Simulation Priors for Data-Efficient Deep Learning

Andreas Krause; Bhavya Sukhija; Florian D\"orfler; Jonas Rothfuss; Lenart Treven; Stelian Coros

REVIEW 4 major objections 6 minor 81 references

Simulation Priors for Data-Efficient Deep Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A simulator prior lets Bayesian neural nets learn RC-car drifting with roughly half the data.

desk verdict A sensible, well-written method for simulator-informed Bayesian deep learning; the empirical claims would be stronger with error bars and explicit GP-prior hyperparameters. read the letter →

arxiv 2509.05732 v1 pith:HLL5LGE7 submitted 2025-09-06 cs.LG cs.AI

Lenart Treven , Bhavya Sukhija , Jonas Rothfuss , Stelian Coros , Florian Dörfler , Andreas Krause This is my paper · ORCID

classification cs.LGcs.AI

keywords Bayesiandeeplearningsimulationpriorssim-to-realgapfunctionalSteinvariationalgradientdescentsample-efficientmodel-basedreinforcementsystemidentificationGaussianprocess

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the sim-to-real gap can be closed not by fitting a simulator's parameters or training a correction model on top, but by using the simulator itself, with randomized parameters, as the prior distribution over functions in a Bayesian neural network. If true, any black-box simulator becomes reusable knowledge that lets a learned dynamics model reach useful accuracy and calibrated uncertainty from a fraction of the data. The paper demonstrates the payoff on its hardest testbed: a physical RC car learns a drifting reverse-parking maneuver with about twice the sample efficiency of prior-free Bayesian deep learning, both offline and online.

What carries the argument

The central object is a stochastic-process prior p(h) built from two additive parts: simulator samples g(x, phi) with phi drawn from a parameter prior, plus a zero-mean Gaussian process with an isotropic kernel meant to absorb the sim-to-real gap. The marginal prior score, the gradient of log p(h^X), is estimated from samples using a Gaussian, KDE, SSGE, or the nu-method, and then fed into FSVGD's function-space update, which projects the functional gradient into neural-network parameter space. The additive GP is the piece that lets the prior remain informative near the simulator while still allowing the model to adapt to real dynamics as data accumulates.

What would settle it

Run SimPEL against FSVGD on a system whose true dynamics differ from the simulator by a non-smooth, state-dependent residual, such as a hard friction threshold in velocity, with the GP gap hyperparameters fixed before seeing data. The central claim fails if SimPEL's held-out negative log-likelihood is not below FSVGD's at small sample sizes, or if its credible intervals systematically miss the true residual.

Watch

Extended reading notes

Core claim

The paper claims that Bayesian neural networks can be trained with a prior placed directly on the learned function rather than on network weights, where that function prior is a stochastic process derived from a low-fidelity simulator with randomized parameters plus a Gaussian-process correction for the simulator's misspecification. Given samples from the simulator's parameter distribution, the prior score, the gradient of the log-density of the prior marginals, can be estimated cheaply, and functional Stein variational gradient descent then carries posterior particles from simulated behavior toward the real dynamics as data arrives. The consequence is a single method that behaves like the s

Load-bearing premise

The sim-to-real gap is modeled as a zero-mean Gaussian process with a stationary, isotropic kernel whose variance and lengthscale are set by hand, so the low-data gains depend on that prior being roughly right.

Editorial extensions

If this is right

Any queryable simulator, with no analytic gradient, can be converted into a functional prior: first-principles codes can be dropped into Bayesian neural-network training as-is.
The GP gap term is load-bearing in exactly the settings where parameter randomization cannot cover the misspecification; the pendulum experiment shows that removing it degrades high-data performance.
In offline RL, SimPEL produces useful parking policies from far fewer transitions, and a high-fidelity simulator prior matches FSVGD trained on roughly five times more data.
In online RL on the RC car, about twice fewer episodes are needed, with directed exploration visible from the first episode.
Gaussian prior-score estimation is a cheap default, while KDE, SSGE, and the nu-method trade compute and flexibility for capturing multimodality in the prior.

Reading between the lines

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

The recipe is not tied to FSVGD: any inference algorithm that consumes functional prior scores could take the same simulator-derived prior, so the sample-efficiency gains may transfer across inference families.
A natural test is to replace the hand-set GP gap prior with one whose hyperparameters are learned from a small real-data batch; if the low-data gains persist, SimPEL becomes much less sensitive to its weakest assumption.
For contact-rich or mode-switching systems, where the sim-to-real gap is non-smooth, the Gaussian prior-score estimator is likely the bottleneck and the nu-method the better candidate, a pattern the paper's own multi-modal prior example points to.
The roughly two-fold reduction in physical episodes on one RC car task suggests a broader economic claim: wherever a low-fidelity model exists, simulator priors could roughly halve the cost of experiment collection for model-based RL.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: SimPEL is a legitimate extension of functional BNN inference where the prior is a low-fidelity simulator plus an additive GP for the sim-to-real gap. The core idea is clear and the paper is readable. What earns credit: the derivation of the score-based FSVGD update is coherent, they ablate different score estimators, and they go to the trouble of testing on hardware, not just simulators. The 2x data reduction on the RC car parking task is the kind of result that would matter if it holds up.

The soft spots are empirical. The main curves in Figures 4 and 5 have no error bars, no seed repetitions, and the hardware experiments appear unreplicated. That makes the magnitude of the claimed gains hard to assess. The stress-test concern about the GP gap hyperparameters is legitimate: κ² and ℓ are load-bearing because in the low-data regime the GP term shapes the posterior, and the paper never reports what values were used or any sensitivity analysis. That is a real omission, not a nit. The Gaussian score approximation is a less serious issue—the paper itself shows a multimodal case where it fails and recommends the ν-method, but then uses Gaussian throughout. That's a consistency point, not a fatal flaw.

I also had trouble reproducing from the artifact: the code URL in the paper has a space in it ('simulation transfer'), which is a small but telling sign of a rushed release.

Bottom line: this is a solid, serious paper, but the evidence is not yet conclusive. It deserves peer review, and a good referee should insist on error bars, repeated hardware trials, and explicit reporting of the GP prior hyperparameters and their sensitivity. I'd bring it to a reading group on simulator-informed inference, but I wouldn't cite the empirical gains yet.

Referee Report

4 major / 6 minor

Summary. The paper proposes SimPEL, a method for imposing functional priors on Bayesian neural networks (BNNs) in order to make dynamics learning and model-based RL more sample-efficient. The prior is constructed by combining a low-fidelity simulator with random physical parameters, g(x,ϕ), and an additive zero-mean Gaussian process that models the sim-to-real gap. Since the resulting stochastic-process prior is intractable, SimPEL samples function values on a measurement set and estimates the prior score from those samples, using a Gaussian approximation as the default and KDE/SSGE/ν-method as alternatives. The score is then plugged into the FSVGD functional-space particle update. Experiments cover synthetic pendulum, gene-expression, greenhouse, and race-car dynamics, plus an RC-car hardware platform, in both supervised system identification and model-based RL (offline and online). The paper claims that SimPEL achieves lower negative log-likelihood in low-data regimes, better uncertainty estimates than FSVGD/SVGD/GREYBOX, and a roughly 2× reduction in hardware episodes for a dynamic parking maneuver.

Significance. If the empirical claims hold, SimPEL is a practically valuable and conceptually clean way to inject first-principles simulator knowledge into BNN inference without fitting the prior to the downstream data. The paper's strengths include a coherent algorithmic derivation (Eqs. 6–7), the use of an external low-fidelity simulator as an input rather than as a fitted component, comparison against several relevant baselines (FSVGD, SVGD, GREYBOX, SysID), and real-hardware RC-car experiments, which are rare in this literature. The provided code link is a useful reproducibility asset. However, the central sample-efficiency claim currently rests on a few unspecified prior hyperparameters and on empirical curves shown without error bars or repeated trials, so the significance cannot be fully assessed from the manuscript as written.

major comments (4)

['Sim-to-Real Gap Prior', Eq. (2)] The GP gap prior is k(x,x')=κ²ρ(||x−x'||/ℓ), with zero mean, and the paper states that κ² and ℓ 'encode prior beliefs' but never reports their values, elicitation, or selection procedure for any experiment. In the low-data regime the posterior is dominated by this prior, so the claimed 'substantially less data' advantage depends directly on these choices: too large κ² degenerates toward FSVGD, while too small κ² or mismatched ℓ can make the prior confidently wrong. The zero-mean assumption is especially consequential because it can only add uncertainty around the simulator output and cannot correct systematic bias until data arrive. A sensitivity analysis over κ² and ℓ, or at minimum a precise statement of how they were set for each benchmark, is required to rule out optimistic results from favorable prior tuning.
[Figs. 4 and 5] The main empirical evidence consists of learning curves plotted without error bars, confidence bands, or any indication of the number of seeds/repeats. The central claims—'SimPEL outperforms all baselines' and 'ca. 2× fewer episodes/hardware experiments'—are not supported by a single run per setting. Particularly for hardware experiments, where noise and run-to-run variability are substantial, the authors should report means and standard errors across multiple trials, or at least show individual runs if repeats are infeasible. Without this, the reported NLL and reward differences cannot be distinguished from stochastic variation.
[Methods, 'Comparison of Score Estimation Techniques', Fig. 6] All main experiments use the Gaussian approximation for the prior score, but Fig. 6 demonstrates that this approximation captures only one mode of a bimodal functional prior, while the ν-method captures both. Since the simulation prior p(h_X) is a mixture of simulator outputs and GP draws, multimodality is plausible in several of the reported systems. The text acknowledges this limitation and advises the ν-method when multimodality is known, but no main-benchmark results with the ν-method are provided. This leaves open the possibility that the reported low-data gains are an artifact of the approximate score rather than a robust property of the proposed prior. At minimum, the authors should report on which systems they checked that the Gaussian approximation is adequate, or include main-result comparisons with the ν-method.
[Algorithm 1 and 'Combining the Stochastic Processes'] The number of prior samples N, the measurement-set size and distribution ζ, and the likelihood noise σ² are all free inputs to SimPEL, but the manuscript does not describe how these were chosen in the experiments or how sensitive the results are to them. The algorithm's pseudocode lists these as inputs, yet the empirical section is silent on their values. This matters because the prior-score estimate—and hence the functional prior actually used—is a finite-sample approximation controlled by N and the measurement points. The authors should report these settings per experiment and, if feasible, a brief sensitivity study for at least one benchmark.

minor comments (6)

[Throughout] Notation is inconsistent: the figure uses 'SimFSVGD' while the text uses 'SimFSVGD[Only sim prior]' and 'SIMFSVGD'; the abstract uses 'substantially less data' without a precise quantity, and the quantitative claim '2× reduction' appears only later. Please unify the notation and give exact numbers in the abstract or first results paragraph.
[Algorithm 1] Typo: 'measurment distribution' should be 'measurement distribution'.
[Introduction, paragraph after Fig. 1] Typo: 'We incoproate the knowledge' should be 'We incorporate the knowledge'.
['Sim-to-Real Gap Prior'] The correlation function ρ is said to be 'e.g., squared exponential or Matérn', but the actual choice used in the experiments is not reported. This is a small but relevant implementation detail for reproducibility.
[Fig. 4 and Fig. 5] The caption of Fig. 5 states part (a) is offline RL and part (b) online RL, but the main text labels them differently and the curves have no axis labels on the right panels. Please make the figure/panel labels consistent and self-contained.
[RC Car System Identification on Hardware] The description says the car has 'a highly dynamics' system; this should read 'highly dynamic'. Also, the reported NLL/RMSE curves would benefit from a statement about how the hardware data were split across repeated trials, if any.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the simulator and GP gap are external priors, and the reported gains are empirical out-of-sample comparisons rather than quantities derived from fitted values.

full rationale

SimPEL's construction takes the low-fidelity simulator g(x, phi) and a zero-mean GP with kernel k(x,x') = kappa^2 rho(||x-x'||/ell) as external inputs in the 'Sim-to-Real Gap Prior' section, not as parameters fitted to the evaluation data. The posterior update (Eq. 6) is the standard FSVGD function-space update, cited to independent work [5,6], and the prior score is estimated from samples drawn from the simulator prior. The likelihood is a standard Gaussian observation model. No equation in the paper defines the prediction in terms of fitted constants, and the central sample-efficiency claims are empirical comparisons on held-out test sets, as the paper states: 'we train the algorithms on train sets of increasing size and have a separate test set on which we evaluate the Negative Log-Likelihood.' The RC-car 'substantially less data' claim is likewise an experimental comparison against baselines, not a reduction. The paper does contain self-citations (e.g., [9,10,30,31,50,58]) but none is load-bearing for the simulator-prior construction; the inference and score-estimation machinery is externally cited. The skeptic's concern about unspecified GP hyperparameters kappa^2 and ell is a robustness or assumption issue, not a circularity, because those values are treated as prior beliefs rather than as fitted predictions. No circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

SimPEL does not derive the prior from data; it packages user-supplied inputs (simulator g, parameter distribution p(phi), GP kernel) into a functional prior. The central claim inherits the quality of these inputs, especially the GP gap model and its hyperparameters.

free parameters (5)

GP output variance κ²
User-chosen prior belief for the magnitude of the sim-to-real gap; not fitted, but directly controls how much the model trusts the simulator.
GP lengthscale ℓ
User-chosen prior belief for the smoothness of the gap; not fitted.
Likelihood noise variance σ²
The Gaussian likelihood N(y|h, σ²I) requires σ²; the paper does not state whether it is known, estimated, or tuned.
Measurement set size and distribution ζ
The functional prior is defined on a measurement set X drawn from ζ; choice of ζ and m affects the prior score.
Number of prior samples N
The prior score is estimated from N samples; larger N reduces Monte Carlo error but costs compute.

assumptions (4)

standard math Kolmogorov extension theorem allows defining the stochastic-process prior via finite-dimensional marginals.
Invoked in Methods, 'BNN Inference in Function Space', to justify working with p(h^X).
domain assumption The sim-to-real gap is a zero-mean GP with a stationary isotropic kernel.
Stated in 'Sim-to-Real Gap Prior' as a modeling choice for generality and interpretability.
domain assumption Output dimensions are independent in the prior, p(h)=∏ p(h_i).
Stated in 'Incorporating Simulators as Functional Priors'; allows treating each output as a scalar function.
ad hoc to paper Approximating the mixture prior score with a single Gaussian is adequate for the tested problems.
Acknowledged in 'Comparison of Score Estimation Techniques'; Gaussian approximation is used in all main experiments although Fig. 6 shows it fails on a bi-modal prior.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Simulation Priors for Data-Efficient Deep Learning." pith.science (2026). https://pith.science/paper/HLL5LGE7

@misc{pith2026250905732,
  author       = {Pith},
  title        = {Pith review of: Simulation Priors for Data-Efficient Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLL5LGE7}},
  note         = {Machine review of arXiv:2509.05732}
}

read the original abstract

How do we enable AI systems to efficiently learn in the real-world? First-principles models are widely used to simulate natural systems, but often fail to capture real-world complexity due to simplifying assumptions. In contrast, deep learning approaches can estimate complex dynamics with minimal assumptions but require large, representative datasets. We propose SimPEL, a method that efficiently combines first-principles models with data-driven learning by using low-fidelity simulators as priors in Bayesian deep learning. This enables SimPEL to benefit from simulator knowledge in low-data regimes and leverage deep learning's flexibility when more data is available, all the while carefully quantifying epistemic uncertainty. We evaluate SimPEL on diverse systems, including biological, agricultural, and robotic domains, showing superior performance in learning complex dynamics. For decision-making, we demonstrate that SimPEL bridges the sim-to-real gap in model-based reinforcement learning. On a high-speed RC car task, SimPEL learns a highly dynamic parking maneuver involving drifting with substantially less data than state-of-the-art baselines. These results highlight the potential of SimPEL for data-efficient learning and control in complex real-world environments.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

81 extracted references · 76 canonical work pages

[1]

Cell systems11(3), 252–271 (2020)

Dibaeinia, P., Sinha, S.: Sergio: a single-cell expression simulator guided by gene regulatory networks. Cell systems11(3), 252–271 (2020)

2020
[2]

Wageningen University and Research, Wageningen, The Netherlands (2000)

Tap, F.: Economics-based Optimal Control of Greenhouse Tomato Crop Production. Wageningen University and Research, Wageningen, The Netherlands (2000)

work page 2000
[3]

Journal of political economy (1973)

Black, F., Scholes, M.: The pricing of options and corporate liabilities. Journal of political economy (1973)

work page 1973
[4]

In: Siciliano, B., Khatib, O

Siciliano, B., Khatib, O.: Robotics and the handbook. In: Siciliano, B., Khatib, O. (eds.) Springer Handbook of Robotics, pp. 1–6. Springer, Cham, Switzerland (2016)

work page 2016
[5]

In: International Conference on Learning Representations (2019)

Wang, Z., Ren, T., Zhu, J., Zhang, B.: Function space particle optimization for Bayesian neural networks. In: International Conference on Learning Representations (2019)

work page 2019
[6]

In: International Conference on Learning Representations (2019)

Sun, S., Zhang, G., Shi, J., Grosse, R.: Functional variational bayesian neural networks. In: International Conference on Learning Representations (2019)

work page 2019
[7]

In: Advances in Neural Information Processing Systems (2018)

Chua, K., Calandra, R., McAllister, R., Levine, S.: Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models. In: Advances in Neural Information Processing Systems (2018)

work page 2018
[8]

Advances in Neural Information Processing Systems (2020)

Curi, S., Berkenkamp, F., Krause, A.: Efficient model-based reinforcement learn- ing through optimistic policy search and planning. Advances in Neural Information Processing Systems (2020)

work page 2020

Show all 81 references

[9]

Advances in Neural Information Processing Systems (2024)

Sukhija, B., Treven, L., Sancaktar, C., Blaes, S., Coros, S., Krause, A.: Optimistic active exploration of dynamical systems. Advances in Neural Information Processing Systems (2024)

2024
[10]

In: 7th Robot Learning Workshop: Towards Robots with Human-Level Abilities (2025)

Sukhija, B., Treven, L., Sferrazza, C., Dorfler, F., Abbeel, P., Krause, A.: Optimism via intrinsic rewards: Scalable and principled exploration for model-based reinforce- ment learning. In: 7th Robot Learning Workshop: Towards Robots with Human-Level Abilities (2025)

2025
[11]

Settles, B.: Active learning literature survey (2009)

2009
[12]

Journal of Machine Learning Research9(2) (2008)

Krause, A., Singh, A., Guestrin, C.: Near-optimal sensor placements in gaussian pro- cesses: Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research9(2) (2008)

2008
[13]

Machine learning80, 111–139 (2010)

Balcan, M.-F., Hanneke, S., Vaughan, J.W.: The true sample complexity of active learning. Machine learning80, 111–139 (2010)

2010
[14]

Advances in Neural Information Processing Systems (2024) 18

H ¨ubotter, J., Sukhija, B., Treven, L., As, Y ., Krause, A.: Transductive active learning: Theory and applications. Advances in Neural Information Processing Systems (2024) 18

2024
[15]

In: ICML (2015)

Sui, Y ., Gotovos, A., Burdick, J., Krause, A.: Safe exploration for optimization with gaussian processes. In: ICML (2015)

2015
[16]

https://openreview.net/pdf? id=fdZvTFn8Yq (2021)

Berkenkamp, F., Eivazi, A., Grossberger, L., Skubch, K., Spitz, J., Daniel, C., Falkner, S.: Probabilistic Meta-Learning for Bayesian Optimization. https://openreview.net/pdf? id=fdZvTFn8Yq (2021)

2021
[17]

In: IEEE International Conference on Connected Vehicles and Expo (2019)

Wischnewski, A., Betz, J., Lohmann, B.: A model-free algorithm to safely approach the handling limit of an autonomous racecar. In: IEEE International Conference on Connected Vehicles and Expo (2019)

2019
[18]

In: IJCAI (2019)

Fiducioso, M., Curi, S., Schumacher, B., Gwerder, M., Krause, A.: Safe contextual Bayesian optimization for sustainable room temperature PID control tuning. In: IJCAI (2019)

2019
[19]

Artificial Intelligence320(2023)

Sukhija, B., Turchetta, M., Lindner, D., Krause, A., Trimpe, S., Baumann, D.: Gosafeopt: Scalable safe exploration for global optimization of dynamical systems. Artificial Intelligence320(2023)

2023
[20]

In: Conference on Robot Learning, pp

Widmer, D., Kang, D., Sukhija, B., H ¨ubotter, J., Krause, A., Coros, S.: Tuning legged locomotion controllers via safe bayesian optimization. In: Conference on Robot Learning, pp. 2444–2464 (2023). PMLR

2023
[21]

Science Robotics (2019)

Hwangbo, J., Lee, J., Dosovitskiy, A., Bellicoso, D., Tsounis, V ., Vladlen Koltun, M.H.: Learning agile and dynamic motor skills for legged robots. Science Robotics (2019)

2019
[22]

IEEE Transactions on Neural Networks (1990)

Narendra, K.S., Parthasarathy, K.: Identification and control of dynamical systems using neural networks. IEEE Transactions on Neural Networks (1990)

1990
[23]

IFAC Proceedings V olumes (1994)

Sj ¨oberg, J., Hjalmarsson, H., Ljung, L.: Neural networks in system identification. IFAC Proceedings V olumes (1994)

1994
[24]

In: ICRA (2018)

Nagabandi, A., Kahn, G., Fearing, R.S., Levine, S.: Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning. In: ICRA (2018)

2018
[25]

In: RoboSoft (2020)

Bern, J.M., Schnider, Y ., Banzet, P., Kumar, N., Coros, S.: Soft Robot Control With a Learned Differentiable Model. In: RoboSoft (2020)

2020
[26]

In: ICRA (2023)

Sukhija, B., K ¨ohler, N., Zamora, M., Zimmermann, S., Curi, S., Krause, A., Coros, S.: Gradient-based trajectory optimization with learned dynamics. In: ICRA (2023)

2023
[27]

Advances in neural information processing systems25 (2012)

Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. Advances in neural information processing systems25 (2012)

2012
[28]

Machine Learning and Knowledge Extraction6(4), 2639–2658 (2024) 19

Ngartera, L., Issaka, M.A., Nadarajah, S.: Application of bayesian neural networks in healthcare: Three case studies. Machine Learning and Knowledge Extraction6(4), 2639–2658 (2024) 19

2024
[29]

In: International Conference on Learning Representations (2020)

Sekar, R., Rybkin, O., Daniilidis, K., Abbeel, P., Hafner, D., Pathak, D.: Planning to explore via self-supervised world models. In: International Conference on Learning Representations (2020)

2020
[30]

Advances in Neural Information Processing Systems (2024)

Treven, L., H ¨ubotter, J., Sukhija, B., D ¨orfler, F., Krause, A.: Efficient exploration in continuous-time model-based reinforcement learning. Advances in Neural Information Processing Systems (2024)

2024
[31]

Uncertainty in Artificial Intelligence (2023)

Rothfuss, J., Sukhija, B., Birchler, T., Kassraie, P., Krause, A.: Hallucinated adversarial control for conservative offline policy evaluation. Uncertainty in Artificial Intelligence (2023)

2023
[32]

PhD thesis, California Institute of Technology (1992)

MacKay, D.J.: Bayesian methods for adaptive models. PhD thesis, California Institute of Technology (1992)

1992
[33]

Springer, New York, NY (2012)

Neal, R.M.: Bayesian Learning for Neural Networks. Springer, New York, NY (2012)

2012
[34]

NeurIPS (2011)

Graves, A.: Practical variational inference for neural networks. NeurIPS (2011)

2011
[35]

In: ICML (2015)

Blundell, C., Cornebise, J., Kavukcuoglu, K., Wierstra, D.: Weight uncertainty in neural network. In: ICML (2015)

2015
[36]

Case Studies in Applied Bayesian Data Science (2020)

Goan, E., Fookes, C.: Bayesian neural networks: An introduction and survey. Case Studies in Applied Bayesian Data Science (2020)

2020
[37]

In: ICML (2016)

Gal, Y ., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: ICML (2016)

2016
[38]

arXiv (2018)

Zellers, R., Bisk, Y ., Schwartz, R., Choi, Y .: Swag: A large-scale adversarial dataset for grounded commonsense inference. arXiv (2018)

2018
[39]

NeurIPS (2017)

Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncer- tainty estimation using deep ensembles. NeurIPS (2017)

2017
[40]

In: Advances in Neural Information Processing Systems (2016)

Liu, Q., Wang, D.: Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm. In: Advances in Neural Information Processing Systems (2016)

2016
[41]

NeurIPS (2021)

D’Angelo, F., Fortuin, V .: Repulsive deep ensembles are bayesian. NeurIPS (2021)

2021
[42]

Journal of Machine Learning Research (2022)

Tran, B.-H., Rossi, S., Milios, D., Filippone, M.: All you need is a good functional prior for bayesian deep learning. Journal of Machine Learning Research (2022)

2022
[43]

In: Artificial Intelligence and Statistics, pp

Duvenaud, D., Rippel, O., Adams, R., Ghahramani, Z.: Avoiding pathologies in very deep networks. In: Artificial Intelligence and Statistics, pp. 202–210 (2014). PMLR

2014
[44]

arXiv preprint arXiv:1804.11271 (2018) 20

Matthews, A.G.d.G., Rowland, M., Hron, J., Turner, R.E., Ghahramani, Z.: Gaussian process behaviour in wide deep neural networks. arXiv preprint arXiv:1804.11271 (2018) 20

2018 arXiv
[45]

International Statistical Review 90(3), 563–591 (2022)

Fortuin, V .: Priors in bayesian deep learning: A review. International Statistical Review 90(3), 563–591 (2022)

2022
[46]

In: International Conference on Machine Learning (2018)

Shi, J., Sun, S., Zhu, J.: A spectral approach to gradient estimation for implicit distributions. In: International Conference on Machine Learning (2018)

2018
[47]

In: International Conference on Machine Learning, pp

Zhou, Y ., Shi, J., Zhu, J.: Nonparametric score estimators. In: International Conference on Machine Learning, pp. 11513–11522 (2020)

2020
[48]

In: Conference on Robot Learning, pp

Clavera, I., Rothfuss, J., Schulman, J., Fujita, Y ., Asfour, T., Abbeel, P.: Model- based reinforcement learning via meta-policy optimization. In: Conference on Robot Learning, pp. 617–629 (2018). PMLR

2018
[49]

In: IEEE SSCI (2020)

Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In: IEEE SSCI (2020)

2020
[50]

Advances in Neural Information Processing Systems34, 280–293 (2021)

Rothfuss, J., Heyn, D., Krause, A.,et al.: Meta-learning reliable priors in the function space. Advances in Neural Information Processing Systems34, 280–293 (2021)

2021
[51]

In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain random- ization for transferring deep neural networks from simulation to the real world. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 23–30 (2017). IEEE

2017
[52]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

Yue, X., Zhang, Y ., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision,...

2019
[53]

In: Conference on Robot Learning, pp

Mehta, B., Diaz, M., Golemo, F., Pal, C.J., Paull, L.: Active domain randomization. In: Conference on Robot Learning, pp. 1162–1176 (2020). PMLR

2020
[54]

arXiv preprint arXiv:2110.03239 (2021)

Chen, X., Hu, J., Jin, C., Li, L., Wang, L.: Understanding domain randomization for sim-to-real transfer. arXiv preprint arXiv:2110.03239 (2021)

2021 arXiv
[55]

In: International Conference on Learning Representations (2017)

Duan, Y ., Schulman, J., Chen, X., Bartlett, P.L., Sutskever, I., Abbeel, P.: Rl2: Fast reinforcement learning via slow reinforcement learning. In: International Conference on Learning Representations (2017)

2017
[56]

In: International Conference on Machine Learning (2017)

Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International Conference on Machine Learning (2017)

2017
[57]

In: Advances in Neural Information Processing Systems (2018)

Finn, C., Xu, K., Levine, S.: Probabilistic model-agnostic meta-learning. In: Advances in Neural Information Processing Systems (2018)

2018
[58]

arXiv (2023) 21

Bhardwaj, A., Rothfuss, J., Sukhija, B., As, Y ., Hutter, M., Coros, S., Krause, A.: Data- efficient task generalization via probabilistic model-based meta reinforcement learning. arXiv (2023) 21

2023
[59]

Cambridge University Press (2008)

Taubes, C.H.: Modeling differential equations in biology. Cambridge University Press (2008)

2008
[60]

Lecture Notes (2018)

Hutter, M., Stastny, T.: Robot dynamics. Lecture Notes (2018)

2018
[61]

In: Webster, J.G

Ljung, L.: System identification. In: Webster, J.G. (ed.) Wiley Encyclopedia of Electrical and Electronics Engineering. John Wiley & Sons, New York, NY (1999). https://doi. org/10.1002/047134608X.W1037

1999 doi
[62]

science (2013)

Li, C., Zhang, T., Goldman, D.I.: A terradynamics of legged locomotion on granular media. science (2013)

2013
[63]

In: IROS, pp

Moeckel, R., Perov, Y .N., Nguyen, A.T., Vespignani, M., Bonardi, S., Pouya, S., Sproe- witz, A., Kieboom, J., Wilhelm, F., Ijspeert, A.J.: Gait optimization for roombots modular robots—matching simulation and reality. In: IROS, pp. 3265–3272 (2013)

2013
[64]

In: IROS, pp

Tan, J., Xie, Z., Boots, B., Liu, C.K.: Simulation-based design of dynamic controllers for humanoid balancing. In: IROS, pp. 2729–2736 (2016). IEEE

2016
[65]

IJCAI (2017)

Zhu, S., Kimmel, A., Bekris, K.E., Boularias, A.: Model identification via physics engines for improved policy search. IJCAI (2017)

2017
[66]

In: ICRA (2021)

Zimmermann, S., Poranne, R., Coros, S.: Go fetch! - dynamic grasps using boston dynamics spot with external robotic arm. In: ICRA (2021)

2021
[67]

In: ICRA (2013)

Pastor, P., Kalakrishnan, M., Binney, J., Kelly, J., Righetti, L., Sukhatme, G., Schaal, S.: Learning task error models for manipulation. In: ICRA (2013)

2013
[68]

In: ICRA (2015)

Ha, S., Yamane, K.: Reducing hardware experiments for model learning and policy optimization. In: ICRA (2015)

2015
[69]

IFAC (2017)

Huang, H., He, W., Zhang, S.: Neural network control of a robotic manipulator with time-varying output constraints by state feedback. IFAC (2017)

2017
[70]

In: UR (2023)

Schperberg, A., Tanaka, Y ., Xu, F., Menner, M., Hong, D.: Real-to-sim: Predicting resid- ual errors of robotic systems with sparse data using a learning-based unscented kalman filter. In: UR (2023)

2023
[71]

Renewable and Sustainable Energy Reviews146, 111174 (2021)

Li, Y ., O’Neill, Z., Zhang, L., Chen, J., Im, P., DeGraw, J.: Grey-box modeling and application for building energy simulations-a critical review. Renewable and Sustainable Energy Reviews146, 111174 (2021)

2021
[72]

Journal of Field Robotics (2020)

Kabzan, J., Valls, M.I., Reijgwart, V .J., Hendrikx, H.F., Ehmke, C., Prajapat, M., B¨uhler, A., Gosala, N., Gupta, M., Sivanesan, R., et al.: Amz driverless: The full autonomous racing system. Journal of Field Robotics (2020)

2020
[73]

Computers and electronics in agriculture70(1), 105–116 (2010)

Fitz-Rodr ´ıguez, E., Kubota, C., Giacomelli, G.A., Tignor, M.E., Wilson, S.B., McMa- hon, M.: Dynamic modeling and simulation of greenhouse environments under several 22 scenarios: A web-based application. Computers and electronics in agriculture70(1), 105–116 (2010)

2010
[74]

Software Impacts6, 100022 (2020)

Tunyasuvunakool, S., Muldal, A., Doron, Y ., Liu, S., Bohez, S., Merel, J., Erez, T., Lillicrap, T., Heess, N., Tassa, Y .: dmcontrol: Software and tasks for continuous control. Software Impacts6, 100022 (2020)

2020
[75]

Vehicle system dynamics, 1–18 (1992)

Pacejka, H.B., Bakker, E.: The magic formula tyre model. Vehicle system dynamics, 1–18 (1992)

1992
[76]

In: International Conference on Machine Learning, pp

Haarnoja, T., Zhou, A., Abbeel, P., Levine, S.: Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor. In: International Conference on Machine Learning, pp. 1861–1870 (2018). Pmlr

2018
[77]

Journal of the American Statistical Association112, 859–877 (2017)

Blei, D.M., Kucukelbir, A., McAuliffe, J.D.: Variational inference: A review for statisticians. Journal of the American Statistical Association112, 859–877 (2017)

2017
[78]

Chen, C., Zhang, R., Wang, W., Li, B., Chen, L.: A unified particle-optimization framework for scalable bayesian sampling (2018)

2018
[79]

Advances in Neural Information Processing Systems34, 20851–20862 (2021)

Noci, L., Bachmann, G., Roth, K., Nowozin, S., Hofmann, T.: Precise characterization of the prior predictive distribution of deep relu networks. Advances in Neural Information Processing Systems34, 20851–20862 (2021)

2021
[80]

MIT Press, Cambridge, MA (2006)

Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press, Cambridge, MA (2006). http://gaussianprocess.org/gpml/

2006
[81]

82, (2000)

Øksendal, B.: Stochastic Differential Equations: An Introduction with Applications vol. 82, (2000). https://doi.org/10.1007/978-3-662-03185-8 23

2000 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Cell systems11(3), 252–271 (2020)

Dibaeinia, P., Sinha, S.: Sergio: a single-cell expression simulator guided by gene regulatory networks. Cell systems11(3), 252–271 (2020)

2020

[2] [2]

Wageningen University and Research, Wageningen, The Netherlands (2000)

Tap, F.: Economics-based Optimal Control of Greenhouse Tomato Crop Production. Wageningen University and Research, Wageningen, The Netherlands (2000)

work page 2000

[3] [3]

Journal of political economy (1973)

Black, F., Scholes, M.: The pricing of options and corporate liabilities. Journal of political economy (1973)

work page 1973

[4] [4]

In: Siciliano, B., Khatib, O

Siciliano, B., Khatib, O.: Robotics and the handbook. In: Siciliano, B., Khatib, O. (eds.) Springer Handbook of Robotics, pp. 1–6. Springer, Cham, Switzerland (2016)

work page 2016

[5] [5]

In: International Conference on Learning Representations (2019)

Wang, Z., Ren, T., Zhu, J., Zhang, B.: Function space particle optimization for Bayesian neural networks. In: International Conference on Learning Representations (2019)

work page 2019

[6] [6]

In: International Conference on Learning Representations (2019)

Sun, S., Zhang, G., Shi, J., Grosse, R.: Functional variational bayesian neural networks. In: International Conference on Learning Representations (2019)

work page 2019

[7] [7]

In: Advances in Neural Information Processing Systems (2018)

Chua, K., Calandra, R., McAllister, R., Levine, S.: Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models. In: Advances in Neural Information Processing Systems (2018)

work page 2018

[8] [8]

Advances in Neural Information Processing Systems (2020)

Curi, S., Berkenkamp, F., Krause, A.: Efficient model-based reinforcement learn- ing through optimistic policy search and planning. Advances in Neural Information Processing Systems (2020)

work page 2020