline comparisons; the paper notes the equivalence but the two settings are not unified under a single protocol.
  • Action penalty weight λ_a = 0 or 0.05 depending on configuration
    Used in J(q) / J(lat) costs; differs between tracking-only, reward comparison, and aperiodic suite.
  • Composite training loss weights = λ_var=1, λ_cov=1, λ_str=0.1, λ_Δa=10
    Chosen hyperparameters that shape the latent geometry; the paper does not report sensitivity to these weights.
  • MPPI sampling parameters = 512 samples, 3 iterations, noise 0.6, horizon 5
    Fixed for all experiments but not swept; could affect the comparison if the controller is not converged.
axioms (5)
  • domain assumption Kinetic energy is injective over the reachable state/action family
    Invoked in §Discussion ('where kinetic energy determines state exactly') and in the abstract's 'provided injectivity'; the paper asserts this from the single positive scalar actuator but does not verify it directly.
  • domain assumption The clean-calibrated KE probe remains valid for clean current-state encodings
    The probe is fit on clean predictor-rollout latents and applied to current-state encodings; §5.3 shows this fails under noise and missing pixels, so the clean-case validity is a load-bearing assumption.
  • domain assumption Offline training distribution covers the closed-loop deployment targets
    The controller is deployed on the benchmark reference and three aperiodic signals; the paper audits the aperiodic triplet in action space but the closed-loop state distribution is not fully characterized.
  • domain assumption Masked training makes the encoder robust to partial/complete observations alike
    The masking protocol trains the encoder on incomplete fields, but §5.3 shows the probe is not robust to corrupted measurements; the encoder-side robustness is not separately established.
  • standard math MPPI exponential weighting converges to a reasonable planner solution
    Standard MPPI convergence properties are assumed from Williams et al. [36]; the paper uses only 3 iterations and 512 samples without a convergence check.

pith-pipeline@v1.3.0-alltime-deepseek · 243 in / 10177 out tokens · 131160 ms · 2026-08-01T12:20:38.179886+00:00 · methodology

0 comments
read the original abstract

We present a goal-agnostic control framework for partial differential equations (PDEs) built around a joint-embedding predictive architecture (JEPA). The small 2D ViT encoder and action-conditioned latent dynamics are trained offline without a reward or downstream goal, frozen, and reused by a model-predictive path integral (MPPI) controller. We find that when available, the control objective is better applied to an explicit physical observable (provided injectivity) than to minimizing raw Euclidean distance ($L^2$) in the learned latent space. For a learned linear kinetic-energy (KE) probe on frozen latent rollouts we can reproduce held-out trajectories with $R^2=0.989$, while requiring no change to the underlying world model. On the PDE Control Gym 2D Navier--Stokes benchmark, using KE-probe planning improves the matched 50-episode native reward from $-12.08\pm0.86$ for latent-$L^2$ planning to $-10.90\pm0.91$ (95\% CI), while lowering last-quarter velocity-field RMSE from $0.0765$ to $0.0692$. Across three intentionally withheld, dissimilar, aperiodic targets, KE planning lowers late field RMSE by $53\%$ relative to latent-$L^2$ planning ($0.0220$ versus $0.0469$), winning all 30 paired episodes. The same frozen model also supports controls targeting stabilization around a steady configuration via direct regulation of KE achieving $2.7\%$ mean relative error. While the latent probe is brittle to measurement noise and missing pixels, we believe the results support the claim that latent dynamics can remain both dynamic and goal-agnostic while calibrated observables (granted they guarantee unique continuation) may be a better objective for state control

Figures

Figures reproduced from arXiv: 2607.21644 by Jonathan Gallagher, Roberto Guglielmi.

Figure 1
Figure 1. Figure 1: Masking protocol: contiguous rectangles, random pixels, and row strips, each shown hiding roughly [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Joint-embedding training process, in which ground-truth and partially masked observations are [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Composition of the offline training corpus, where (a) denotes the categories of control actions, (b) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training and validation diagnostics over the full 0–100k schedule. The action decoder remains [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Clean linear-reference KE cost-scale screen over ten matched episodes (mean and 95% CI). The [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Action-space out-of-distribution audit for the aperiodic triplet. Left: the frozen triplet schedule and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Clean tracking of three aperiodic targets over ten matched initial conditions per signal (mean with [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Aperiodic pulse reconstruction for case A: Up-down-up triplet, where both the learned and true [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Derived-quantity control with a rollout-latent kinetic-energy probe. From ten random initial [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Secondary corruption diagnostic over ten matched episodes. Fixed-field stabilization remains [PITH_FULL_IMAGE:figures/full_fig_p016_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

40 extracted references · 1 canonical work pages

  1. [1]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. URLhttps: //arxiv.org/abs/2301.08243

    Pith/arXiv arXiv 2023

  2. [2]

    V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    Mahmoud Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025. URL https://arxiv.org/abs/2506.09985

    Pith/arXiv arXiv 2025

  3. [3]

    VICReg: Variance-invariance-covariance regularization for self- supervised learning

    Adrien Bardes, Jean Ponce, and Yann LeCun. VICReg: Variance-invariance-covariance regularization for self- supervised learning. InInternational Conference on Learning Representations, 2022. URLhttps://arxiv.org/ abs/2105.04906

    Pith/arXiv arXiv 2022

  4. [4]

    PDE Control Gym: A benchmark for data- driven boundary control of partial differential equations.arXiv preprint arXiv:2405.11401, 2024

    Luke Bhan, Yuexin Bian, Miroslav Krstic, and Yuanyuan Shi. PDE Control Gym: A benchmark for data- driven boundary control of partial differential equations.arXiv preprint arXiv:2405.11401, 2024. URLhttps: //arxiv.org/abs/2405.11401

    Pith/arXiv arXiv 2024

  5. [5]

    Neural operators for bypassing gain and control computations in PDE backstepping.IEEE Transactions on Automatic Control, 69:5310–5325, 2024

    Luke Bhan, Yuanyuan Shi, and Miroslav Krstic. Neural operators for bypassing gain and control computations in PDE backstepping.IEEE Transactions on Automatic Control, 69:5310–5325, 2024. URLhttps://arxiv.org/ abs/2302.14265

    Pith/arXiv arXiv 2024

  6. [6]

    Brunton and Bernd R

    Steven L. Brunton and Bernd R. Noack. Closed-loop turbulence control: Progress and challenges.Applied Mechanics Reviews, 67(5):050801, 2015. doi: 10.1115/1.4031175

    doi:10.1115/1.4031175 2015

  7. [7]

    Control of chaotic systems by deep reinforcement learning.Proceedings of the Royal Society A, 475(2231):20190351, 2019

    Michele Alessandro Bucci, Onofrio Semeraro, Alexandre Allauzen, Guillaume Wisniewski, Laurent Cordier, and Lionel Mathelin. Control of chaotic systems by deep reinforcement learning.Proceedings of the Royal Society A, 475(2231):20190351, 2019. doi: 10.1098/rspa.2019.0351

    arXiv 2019

  8. [8]

    Deep reinforcement learning in a handful of trials using probabilistic dynamics models

    Kurtland Chua, Roberto Calandra, Rowan McAllister, and Sergey Levine. Deep reinforcement learning in a handful of trials using probabilistic dynamics models. InAdvances in Neural Information Processing Systems,

  9. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, 2021. URLhttps:/...

    Pith/arXiv arXiv 2021

  10. [10]

    Nikovski

    Amir-massoud Farahmand, Saleh Nabi, and Daniel N. Nikovski. Deep reinforcement learning for partial differential equation control. In2017 American Control Conference, pages 3120–3127, 2017. doi: 10.23919/ACC.2017.7963427

    arXiv 2017

  11. [11]

    Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In Advances in Neural Information P...

    Pith/arXiv arXiv 2020

  12. [12]

    World models.arXiv preprint arXiv:1803.10122, 2018

    David Ha and Jürgen Schmidhuber. World models.arXiv preprint arXiv:1803.10122, 2018. URL https: //arxiv.org/abs/1803.10122

    Pith/arXiv arXiv 2018

  13. [13]

    Learning latent dynamics for planning from pixels

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. InInternational Conference on Machine Learning, 2019. URL https://arxiv.org/abs/1811.04551

    Pith/arXiv arXiv 2019

  14. [14]

    Dream to control: Learning behaviors by latent imagination

    Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. InInternational Conference on Learning Representations, 2020. URLhttps://arxiv. org/abs/1912.01603

    Pith/arXiv arXiv 2020

  15. [15]

    Temporal difference learning for model predictive control

    Nicklas Hansen, Xiaolong Wang, and Hao Su. Temporal difference learning for model predictive control. In International Conference on Machine Learning, 2022. URLhttps://arxiv.org/abs/2203.04955

    Pith/arXiv arXiv 2022

  16. [16]

    TD-MPC2: Scalable, robust world models for continuous control

    Nicklas Hansen, Hao Su, and Xiaolong Wang. TD-MPC2: Scalable, robust world models for continuous control. InInternational Conference on Learning Representations, 2024. URLhttps://arxiv.org/abs/2310.16828. 17

    Pith/arXiv arXiv 2024

  17. [17]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. URLhttps://arxiv.org/abs/2111.06377

    Pith/arXiv arXiv 2022

  18. [18]

    Learning to control PDEs with differentiable physics

    Philipp Holl, Vladlen Koltun, and Nils Thuerey. Learning to control PDEs with differentiable physics. In International Conference on Learning Representations, 2020. URLhttps://arxiv.org/abs/2001.07457

    Pith/arXiv arXiv 2020

  19. [19]

    Solving PDE-constrained control problems using operator learning

    Rakhoon Hwang, Jae Yong Lee, Jin Young Shin, and Hyung Ju Hwang. Solving PDE-constrained control problems using operator learning. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4504–4512, 2022. URLhttps://arxiv.org/abs/2111.04941

    Pith/arXiv arXiv 2022

  20. [20]

    Nathan Kutz, and Steven L

    Eurika Kaiser, J. Nathan Kutz, and Steven L. Brunton. Sparse identification of nonlinear dynamics for model predictive control in the low-data limit.Proceedings of the Royal Society A, 474(2219):20180335, 2018. doi: 10.1098/rspa.2018.0335

    arXiv 2018

  21. [21]

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

    Milan Korda and Igor Mezić. Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control.Automatica, 93:149–160, 2018. doi: 10.1016/j.automatica.2018.03.046

    doi:10.1016/j.automatica.2018.03.046 2018

  22. [22]

    A path towards autonomous machine intelligence, 2022

    Yann LeCun. A path towards autonomous machine intelligence, 2022. URLhttps://openreview.net/forum? id=BZ5a1r-kVsf. OpenReview preprint, version 0.9.2

    2022

  23. [23]

    Nathan Kutz, and Steven L

    Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics.Nature Communications, 9:4950, 2018. doi: 10.1038/s41467-018-07210-0

    doi:10.1038/s41467-018-07210-0 2018

  24. [24]

    LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026. URL https://arxiv.org/abs/2603.19312

    Pith/arXiv arXiv 2026

  25. [25]

    Kochenderfer, and Freddie Witherden

    Jeremy Morton, Antony Jameson, Mykel J. Kochenderfer, and Freddie Witherden. Deep dynamical modeling and control of unsteady fluid flows. InAdvances in Neural Information Processing Systems, 2018. URL https://arxiv.org/abs/1805.07472

    Pith/arXiv arXiv 2018

  26. [26]

    Optimal control of PDEs using physics-informed neural networks.Journal of Computational Physics, 473:111731, 2023

    Saviz Mowlavi and Saleh Nabi. Optimal control of PDEs using physics-informed neural networks.Journal of Computational Physics, 473:111731, 2023. URLhttps://arxiv.org/abs/2111.09880

    Pith/arXiv arXiv 2023

  27. [27]

    Koopman operator-based model reduction for switched-system control of PDEs

    Sebastian Peitz and Stefan Klus. Koopman operator-based model reduction for switched-system control of PDEs. Automatica, 106:184–191, 2019. doi: 10.1016/j.automatica.2019.05.016

    work page doi:10.1016/j.automatica.2019.05.016 2019

  28. [28]

    Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control.Journal of Fluid Mechanics, 865:281–302, 2019

    Jean Rabault, Miroslav Kuchta, Atle Jensen, Ulysse Réglade, and Nicolas Cerardi. Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control.Journal of Fluid Mechanics, 865:281–302, 2019. doi: 10.1017/jfm.2019.62

    doi:10.1017/jfm.2019.62 2019

  29. [29]

    Skyjepa: Learning long-horizon world models for zero-shot sim-to-real control of quadrotors, 2026

    Pratyaksh Rao, Wancong Zhang, Randall Balestriero, Yann LeCun, and Giuseppe Loianno. Skyjepa: Learning long-horizon world models for zero-shot sim-to-real control of quadrotors, 2026. URLhttps://arxiv.org/abs/ 2606.23444

    Pith/arXiv arXiv 2026

  30. [30]

    Mastering Atari, Go, chess and shogi by planning with a learned model.Nature, 588:604–609, 2020

    Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy Lillicrap, and David Silver. Mastering Atari, Go, chess and shogi by planning with a learned model.Nature, 588:604–609, 2020. doi: 10.1038/s41586-020-03051-4

    doi:10.1038/s41586-020-03051-4 2020

  31. [31]

    Latent feedback control of distributed systems in multiple scenarios through deep learning-based reduced order models.arXiv preprint arXiv:2412.09942, 2024

    Matteo Tomasetto, Francesco Braghin, and Andrea Manzoni. Latent feedback control of distributed systems in multiple scenarios through deep learning-based reduced order models.arXiv preprint arXiv:2412.09942, 2024. URLhttps://arxiv.org/abs/2412.09942

    Pith/arXiv arXiv 2024

  32. [32]

    Efficient collective swimming by harnessing vortices through deep reinforcement learning.Proceedings of the National Academy of Sciences, 115(23):5849–5854, 2018

    Siddhartha Verma, Guido Novati, and Petros Koumoutsakos. Efficient collective swimming by harnessing vortices through deep reinforcement learning.Proceedings of the National Academy of Sciences, 115(23):5849–5854, 2018. doi: 10.1073/pnas.1800923115

    doi:10.1073/pnas.1800923115 2018

  33. [33]

    Ying Wang, Oumayma Bounou, Gaoyue Zhou, Randall Balestriero, Tim G. J. Rudner, Yann LeCun, and Mengye Ren. Temporal straightening for latent planning, 2026. URLhttps://arxiv.org/abs/2603.12231

    arXiv 2026

  34. [34]

    Embed to control: A locally linear latent dynamics model for control from raw images

    Manuel Watter, Jost Tobias Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally linear latent dynamics model for control from raw images. InAdvances in Neural Information Processing Systems, 2015. URLhttps://arxiv.org/abs/1506.07365. 18

    Pith/arXiv arXiv 2015

  35. [35]

    DiffPhyCon: A generative approach to control complex physical systems

    Long Wei, Peiyan Hu, Ruiqi Feng, Haodong Feng, Yixuan Du, Tao Zhang, Rui Wang, Yue Wang, Zhi-Ming Ma, and Tailin Wu. DiffPhyCon: A generative approach to control complex physical systems. InAdvances in Neural Information Processing Systems, 2024. URLhttps://arxiv.org/abs/2407.06494

    Pith/arXiv arXiv 2024

  36. [36]

    Rehg, Byron Boots, and Evangelos A

    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, pages 1714–1721, 2017. doi: 10.1109/ICRA.2017.7989202. URL https://ieeexplore.ieee.org/document/7989202

    arXiv 2017

  37. [37]

    Learning to accelerate partial differential equations via latent global evolution.arXiv preprint arXiv:2206.07681, 2022

    Tailin Wu, Takashi Maruyama, and Jure Leskovec. Learning to accelerate partial differential equations via latent global evolution.arXiv preprint arXiv:2206.07681, 2022. URLhttps://arxiv.org/abs/2206.07681

    Pith/arXiv arXiv 2022

  38. [38]

    Delta-JEPA: Learning action-sensitive world models via latent difference decoding.arXiv preprint arXiv:2606.31232, 2026

    Zhenghao Zhang, Yuanxiang Wang, Zhenyu Guan, Yujia Yang, Bingkang Shi, Tianyu Zong, Hongzhu Yi, Guoqing Chao, Xingchen Chen, Tiankun Yang, Chenxi Bao, Tao Yu, Jingjing Zhou, and Jungang Xu. Delta-JEPA: Learning action-sensitive world models via latent difference decoding.arXiv preprint arXiv:2606.31232, 2026. URLhttps://arxiv.org/abs/2606.31232

    Pith/arXiv arXiv 2026

  39. [39]

    DINO-WM: World models on pre-trained visual features enable zero-shot planning.arXiv preprint arXiv:2411.04983, 2024

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. DINO-WM: World models on pre-trained visual features enable zero-shot planning.arXiv preprint arXiv:2411.04983, 2024. URLhttps://arxiv.org/abs/2411. 04983. 19

    Pith/arXiv arXiv 2024

  40. [2018]

    URLhttps://arxiv.org/abs/1805.12114

    Pith/arXiv arXiv