Pith. sign in

REVIEW 3 major objections 6 minor 30 references

VILP: Imitation Learning with Latent Video Planning

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Latent video diffusion lets a robot plan actions in real time

desk verdict Solid latent-space video planning system with real-time closed-loop control, but the 'first' and 'temporal consistency' claims need quantitative support. read the letter →

arxiv 2502.01784 v1 pith:WINV7XAN submitted 2025-02-03 cs.RO cs.CV

Zhengtong Xu , Qiang Qiu , Yu She This is my paper · ORCID
classification cs.ROcs.CV
keywords latentvideoplanningimitationlearningdiffusionmodelrecedinghorizoncontrolmulti-viewgenerationrobotmanipulationreal-timemulti-modalactiondistribution
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

VILP (imitation learning with latent video planning) sets out to show that video generation can be a real-time planner for robots, not just an offline preview. Its latent video diffusion model predicts future frames in a compressed latent space, and those frames are converted directly into action sequences by a small CNN-plus-MLP policy. Running diffusion in latent space makes one inference take about 0.073 seconds on the tested hardware, so the policy can replan at roughly 14 Hz (and 5 Hz for two-view, six-frame 96x160 videos). The paper reports that this approach beats the pixel-space video planner UniPi on training cost, inference speed, temporal consistency, and task success, and that it matches or exceeds diffusion policy while using fewer action-labelled demonstrations. If these results transfer, imitation learning could draw on plentiful video data and still represent multi-modal actions and time-aligned multi-view observations.

What carries the argument

The load-bearing mechanism is the latent video diffusion model: a 3D-UNet with temporal and spatial convolutions that denoises latent future-frame stacks, combined with a VQGAN autoencoder that compresses images (for example, from 96x160 pixels to a 12x20 latent grid). Cross-attention conditioning embeds observation vectors into intermediate layers of the U-Net, which the paper argues gives better control than concatenating condition images. This machinery carries the speed claim by shrinking the denoising manifold and by working with DDIM so that only 4 to 16 denoising steps are enough; it carries the multi-view claim by training one diffusion model per view with fused multi-view conditioning, aligning generated frames across views.

What would settle it

Take a task where the planner's frames are fed to the low-level policy, and compare rollouts against the same policy fed ground-truth future frames: a large success gap would show that the video-to-action bridge, not the action module, is the bottleneck. Concretely, on Push-T or Nut-Assembly, one could swap generated frames for real future frames at test time and measure the drop in success rate; if success stays high with real frames but VILP fails with its own generated frames, the paper's central planning claim is falsified.

Watch

Extended reading notes

Core claim

At the center of VILP is a latent video diffusion planner. Given current observations, per-view visual encoders produce a condition vector; a 3D-UNet denoises a stack of future latent frames conditioned on that vector through cross-attention; and a fixed VQGAN decoder maps the denoised latents back into predicted frames. Two adjacent predicted frames are then fed into a small CNN-plus-MLP action module that outputs an action chunk, with receding-horizon control executing only the first few actions before replanning. The paper's central claim is that this design makes video generation fast enough for real-time robot policies, that generated videos from multiple views stay temporally aligned, and that policies built this way need fewer action-labelled demonstrations than direct imitation learning. In experiments across simulation and a real-world block-arranging task, VILP outperforms UniPi on generation quality, speed, and training memory and is competitive with diffusion policy while using less action data.

Load-bearing premise

The whole pipeline rests on the assumption that videos generated by the latent diffusion model are accurate enough simulations of the robot's future that a low-level policy trained on real frames can map them to correct actions, despite the artifacts the paper acknowledges in generated videos.

Editorial extensions

If this is right

  • Real-time closed-loop video planning becomes practical: VILP reports roughly 14 Hz inference for single-view five-frame videos and 5 Hz for two-view six-frame 96x160 videos, so policies can replan every few control steps.
  • Video data can substitute for action labels: on the 'Hybrid' datasets, VILP trained with a small action set supplemented by off-target demonstrations outperforms diffusion policy trained on the same action data.
  • Multi-view video generation can be time-aligned, which matters for policies that fuse observations from several cameras.
  • Multi-modal action distributions, such as the variety of ways to arrange blocks in the real-world task, are represented by the generative video planner rather than by a hand-designed action distribution.
  • Training memory for video planning drops: VILP trains in 6 to 10 GB versus 20 to 82 GB for UniPi at batch size 16, lowering the hardware barrier.

Reading between the lines

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

  • Beyond the paper, if the latent planner is a trustworthy world model, it could be reused as the dynamics model in model-predictive control or model-based RL, scoring candidate actions by predicted goal attainment.
  • Beyond the paper, the 'Hybrid' results suggest a data-economics principle: off-task or low-quality action data combined with task-relevant video can produce a strong policy, which could guide dataset collection for robotic manipulation.
  • Beyond the paper, a universal video-to-action mapper would let a web-scale video generator assume planning, with the small CNN-plus-MLP module acting as a readout; VILP's decoupling of video and action training makes this a testable next step.
  • A natural stress test the paper does not run is measuring success when generated frames are replaced by real future frames, isolating how much of the policy's error comes from video generation versus action mapping.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes VILP, an imitation-learning method in which a latent video diffusion model generates future multi-view video clips conditioned on current observations, and a low-level CNN-MLP policy maps pairs of adjacent generated frames into action sequences under receding-horizon control. The video planner is trained on video-only data, while the action mapping is trained on action-labeled data; the authors claim this reduces reliance on action-labeled data, enables real-time planning (e.g., about 14 Hz in one configuration), produces multi-view videos with temporal consistency, and outperforms UniPi in training cost, inference speed, temporal consistency, and policy success rate. Experiments include three video-prediction datasets, four simulation manipulation tasks, and one real-robot task, with ablations over denoising steps, video horizon, action horizon, and conditioning design.

Significance. If the claims hold, the paper is a useful practical demonstration that a latent video diffusion model can serve as a real-time receding-horizon planner and that video pretraining can reduce the amount of action-labeled data needed for a downstream action mapping. The experimental scope is broader than many system papers: three-seed rollouts over 50 environments, an ablation of denoising steps and horizons, a real-world Franka task, and public code. The main gap is that the two load-bearing claims, temporal consistency and robust video-to-action transfer, are not directly measured; FID/FVD and success rates on the tested tasks support but do not fully certify those claims. The paper also makes a strong first-real-time-video-policy claim without benchmarking against close methods.

major comments (3)
  1. [Section V and Section VI-B] The load-bearing video-to-action transfer is asserted but not directly measured. In Section V, actions are computed as \hat{a}_{t+n\Delta t} = \pi(\hat{o}_{t+n\Delta t}, \hat{o}_{t+(n+1)\Delta t}), where \pi is trained on real demonstration frames, but at deployment the inputs are VQGAN-decoded outputs of the latent diffusion model. Section VI-B.3 acknowledges artifacts in generated videos, and Tables III-VII show that the full pipeline succeeds on the tested tasks, but there is no action-level experiment that quantifies this distribution shift. Please add a comparison of \pi applied to ground-truth future frames versus generated frames on the same held-out episodes (e.g., action prediction error or success-rate ceiling), or otherwise measure the sensitivity of the low-level policy to generated-video artifacts.
  2. [Section VI-A, Table I] The temporal-consistency claim is supported only by FVD and qualitative frames (Fig. 5). FVD is a distribution-level metric and does not certify within-video temporal coherence or cross-view alignment; a video generator with correct marginals but independently sampled frames can achieve low FVD. Since Section V feeds adjacent generated frame pairs to the low-level policy, temporal incoherence is precisely the failure mode that would break the pipeline. Please add a quantitative temporal-consistency metric, such as forecast-frame accuracy over a fixed horizon, optical-flow warping error, frame-to-frame LPIPS, or cross-view temporal-alignment error, and report it in or around Table I.
  3. [Section I and Section VI-B.2] The claim that VILP is 'the first work capable of real-time video generation for robotic policies' is not supported by a comparison with other real-time video-planning methods. The paper compares only with UniPi, a pixel-space method not designed for real-time control, and reports total inference time in Tables IV and VII without a component-wise latency breakdown (observation encoding, denoising, VQGAN decoding, action inference) and without comparing to other latent-video policies with real-time goals. Please either soften the claim to be relative to the evaluated baselines or benchmark against the closest real-time video-planning methods and report end-to-end control frequency on the same hardware.
minor comments (6)
  1. [Section VI] There is a typo in the experimental setup paragraph: 'VLIP' should be 'VILP'.
  2. [Table II] Table II compares GPU memory at different frame counts for some tasks (e.g., Move-the-Stack uses 8 frames for VILP and 5/6 for UniPi); please state the exact frame counts inside the table or in the caption so that the training-cost comparison is transparent.
  3. [Tables III and IV] The 'max success rate' metric is the best value over training, which is optimistic; please justify its use or move it to supplementary material, and make the caption self-contained by defining 'Score' explicitly in Table IV.
  4. [Table V] The baselines in Table V are 'directly adapted from [9]'; please state explicitly that the training data, observation modalities, and evaluation protocol for VILP are identical to those used for the baselines, or note the differences.
  5. [Section V] The notation \hat{o}_{t+n\Delta t} is overloaded because o is originally defined as a combination of multiview images; please clarify that \hat{o} denotes the VQGAN-decoded generated frame corresponding to the relevant view.
  6. [Section VI-B.1] The text says VILP results are 'comparable' to diffusion policy on Can-PickPlace; given the reported values (95.7/92.2 vs. 97/98), please make the comparison statement more precise about the seed-level spread and the difference between input modalities.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VILP's predictions are evaluated on held-out episodes and physical rollouts, and no equation reduces the claimed results to fitted inputs.

full rationale

The paper's derivation chain is self-contained. The video planner is trained with a standard latent diffusion objective (Eq. 2) conditioned on observations, and the low-level policy maps generated frames to actions via π(ô_{t+nΔt}, ô_{t+(n+1)Δt}). The reported success rates come from environment rollouts, not from the training loss, so no 'prediction' is equivalent to a fitted input by construction. The video-quality metrics (FID/FVD) are computed on a held-out 10% of episodes, and policy rollouts use environment feedback. The comparison against UniPi uses the same low-level policy and controlled U-Net sizes, so the comparison is not circular. The self-citations ([26], [27]) provide datasets or examples of multimodal observations but are not load-bearing for the central claims; no uniqueness theorem or ansatz is smuggled in via self-citation. The acknowledged artifacts in generated videos (Section VI-B.3) are a robustness concern, not a circularity, because the policy is still evaluated against external task success.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rest on the assumptions listed above, plus standard tools (DDIM, VQGAN, cross-attention) taken from prior literature. The only 'free' choices are hyperparameters (denoising steps, horizons) tuned per task; no scientific constants are fitted.

free parameters (3)
  • Denoising steps K = 4, 8, or 16 depending on experiment
    Number of DDIM denoising steps used at inference; ablated in Table IV. Lower K enables real-time speeds but may reduce video quality. It is a hyperparameter chosen per experiment, not a constant derived from the method.
  • Video planning horizon N = 5, 6, 8, or 12 depending on task
    Length of the predicted video (number of future frames). Ablated in Fig. 7; optimal around 6 or 12. Chosen by hand per task to balance cost and performance.
  • Action horizon N_e = 4 to 16 depending on task
    Number of executed actions per receding horizon cycle. Ablated in Fig. 7; larger values improve performance but are limited by the video horizon. Tuned per task.
assumptions (3)
  • domain assumption The latent video diffusion model, trained on demonstration videos, generates future frames that faithfully simulate the robot's future dynamics and object interactions.
    Invoked in Sections IV-C and V. The generated video is the basis for action generation. If the predicted frames are physically incorrect, the policy fails. The paper provides only qualitative support (Fig. 5).
  • domain assumption The low-level policy, trained on real observation frames, transfers to generated (slightly artifact-laden) frames.
    Section V trains the action mapper on real frames; inference uses generated frames. The paper's own admission that generated videos contain minor artifacts (Section VI-B.3) makes this an explicit load-bearing assumption, tested empirically but not guaranteed.
  • domain assumption Separate per-view video diffusion models produce time-aligned videos because they share a common conditioning embedding.
    Section IV-D asserts cross-view temporal alignment; no quantitative metric is reported, so this is an architectural assumption rather than a measured result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VILP: Imitation Learning with Latent Video Planning." pith.science (2026). https://pith.science/paper/WINV7XAN

@misc{pith2026250201784,
  author       = {Pith},
  title        = {Pith review of: VILP: Imitation Learning with Latent Video Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WINV7XAN}},
  note         = {Machine review of arXiv:2502.01784}
}
read the original abstract

In the era of generative AI, integrating video generation models into robotics opens new possibilities for the general-purpose robot agent. This paper introduces imitation learning with latent video planning (VILP). We propose a latent video diffusion model to generate predictive robot videos that adhere to temporal consistency to a good degree. Our method is able to generate highly time-aligned videos from multiple views, which is crucial for robot policy learning. Our video generation model is highly time-efficient. For example, it can generate videos from two distinct perspectives, each consisting of six frames with a resolution of 96x160 pixels, at a rate of 5 Hz. In the experiments, we demonstrate that VILP outperforms the existing video generation robot policy across several metrics: training costs, inference speed, temporal consistency of generated videos, and the performance of the policy. We also compared our method with other imitation learning methods. Our findings indicate that VILP can rely less on extensive high-quality task-specific robot action data while still maintaining robust performance. In addition, VILP possesses robust capabilities in representing multi-modal action distributions. Our paper provides a practical example of how to effectively integrate video generation models into robot policies, potentially offering insights for related fields and directions. For more details, please refer to our open-source repository https://github.com/ZhengtongXu/VILP.

Figures

Figures reproduced from arXiv: 2502.01784 by the authors.

Figure 1
Figure 1. VILP is capable of generating predictive robot videos that adhere [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our proposed video planning pipeline. For the architec [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the low-level policy that maps the predicted video to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Example frames extracted from the generated videos. For more [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Illustration of “Hybrid” datasets. B. Imitation Learning with Video Planning 1) Simulation Tasks: In this section, we present the results of implementing VILP as policies in four simulation tasks: Nut-Assembly [29], Can-PickPlace[29], Sim Push-T [30], and Arrange-Block…
Figure 7
Figure 7. Figure 7: Horizon Ablation Study: The results are based on Sim Push-T task. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Snapshots of the VILP rollout in the Real-Arrange-Blocks task. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 11 canonical work pages

  1. [1]

    Video generation models as world simulators,

    T. Brooks, B. Peebles, C. Homes, W. DePue, Y . Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman et al., “Video generation models as world simulators,” 2024

    work page 2024

  2. [2]

    Video language planning,

    Y . Du, M. Yang, P. Florence, F. Xia, A. Wahid, B. Ichter, P. Sermanet, T. Yu, P. Abbeel, J. B. Tenenbaum et al. , “Video language planning,” arXiv preprint arXiv:2310.10625 , 2023

    arXiv 2023

  3. [3]

    Learning universal policies via text-guided video generation,

    Y . Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schuur- mans, and P. Abbeel, “Learning universal policies via text-guided video generation,” Advances in Neural Information Processing Systems , vol. 36, 2024

    work page 2024

  4. [4]

    Learn- ing to Act from Actionless Video through Dense Correspondences,

    P.-C. Ko, J. Mao, Y . Du, S.-H. Sun, and J. B. Tenenbaum, “Learn- ing to Act from Actionless Video through Dense Correspondences,” arXiv:2310.08576, 2023

    arXiv 2023

  5. [5]

    Long video generation with time-agnostic vqgan and time- sensitive transformer,

    S. Ge, T. Hayes, H. Yang, X. Yin, G. Pang, D. Jacobs, J.-B. Huang, and D. Parikh, “Long video generation with time-agnostic vqgan and time- sensitive transformer,” in European Conference on Computer Vision . Springer, 2022, pp. 102–118

    work page 2022

  6. [6]

    Magvit: Masked gen- erative video transformer,

    L. Yu, Y . Cheng, K. Sohn, J. Lezama, H. Zhang, H. Chang, A. G. Hauptmann, M.-H. Yang, Y . Hao, I. Essaet al., “Magvit: Masked gen- erative video transformer,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 2023, pp. 10 459– 10 469

    work page 2023

  7. [7]

    Videogpt: Video generation using vq-vae and transformers,

    W. Yan, Y . Zhang, P. Abbeel, and A. Srinivas, “Videogpt: Video generation using vq-vae and transformers,” arXiv preprint arXiv:2104.10157, 2021

    arXiv 2021

  8. [8]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” in Conference on Robot Learning . PMLR, 2022, pp. 158– 168

    work page 2022

Show all 30 references
  1. [9]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” in Proceedings of Robotics: Science and Systems (RSS) , 2023

    2023

  2. [10]

    Behavior generation with latent actions,

    S. Lee, Y . Wang, H. Etukuru, H. J. Kim, N. M. M. Shafiullah, and L. Pinto, “Behavior generation with latent actions,” arXiv preprint arXiv:2403.03181, 2024

    2024 arXiv

  3. [11]

    Deep generative models in robotics: A survey on learning from multimodal demonstrations,

    J. Urain, A. Mandlekar, Y . Du, M. Shafiullah, D. Xu, K. Fragkiadaki, G. Chalvatzaki, and J. Peters, “Deep generative models in robotics: A survey on learning from multimodal demonstrations,” arXiv preprint arXiv:2408.04380, 2024

    2024 arXiv

  4. [12]

    Learning interactive real-world simulators,

    M. Yang, Y . Du, K. Ghasemipour, J. Tompson, D. Schuurmans, and P. Abbeel, “Learning interactive real-world simulators,” arXiv preprint arXiv:2310.06114, 2023

    2023 arXiv

  5. [13]

    Seer: Language in- structed video prediction with latent diffusion models,

    X. Gu, C. Wen, W. Ye, J. Song, and Y . Gao, “Seer: Language in- structed video prediction with latent diffusion models,” arXiv preprint arXiv:2303.14897, 2023

    2023 arXiv

  6. [14]

    Dreamitate: Real-world visuomotor policy learning via video generation,

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

    2024 arXiv

  7. [15]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning . PMLR, 2015, pp. 2256–2265

    2015

  8. [16]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

    2020

  9. [17]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

    2010 arXiv

  10. [18]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  11. [19]

    Taming transformers for high-resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 12 873–12 883

    2021

  12. [20]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  13. [21]

    Vector-quantized image modeling with improved vqgan,

    J. Yu, X. Li, J. Y . Koh, H. Zhang, R. Pang, J. Qin, A. Ku, Y . Xu, J. Baldridge, and Y . Wu, “Vector-quantized image modeling with improved vqgan,” arXiv preprint arXiv:2110.04627 , 2021

    2021 arXiv

  14. [22]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

    2021

  15. [23]

    U-net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional net- works for biomedical image segmentation,” in Medical image comput- ing and computer-assisted intervention–MICCAI 2015: 18th interna- tional conference, Munich, Germany, October 5-9, 2015, proceedings, part III...

    2015

  16. [24]

    Perceiver: General perception with iterative attention,

    A. Jaegle, F. Gimeno, A. Brock, O. Vinyals, A. Zisserman, and J. Carreira, “Perceiver: General perception with iterative attention,” in International conference on machine learning . PMLR, 2021, pp. 4651–4664

    2021

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023

    2023

  18. [26]

    UniT: Unified tactile representation for robot learning,

    Z. Xu, R. Uppuluri, X. Zhang, C. Fitch, P. G. Crandall, W. Shou, D. Wang, and Y . She, “UniT: Unified tactile representation for robot learning,” 2024. [Online]. Available: https://arxiv.org/abs/2408.06481

    2024 arXiv

  19. [27]

    Leto: Learning constrained visuomotor pol- icy with differentiable trajectory optimization,

    Z. Xu and Y . She, “Leto: Learning constrained visuomotor pol- icy with differentiable trajectory optimization,” arXiv preprint arXiv:2401.17500, 2024

    2024 arXiv

  20. [28]

    Transporter networks: Rearranging the visual world for robotic manipulation,

    A. Zeng, P. Florence, J. Tompson, S. Welker, J. Chien, M. Attarian, T. Armstrong, I. Krasin, D. Duong, V . Sindhwani et al. , “Transporter networks: Rearranging the visual world for robotic manipulation,” in Conference on Robot Learning . PMLR, 2021, pp. 726–747

    2021

  21. [29]

    What matters in learning from offline human demonstrations for robot manipulation,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart ´ın-Mart´ın, “What matters in learning from offline human demonstrations for robot manipulation,” in Proc. Conf. Robot Learn. , 2022, pp. 1678–1690

    2022

  22. [30]

    Iterative Residual Policy: for goal-conditioned dynamic manipulation of de- formable objects,

    C. Chi, B. Burchfiel, E. Cousineau, S. Feng, and S. Song, “Iterative Residual Policy: for goal-conditioned dynamic manipulation of de- formable objects,” in in Proc. of Robot.: Sci. and Syst , 2022

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.