Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Semi-supervised imitation succeeds because inverse dynamics models predict actions from state pairs that are simpler and less stochastic than the expert policy's state-only view, and two major IDM-based methods learn the same policy in the

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

2026-08-03 05:15 UTC pith:P46YUX63

load-bearing objection A clean formal unification of VM-IDM and IDM labeling, with a plausible but not fully pinned-down explanation of IDM sample efficiency; deserves refereeing. the 3 major comments →

arxiv 2602.02762 v2 pith:P46YUX63 submitted 2026-02-02 cs.LG

On the Sample Efficiency of Inverse Dynamics Models for Semi-Supervised Imitation Learning

Sacha Morin , Moonsub Byeon , Alexia Jolicoeur-Martineau , S\'ebastien Lachapelle This is my paper
classification cs.LG
keywords semi-supervised imitation learninginverse dynamics modelsbehavior cloningIDM labelingVM-IDMsample efficiencylatent action policiesProcGen
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 explain why semi-supervised imitation learning works so well when an inverse dynamics model (IDM) predicts the action from two consecutive states. It shows that the two dominant IDM-based strategies—pairing the IDM with a video model, and using the IDM to label action-free data for behavior cloning—learn exactly the same policy in the infinite-unlabeled-data limit. It then argues, with statistical-learning-theory reasoning and experiments, that the IDM-based policy beats plain behavior cloning because the ground-truth IDM is usually a simpler and less stochastic function than the expert policy. If the explanation is right, practitioners can predict which environments will show the largest gains from IDM-based methods.

Core claim

The paper's central discovery is a unification and an explanation. It proves that with infinitely many unlabeled transitions and sufficient capacity, VM-IDM and IDM labeling recover the same object: the IDM-based policy, formed by marginalizing the inverse dynamics model over next states from the expert's transition distribution. This policy is consistent, recovering the expert when action labels are also infinite. The paper formalizes why IDM learning is more label-efficient: the ground-truth IDM h*(a|s,s') tends to be less complex and less stochastic than the expert policy π*(a|s), and a chain-rule inequality (6) bounds the policy's KL error by the IDM's KL error. Maze, ProcGen, and robot

What carries the argument

The load-bearing object is the IDM-based policy, π̂_{v*,ĥ}(a|s) = ∫ ĥ(a|s,s') v*(s'|s) ds', an inverse dynamics model composed with the expert's next-state distribution. Two mechanisms carry the argument. First, a limit equivalence: under infinite unlabeled data and sufficient capacity, VM-IDM and IDM labeling minimize objectives whose unique solution is this same composed policy. Second, a chain-rule inequality (6), proved in the appendix, shows the policy's KL error against the expert is no larger than the IDM's own KL error, transferring the IDM's complexity and stochasticity advantages to the composed policy.

Load-bearing premise

The argument rests on the unlabeled transitions being i.i.d. samples from the same expert distribution as the labeled ones, plus the unproven regularity that in the target environments the ground-truth IDM really is simpler and less stochastic than the expert policy.

What would settle it

Build a benchmark where the expert policy is a simple, deterministic function but the IDM is hard—for example, stochastic transitions in which several actions can lead to the same next state, making h* stochastic and complex. If behavior cloning then matches or beats IDM-based methods on label efficiency, the complexity/stochasticity explanation is wrong.

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

If this is right

  • VM-IDM and IDM labeling are not competing methods but two estimators of the same target policy; any difference between them is a finite-data artifact, not a fundamental distinction.
  • The label-efficiency gap between IDM-based policies and behavior cloning is predictable: it should be large when the next state disambiguates the action and small when the expert policy is already simple or the dynamics are noisy or aliased.
  • Decoding a latent inverse dynamics model (the proposed LAPO+) is more label-efficient than decoding a latent policy (LAPO), as shown across the 16 ProcGen environments.
  • When the unlabeled dataset is finite, the equivalence breaks down; the paper's manipulation experiments suggest IDM labeling trails VM-IDM modestly, indicating the video model's generalization matters in practice.

Where Pith is reading between the lines

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

  • A testable extension: estimate the conditional entropies H(a|s,s') and H(a|s) from a small labeled sample in any new environment to predict in advance how much IDM labeling will beat behavior cloning; the paper's Figure 5b supports the correlation only imperfectly.
  • The equivalence result suggests a unified training objective: instead of training a video model and an IDM separately, one could directly fit a joint conditional model p(a,s'|s) and extract the policy by marginalization—standard UVA architectures begin to make this feasible.
  • For robotics, the framework implies that the value of action-free video depends on whether the next state reveals the action; in tasks with unobserved confounders or external disturbance, IDM-based methods should lose their edge and behavior cloning may be preferable.

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 / 6 minor

Summary. This paper studies semi-supervised imitation learning (SSIL) with inverse dynamics models (IDMs). It proves that, in the limit of infinite unlabeled data, sufficient model capacity, and global optimization, the VM-IDM policy and the IDM-labeling policy converge to the same policy, called the IDM-based policy, and that this policy is consistent with the expert. The paper then proposes an explanation for the observed sample-efficiency advantage of IDM learning over behavior cloning: the ground-truth IDM is typically less complex and less stochastic than the expert policy, formalized by comparing KL divergences. This explanation is tested in synthetic maze/grid environments and analyzed on ProcGen, and the paper introduces LAPO+, a modification of LAPO that decodes the latent IDM rather than the latent policy, plus UVA-based VM-IDM experiments on Push-T and LIBERO.

Significance. If the empirical claims hold, the paper provides a useful framework for predicting when IDM-based SSIL methods outperform BC, and a practical algorithmic improvement in LAPO+. The theoretical contribution is solid: the Section 3 equivalence is a clean, self-contained derivation, and the proof of inequality (6) in Appendix A is correct. The toy experiments in Section 4 isolate the proposed complexity and stochasticity mechanisms well. However, the direct formal premise of the explanation, inequality (5), is never measured on the benchmarks; the ProcGen evidence relies on a post hoc hand-labeled dynamics split and on entropies of trained models as proxies. These gaps leave the central causal explanation under-supported. LAPO+ is plausible and consistently favorable in the reported averages, but the lack of uncertainty quantification tempers the strength of that claim.

major comments (3)
  1. [§4, Eq. (5)] The central explanatory premise is inequality (5), but neither side of this inequality is ever estimated in any experiment. Section 4 demonstrates in synthetic mazes that low-capacity IDMs achieve better accuracy than low-capacity BC, which is consistent with (5), but it does not measure the actual KL divergences involved. Section 5 uses return gaps and proxies. Because (5) is load-bearing for the claim that IDM-based policies outperform BC due to the sample efficiency of IDM learning, please either (a) directly estimate D_KL(h*||ĥ) and D_KL(π*||π̂_BC) on held-out data at matched DL sizes on at least a subset of the benchmarks, or (b) explicitly demote (5) to a conjecture and present the benchmark evidence as correlational. Without this, alternative explanations—e.g., the IDM's smaller effective output space or the use of DU—are not ruled out.
  2. [§5.1, Fig. 5a, App. C.2.2] The simple/complex dynamics classification is a free parameter in the analysis: the categories in Appendix C.2.2 appear to be assigned by the authors, and the text does not report that the classification was made before inspecting the return gaps. This creates a selection risk for the correlation in Figure 5a. Please provide a pre-defined, deterministic classification rule (or independent annotations with agreement), and show that the qualitative pattern is robust to plausible reclassifications (e.g., moving chaser or leaper between categories).
  3. [§5.1, Fig. 5b] The stochasticity analysis uses the entropies of trained models \hat{\pi}_BC and \hat{h} as proxies for the ground-truth entropies H(\pi*) and H(h*). These quantities are influenced by model capacity, optimization, and dataset sizes, so the observation H(\hat{\pi}_BC)>H(\hat{h}) across all environments does not establish that h* is less stochastic than \pi*. The phrase 'as hinted in Section 4.2' overstates the support. Please temper the claim to 'the trained IDM has lower entropy than the trained BC policy' and, where possible, validate the proxy on environments with known ground-truth stochasticity (as in Figure 3).
minor comments (6)
  1. [App. C.1.1] In the definition of the linear IDM, the bias term is named b but should be c (or consistently named); the current text has 'b\inR^4' after defining c for the IDM.
  2. [Fig. 5b] The axis label appears to be missing a minus sign between H(BC) and H(h); as printed it is ambiguous.
  3. [§5.1] The claim that LAPO+ 'consistently outperforms' LAPO is based on averages over 3 seeds without error bars. Please include seed-level spread or a significance test to support this claim.
  4. [§2.3.2] The IDM labeling objective is written as −(1/N)Sigma E_{\hat{h}} log \hat{\pi}; clarify that \hat{h} is frozen and that the expectation is with respect to the empirical DU distribution, not the true p_\pi*.
  5. [§4.2] The notation H(a|s) as −E_{s,a} log \pi*(a|s) is slightly nonstandard; use H(\pi*(·|s)) or write the conditional entropy explicitly to avoid confusion.
  6. [Appendix A] The distribution p_{v*,\hat{h}}(a,s'|s) is defined as a product; it may help to note explicitly that this is a normalized joint distribution.

Circularity Check

0 steps flagged

No significant circularity: Section 3 is a derived limiting equivalence, Section 4's sample-efficiency explanation is under-tested but not definitionally forced, and there is no load-bearing self-citation chain.

full rationale

The derivation chain is self-contained. In Section 3, the VM-IDM policy is defined by the marginal (4), and the IDM-labeling objective in the infinite-unlabeled-data limit is shown by direct algebra to be the cross-entropy to that same marginal, so its global minimizer is the IDM-based policy. This is a mathematical identity derived from the stated objectives, not a parameter fitted to a target claim, and it does not smuggle in the conclusion. Inequality (6) is proved in Appendix A via the KL chain rule and is not assumed. Section 4 formalizes the previously observed sample-efficiency premise as inequality (5), but never claims to derive (5) from first principles; it explicitly treats it as a premise and then investigates mechanistic explanations (complexity and stochasticity). The toy-maze constructions in Appendices B.1 and B.2 show by construction that h* is simple in those environments, which is valid supporting evidence for the claimed causal story. The ProcGen 'simple/complex' split in Appendix C.2.2 is based on environment mechanics (momentum, gravity, moving platforms) rather than on the measured return gaps, and the entropy analysis in Figure 5b is framed as a proxy with an admitted imperfect correlation. These are evidential limitations, not circular reductions. The LAPO+ variant is a new algorithmic proposal whose improvement over LAPO is an empirical result, not entailed by the training objective. Section 6 candidly acknowledges the curated-DU limitation and the fact that finite-DU differences between VM-IDM and IDM labeling were not studied. There are no self-citations, no imported uniqueness theorems, and no fitted quantities renamed as predictions. The central formal results are honest derivations; the explanatory claims are weaker than the formal ones but not circular.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central derivation rests on standard KL/probability assumptions plus a few domain-specific regularities: same-distribution unlabeled data, deterministic maze dynamics, and the imported simplicity-bias hypothesis. The only hand-chosen quantity that directly shapes a headline result is the simple/complex environment split in Figure 5a.

free parameters (1)
  • Environment complexity classification (simple vs complex) used in Figure 5a = 9 simple / 7 complex hand-labeled ProcGen environments
    Appendix C.2.2 assigns each environment to simple or complex dynamics by hand (momentum, gravity, moving platforms). This labeling is the independent variable behind the correlation in Figure 5a; it is not derived from a systematic algorithm and could be adjusted to change the apparent trend.
axioms (5)
  • domain assumption DU is sampled i.i.d. from the same expert transition distribution p_π*(s,a,s') as DL
    Section 2.3 states this assumption; Section 6 Discussion admits it is a limitation and that a more practical setting would have DU from a different environment/expert.
  • standard math With infinite data and sufficient capacity, ERM recovers the ground-truth conditional distributions (KL minimizers)
    Used throughout Section 3 and Appendix A for the equivalence and consistency results.
  • domain assumption Neural networks exhibit a simplicity bias favoring 'simpler' interpolating functions, which the paper imports from the cited literature
    Invoked in Section 4.1.1 to explain why a high-capacity VM-IDM can beat a high-capacity BC despite the IDM having twice the input dimension; the paper does not measure this bias in its own experiments.
  • domain assumption The variance/stochasticity arguments from linear regression (Gauss-Markov) and classification (Tsybakov noise) transfer to deep-network IDM/BC learning
    Section 4.2 uses these classical results as motivation for Claim 2; the transfer to the neural-net settings tested is an analogy, not a theorem.
  • domain assumption Maze/grid experiments have deterministic dynamics and a deterministic expert that never collides with walls
    Appendix C.1.2 defines the environment; the explicit linear/convolutional constructions of h* in Appendix B depend on this determinism and on the expert avoiding collisions.

pith-pipeline@v1.3.0-alltime-deepseek · 20199 in / 15679 out tokens · 149730 ms · 2026-08-03T05:15:04.080496+00:00 · methodology

0 comments
read the original abstract

Semi-supervised imitation learning (SSIL) consists in learning a policy from a small dataset of action-labeled trajectories and a much larger dataset of action-free trajectories. Some SSIL methods learn an inverse dynamics model (IDM) to predict the action from the current state and the next state. An IDM can act as a policy when paired with a video model (VM-IDM) or as a label generator to perform behavior cloning on action-free data (IDM labeling). In this work, we first show that VM-IDM and IDM labeling learn the same policy in a limit case, which we call the IDM-based policy. We then argue that the previously observed advantage of IDM-based policies over behavior cloning is due to the superior sample efficiency of IDM learning, which we attribute to two causes: (i) the ground-truth IDM tends to be contained in a lower complexity hypothesis class relative to the expert policy, and (ii) the ground-truth IDM is often less stochastic than the expert policy. We argue these claims based on insights from statistical learning theory and novel experiments, including a study of IDM-based policies using recent architectures for unified video-action prediction (UVA). Motivated by these insights, we finally propose an improved version of the existing LAPO algorithm for latent action policy learning. We experiment on the Procgen, Push-T and LIBERO benchmarks.

Figures

Figures reproduced from arXiv: 2602.02762 by Alexia Jolicoeur-Martineau, Moonsub Byeon, Sacha Morin, S\'ebastien Lachapelle.

Figure 1
Figure 1. Figure 1: Varying the environment complexity. Comparing BC and VM∗ -IDM, with different architectures for πˆ and hˆ, respec￾tively: LC is a linear classifier, 5L MLP is a 5-layer multilayer perceptron and nL CNN is an n-layer convolutional neural network. See Appendix C.1 for details. Averaging over 5 seeds. In other words, expressing π ∗ requires a more com￾plex/expressive model than what is needed to express h ∗ w… view at source ↗
Figure 2
Figure 2. Figure 2: Varying the number of goals. BCG is behavior cloning with goal conditioning, VM∗ G-IDM is VM∗ -IDM where the VM is goal conditioned and the IDM is not; and VM∗ G-IDMG is when both the VM and the IDM are goal conditioned. See Appendix C.1 for details. Averaging over 5 seeds. out impacting that of the IDM h ∗ , leading to IDM learning having a statistical advantage over BC. We now argue that the diversity of… view at source ↗
Figure 4
Figure 4. Figure 4: Normalized maximum return achieved during training by BC, IDM labeling, LAPO and LAPO+ in the 16 Atari-like ProcGen environments (Cobbe et al., 2019), varying the number of action-labeled transitions in DL. We show the average of 3 seeds, with DL being randomly sampled from the whole datasets each time. IDM labeling, LAPO and LAPO+ use all 2.6M transitions without action labels as DU . and using zi ∼ h˜(· … view at source ↗
Figure 5
Figure 5. Figure 5: We study the complexity and stochasticity factors from Section 4 across the 16 ProcGen environments. As an aggregate score, we consider the mean return gap r(ˆπIDM) − r(ˆπBC ) be￾tween IDM labeling and BC, which we average over DL sizes ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Policy performance on Push-T (Chi et al., 2025), which requires pushing a T-shaped block in a target position with a cir￾cular end-effector, and Libero10 (Liu et al., 2023), a set of 10 long-horizon robot manipulation tasks. For DL, we use 10 demon￾strations for Push-T and 2 demonstrations per task for Libero10, representing approximately 5% of available demonstrations, which are all included without actio… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Latent Geometry Beyond Search: Amortizing Planning in World Models

    cs.RO 2026-05 unverdicted novelty 6.0

    In regularized latent spaces of world models, planning can be amortized into a goal-conditioned inverse dynamics model that matches CEM performance at 100-130x lower per-decision cost.

  2. Latent Geometry Beyond Search: Amortizing Planning in World Models

    cs.RO 2026-05 unverdicted novelty 6.0

    A Goal-Conditioned Inverse Dynamics Model amortizes planning in pretrained world model latents, matching or exceeding CEM in seven of eight settings at 100-130x lower per-decision cost.

Reference graph

Works this paper leans on

54 extracted references · 14 linked inside Pith · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Saxe, A

    Advani, M. and Saxe, A. High-dimensional dynamics of generalization error in neural networks. Neural Networks, 2017

    2017

  3. [3]

    Video pretraining (vpt): Learning to act by watching unlabeled online videos

    Baker, B., Akkaya, I., Zhokov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., and Clune, J. Video pretraining (vpt): Learning to act by watching unlabeled online videos. Advances in Neural Information Processing Systems, 35: 0 24639--24654, 2022

    2022

  4. [4]

    Bartlett, P. L. and Mendelson, S. Rademacher and gaussian complexities: risk bounds and structural results. J. Mach. Learn. Res., 2003

    2003

  5. [5]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Bengio, Y., L \'e onard, N., and Courville, A. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013

    Pith/arXiv arXiv 2013

  6. [6]

    Bishop, C. M. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer, 2007

    2007

  7. [7]

    D., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., Aytar, Y., Bechtle, S

    Bruce, J., Dennis, M. D., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., Aytar, Y., Bechtle, S. M. E., Behbahani, F., Chan, S. C., Heess, N., Gonzalez, L., Osindero, S., Ozair, S., Reed, S., Zhang, J., Zolna, K., Clune, J., de Freitas, N., Singh, S., and Rockt \"a schel, T. Genie: Generative intera...

    2024

  8. [8]

    S., Brutzkus, A., Srebro, N., and Soudry, D

    Buzaglo, G., Harel, I., Nacson, M. S., Brutzkus, A., Srebro, N., and Soudry, D. How uniform random weights induce non-uniform bias: Typical interpolating neural networks generalize with narrow teachers. In Proceedings of the 41st International Conference on Machine Learning, 2024

    2024

  9. [9]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., and Song, S. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 44 0 (10-11): 0 1684--1704, 2025

    2025

  10. [10]

    Y., Bansal, A., Geiping, J., Goldblum, M., and Goldstein, T

    Chiang, P., Ni, R., Miller, D. Y., Bansal, A., Geiping, J., Goldblum, M., and Goldstein, T. Loss landscapes are all you need: Neural network generalization can be explained without the implicit bias of gradient descent. In The Eleventh International Conference on Learning Representations, 2023

    2023

  11. [11]

    Leveraging procedural generation to benchmark reinforcement learning

    Cobbe, K., Hesse, C., Hilton, J., and Schulman, J. Leveraging procedural generation to benchmark reinforcement learning. arXiv preprint arXiv:1912.01588, 2019

    Pith/arXiv arXiv 1912

  12. [12]

    Learning universal policies via text-guided video generation

    Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J., Schuurmans, D., and Abbeel, P. Learning universal policies via text-guided video generation. Advances in neural information processing systems, 36: 0 9156--9172, 2023

    2023

  13. [13]

    Dziugaite, G. K. and Roy, D. M. The size of teachers as a measure of data complexity: Pac-bayes excess risk bounds and scaling laws. In Proceedings of The 28th International Conference on Artificial Intelligence and Statistics, 2025

    2025

  14. [14]

    Imitating latent policies from observation

    Edwards, A., Sahni, H., Schroecker, Y., and Isbell, C. Imitating latent policies from observation. In Proceedings of the 36th International Conference on Machine Learning, pp.\ 1755--1763, 2019

    2019

  15. [15]

    Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures

    Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning, I., et al. Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. In International conference on machine learning, pp.\ 1407--1416. PMLR, 2018

    2018

  16. [16]

    Implicit regularization of discrete gradient dynamics in linear neural networks

    Gidel, G., Bach, F., and Lacoste-Julien, S. Implicit regularization of discrete gradient dynamics in linear neural networks. 2019

    2019

  17. [17]

    Prediction with action: Visual policy learning via joint denoising process

    Guo, Y., Hu, Y., Zhang, J., Wang, Y.-J., Chen, X., Lu, C., and Chen, J. Prediction with action: Visual policy learning via joint denoising process. Advances in Neural Information Processing Systems, 37: 0 112386--112410, 2024

    2024

  18. [18]

    Video prediction policy: A generalist robot policy with predictive visual representations

    Hu, Y., Guo, Y., Wang, P., Chen, X., Wang, Y.-J., Zhang, J., Sreenath, K., Lu, C., and Chen, J. Video prediction policy: A generalist robot policy with predictive visual representations. arXiv preprint arXiv:2412.14803, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Dreamgen: Unlocking generalization in robot learning through neural trajectories

    Jang, J., Ye, S., Lin, Z., Xiang, J., Bjorck, J., Fang, Y., Hu, F., Huang, S., Kundalia, K., Lin, Y.-C., et al. Dreamgen: Unlocking generalization in robot learning through neural trajectories. arXiv e-prints, pp.\ arXiv--2505, 2025

    2025

  20. [20]

    Ko, P.-C., Mao, J., Du, Y., Sun, S.-H., and Tenenbaum, J. B. Learning to act from actionless videos through dense correspondences. arXiv preprint arXiv:2310.08576, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    and Panchenko, D

    Koltchinskii, V. and Panchenko, D. Rademacher processes and bounding the risk of function learning. In High Dimensional Probability II, 2000

    2000

  22. [22]

    Unified video action model, 2025

    Li, S., Gao, Y., Sadigh, D., and Song, S. Unified video action model, 2025. URL https://arxiv.org/abs/2503.00200

    Pith/arXiv arXiv 2025

  23. [23]

    Autoregressive image generation without vector quantization

    Li, T., Tian, Y., Li, H., Deng, M., and He, K. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 0 56424--56445, 2024

    2024

  24. [24]

    Dreamitate: Real-world visuomotor policy learning via video generation

    Liang, J., Liu, R., Ozguroglu, E., Sudhakar, S., Dave, A., Tokmakov, P., Song, S., and Vondrick, C. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Video generators are robot policies

    Liang, J., Tokmakov, P., Liu, R., Sudhakar, S., Shah, P., Ambrus, R., and Vondrick, C. Video generators are robot policies. arXiv preprint arXiv:2508.00795, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Libero: Benchmarking knowledge transfer for lifelong robot learning

    Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., and Stone, P. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36: 0 44776--44791, 2023

    2023

  27. [27]

    Videoagenttrek: Computer use pretraining from unlabeled videos

    Lu, D., Xu, Y., Wang, J., Wu, H., Wang, X., Wang, Z., Yang, J., Su, H., Chen, J., Chen, J., et al. Videoagenttrek: Computer use pretraining from unlabeled videos. arXiv preprint arXiv:2510.19488, 2025

    arXiv 2025

  28. [28]

    and Tsybakov, A

    Mammen, E. and Tsybakov, A. B. Smooth discrimination analysis. The Annals of Statistics, 1999

    1999

  29. [29]

    Mingard, C., Valle-P \'e rez, G., Skalse, J., and Louis, A. A. Is sgd a bayesian sampler? well, almost. Journal of Machine Learning Research, 2021

    2021

  30. [30]

    MIT Press, 2018

    Mohri, M., Rostamizadeh, A., and Talwalkar, A. MIT Press, 2018

    2018

  31. [31]

    Combining self-supervised learning and imitation for vision-based rope manipulation

    Nair, A., Chen, D., Agrawal, P., Isola, P., Abbeel, P., Malik, J., and Levine, S. Combining self-supervised learning and imitation for vision-based rope manipulation. In 2017 IEEE international conference on robotics and automation (ICRA), pp.\ 2146--2153. IEEE, 2017

    2017

  32. [32]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0

    O’Neill, A., Rehman, A., Maddukuri, A., Gupta, A., Padalkar, A., Lee, A., Pooley, A., Gupta, A., Mandlekar, A., Jain, A., et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 6892--6903. IEEE, 2024

    2024

  33. [33]

    mimic-video: Video-action models for generalizable robot control beyond vlas

    Pai, J., Achenbach, L., Montesinos, V., Forrai, B., Mees, O., and Nava, E. mimic-video: Video-action models for generalizable robot control beyond vlas. arXiv preprint arXiv:2512.15692, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Pomerleau, D. A. ALVINN : An autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems, 1988

    1988

  35. [35]

    State-only imitation learning for dexterous manipulation

    Radosavovic, I., Wang, X., Pinto, L., and Malik, J. State-only imitation learning for dexterous manipulation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 7865--7871. IEEE, 2021

    2021

  36. [36]

    Lecture notes

    Rigollet, P. Lecture notes. Mathematics of Machine Learning (18.657), MIT, 2015. URL https://ocw.mit.edu/courses/18-657-mathematics-of-machine-learning-fall-2015/81406c87dccb9e873cfafa876a4d69c3_MIT18_657F15_LecNote.pdf. Accessed: 2026-01-23

    2015

  37. [37]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  38. [38]

    and Jiang, M

    Schmidt, D. and Jiang, M. Learning to act without actions. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=rvUq3cxpDF

    2024

  39. [39]

    Sen, A. K. and Srivastava, M. S. Regression analysis: Theory, methods, and applications, 1991

    1991

  40. [40]

    and Ben-David, S

    Shalev-Shwartz, S. and Ben-David, S. Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press, 2014

    2014

  41. [41]

    S., Gunasekar, S., and Srebro, N

    Soudry, D., Hoffer, E., Nacson, M. S., Gunasekar, S., and Srebro, N. The implicit bias of gradient descent on separable data. Journal of Machine Learning Research, 2018

    2018

  42. [42]

    Sutton, R. S. and Barto, A. G. Reinforcement Learning: An Introduction. The MIT Press, 2018

    2018

  43. [43]

    Predictive inverse dynamics models are scalable learners for robotic manipulation

    Tian, Y., Yang, S., Zeng, J., Wang, P., Lin, D., Dong, H., and Pang, J. Predictive inverse dynamics models are scalable learners for robotic manipulation. arXiv preprint arXiv:2412.15109, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Behavioral cloning from observation

    Torabi, F., Warnell, G., and Stone, P. Behavioral cloning from observation. arXiv preprint arXiv:1805.01954, 2018

    Pith/arXiv arXiv 2018

  45. [45]

    Recent advances in imitation learning from observation

    Torabi, F., Warnell, G., and Stone, P. Recent advances in imitation learning from observation. arXiv preprint arXiv:1905.13566, 2019

    Pith/arXiv arXiv 1905

  46. [46]

    Q., and Louis, A

    Valle-P \'e rez, G., Camargo, C. Q., and Louis, A. A. Deep learning generalizes because the parameter-function map is biased towards simple functions. International Conference on Learning Representations, 2019

    2019

  47. [47]

    Neural discrete representation learning

    Van Den Oord, A., Vinyals, O., et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017

    2017

  48. [48]

    Vapnik, V. N. and Chervonenkis, A. Y. On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability & Its Applications, 1971

    1971

  49. [49]

    Unleashing large-scale video generative pre-training for visual robot manipulation

    Wu, H., Jing, Y., Cheang, C., Chen, G., Xu, J., Li, X., Liu, M., Li, H., and Kong, T. Unleashing large-scale video generative pre-training for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    Y., et al

    Ye, S., Jang, J., Jeon, B., Joo, S., Yang, J., Peng, B., Mandlekar, A., Tan, R., Chao, Y.-W., Lin, B. Y., et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024

    Pith/arXiv arXiv 2024

  51. [51]

    Become a proficient player with limited data through watching pure videos

    Ye, W., Zhang, Y., Abbeel, P., and Gao, Y. Become a proficient player with limited data through watching pure videos. In The Eleventh International Conference on Learning Representations, 2022

    2022

  52. [52]

    Learning to drive by watching youtube videos: Action-conditioned contrastive policy pretraining

    Zhang, Q., Peng, Z., and Zhou, B. Learning to drive by watching youtube videos: Action-conditioned contrastive policy pretraining. In European Conference on Computer Vision, pp.\ 111--128. Springer, 2022

    2022

  53. [53]

    Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets

    Zhu, C., Yu, R., Feng, S., Burchfiel, B., Shah, P., and Gupta, A. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets. arXiv preprint arXiv:2504.02792, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    mazelab: A customizable framework to create maze and gridworld environments

    Zuo, X. mazelab: A customizable framework to create maze and gridworld environments. https://github.com/zuoxingdong/mazelab, 2018

    2018