Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

CDP: Towards Robust Autoregressive Visuomotor Policy Learning via Causal Diffusion

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Causal Diffusion Policy claims that conditioning action denoising on past robot actions improves manipulation accuracy and robustness, especially when sensor observations are degraded.

desk verdict A plausible causal-caching diffusion policy with an overly broad robustness claim; the experiments are broad but the key evidence for robustness is thin. read the letter →

arxiv 2506.14769 v2 pith:VWIMUOFH submitted 2025-06-17 cs.CV cs.RO

Jiahua Ma , Yiran Qin , Yixiong Li , Xuanqi Liao , Yulan Guo , Ruimao Zhang This is my paper
classification cs.CVcs.RO
keywords roboticmanipulationimitationlearningdiffusionpolicycausaltransformerautoregressiveactiongenerationcachesharingobservationdegradationvisuomotor
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 a robot manipulation policy can predict actions more accurately by treating action generation as a denoising diffusion process conditioned on a history of the robot's own past actions, rather than conditioning only on the current observation. The proposed Causal Diffusion Policy (CDP) uses a causal transformer so that each denoising target attends to prior actions, and it introduces a small amount of noise into those historical actions during training so the policy learns to use their coarse temporal dynamics instead of their exact values. The paper claims this yields higher success rates than standard Diffusion Policy and 3D Diffusion Policy on 2D and 3D manipulation tasks, and that CDP keeps working well when the current observation is degraded by sensor noise. A sympathetic reader would care because real robot deployments face noisy and partial observations, and a policy that can fall back on its own motion history would be more reliable in practice.

What carries the argument

The load-bearing mechanism is the Causal Action Generation Module, built from three parts: a Causal Temporal Attention mask, Historical Actions Re-Denoising, and cache sharing. The attention mask defines what each action can see: historical actions are chunked so only same-chunk history is visible, while every denoising target can attend to the full history, which is what makes prediction autoregressive and temporally coherent. Historical Actions Re-Denoising perturbs the past action sequence with small noise during training, forcing the policy to rely on coarse temporal dynamics rather than memorized exact values and preventing error accumulation at inference. Cache sharing stores the key-value pairs of historical actions once and reuses them across denoising steps, turning the quadratic cost of repeated attention into reused computation and giving latency improvements that grow with history length.

What would settle it

Run CDP on several manipulation tasks with complete camera dropout or persistent occlusion of the workspace rather than point-cloud noise; if success drops to the level of a policy that receives no observation at all, the claim that historical action sequences compensate for degraded observations is falsified.

Watch

Extended reading notes

Core claim

CDP's central claim is that conditioning action denoising on historical action sequences improves both accuracy and robustness in visuomotor imitation learning. During training, historical actions $\tilde A$ are perturbed with small noise $N_\sigma \sim \mathcal{N}(0,\sigma^2)$, concatenated with denoising targets $N$, and fed through a transformer; the objective is $\min \mathbb{E}_{\tilde A,A,N}\|D_\theta([\tilde A,N])-A\|_2^2$ (Eq. 1). A Causal Temporal Attention mask lets each denoising target attend to all historical actions while keeping historical chunks from attending across chunks, and the denoising timestep is injected in the visual-action cross-attention layer so cached features stay reusable. At inference, a cache-sharing mechanism reuses key-value pairs computed in earlier autoregressive steps. The paper argues that CDP thereby leverages temporal continuity to compensate for degraded observations, giving it a roughly 5 to 20 percentage point advantage over the diffusion baselines on the evaluated benchmarks and near-constant success under increasing point-cloud noise where the 3D baseline collapses.

Load-bearing premise

The method assumes that the robot's past actions are informative enough about the hidden state of the task to substitute for degraded or missing spatial information in the current observation, and the paper tests this assumption on only one simulated task with point-cloud noise.

Editorial extensions

If this is right

  • On the Adroit, DexArt, MetaWorld, and RoboFactory benchmarks, CDP reports success rates roughly 5 to 20 percentage points above the corresponding Diffusion Policy baselines with the same visual encoder.
  • On the RoboFactory Lift Barrier task, CDP maintains high success when low-resolution point clouds are corrupted by growing noise, whereas the 3D Diffusion Policy's success drops sharply.
  • The cache-sharing mechanism reduces per-step inference latency, with the speed-up scaling roughly linearly with the length of the historical action sequence.
  • In real-world tests with 50 demonstrations per task, CDP reports higher grasping, placing, and overall success on collecting objects, stacking cubes, and push-T than the Diffusion Policy baseline.

Reading between the lines

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

  • The paper's robustness evidence covers only point-cloud noise on one simulated task; if historical action conditioning is genuinely compensating for missing spatial cues, CDP should also stay accurate under occlusions, missing depth, or camera dropout, but that remains an untested extension.
  • Because the model stores cached key-value pairs from earlier steps, fast-changing scenes or sudden visual discontinuities could make cached observation features stale; testing CDP with abrupt scene changes would probe this boundary.
  • The historical-action perturbation acts like a task-agnostic regularizer; setting its noise scale per task (the paper uses 1/6 or 1/2) suggests a tunable hyperparameter whose optimal value could be predicted from action repeatability or trajectory smoothness.
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

4 major / 4 minor

Summary. The paper proposes Causal Diffusion Policy (CDP), a transformer-based diffusion policy that conditions action denoising on historical action sequences via a causal temporal attention mechanism, and introduces a chunk-wise autoregressive inference scheme with key-value cache sharing to reduce computation. The authors report simulation results on Adroit, DexArt, MetaWorld, and RoboFactory against Diffusion Policy (DP) and 3D Diffusion Policy (DP3), real-world experiments on three manipulation tasks, and an ablation on the RoboFactory Lift Barrier task that claims robustness to degraded point-cloud observations.

Significance. If the central claims are established, conditioning action denoising on action history is a simple and plausible improvement to diffusion policies, and the cache-sharing mechanism could make autoregressive inference more practical. The paper evaluates on several external benchmarks and includes real-robot experiments, which is a strength. However, the evidence as presented does not yet support the strong robustness claim in the abstract, and the quantitative comparisons lack the error bars promised by the stated evaluation protocol.

major comments (4)
  1. [Section 4.1 / Table 1 and Appendix A.1] Table 1 reports single success percentages for every method and task, even though Appendix A.1 states that Adroit, DexArt, and MetaWorld are evaluated over three seeds with the mean and standard deviation computed. Without error bars or seed-level numbers, the claimed 5-to-20-point improvements over baselines cannot be assessed, and at least one entry (CDP 22% vs. DP 29% on RoboFactory Place Food) goes against the paper's overall claim of consistent superiority.
  2. [Section 4.2 / Figure 4a and Appendix A.1/A.2] The robustness experiment is too weak to support the causal story that historical action sequences compensate for degraded observations. The noise level is not specified in units relative to the coordinate range of the 64×3 point clouds, no noise-free condition at the same training settings is shown for DP3 and CDP, and the test covers only one simple task. The maintained performance under added noise could therefore reflect a mild degradation that still leaves spatial structure recoverable, or a baseline performance gap, rather than temporal reasoning.
  3. [Section 3.1 / Equation (1)] Equation (1) is under-specified for a diffusion model: it defines the objective as minimizing the L2 distance between Dθ([A~, N]) and the ground-truth action A, where N is Gaussian noise, but it does not describe a forward noising process on the target actions, a noise schedule, or explicit timestep conditioning. The text later refers to 'denoising timestep t' and multiple denoising steps, so the relationship between Equation (1) and the actual training procedure must be clarified; as written, the equation describes a single-step conditional generator rather than a diffusion training loss.
  4. [Appendix A.3 / Tables 4-6] The key hyperparameters (noise scale, chunk size, and action lengths) are tuned on the RoboFactory Lift Barrier task, and the same benchmark is also used for the main quantitative comparison. The benchmark reference [46] shares authors with this submission, and Table 6 shows CDP3 success rates ranging from 0% to 40% depending on the action-window configuration, indicating high sensitivity. Without a held-out tuning protocol or evaluation on an independent benchmark, the main results may reflect favorable hyperparameter selection rather than a robust algorithmic advantage.
minor comments (4)
  1. [Appendix A.3 / Table 6] The column header '? - 4 - 4 ? - 4 - 8 8 - ? - 8' and the sub-row '4 8 16 4 8 16 2 4 8' are malformed; the three hyperparameter configurations need explicit labels so the reader can map columns to the historical-action length, valid target length, and redundant target length.
  2. [Sections 3.1 and 4.2] The text refers to 'Fig. 4 (a)' and 'Fig. 4 (b)' in Section 3.1 for the module architecture and attention mask, and again in Section 4.2 for the robustness and efficiency plots; these references should be disambiguated because they appear to denote different figures.
  3. [Abstract and Section 4.2] The abstract claims robustness against sensor noise, occlusions, and hardware limitations, but the experiments only add point-cloud noise; no occlusion, missing depth, or camera-dropout test is provided, so the wording overstates the evidence.
  4. [Section 4.1 and Appendix A.2] The phrase 'with an identical visual encoder' is not fully supported by the nonstandard RoboFactory settings in Appendix A.2 (image size 64×64 vs. 84×84, point cloud size 128 vs. 512, noise scale 1/2 vs. 1/6, and epoch 300 vs. 3000); the authors should clarify whether the baselines used these same modified settings in the RoboFactory experiments.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity: the training and inference rules are self-contained; the score reflects a minor self-authored-benchmark overlap in the robustness evidence, not a derivation that reduces to its inputs.

full rationale

The paper's derivation chain is self-contained. Equation (1) defines a conditional denoising objective that maps a concatenation of perturbed historical actions and Gaussian target noise to the ground-truth target action; Equation (2) is a training-time perturbation of the conditioning history; and Equation (6) is the corresponding autoregressive sampling rule. No term in these equations equals a reported success rate by construction, and the cache-sharing speedup follows exactly from reusing timestep-independent key-value pairs, not from fitting a parameter to the reported result. The central robustness claim in Sec. 4.2 is an empirical measurement, not a quantity derived from the method's equations; its main weaknesses are that the only degradation experiment is on RoboFactory Lift Barrier, a benchmark authored by co-authors of this paper, and the added noise level is under-specified, so the breadth of the abstract's robustness claim is not independently established. These are evidence and selection concerns rather than circular derivations. The Conclusion's Limitation paragraph also honestly notes that extremely long horizons were not addressed, further limiting generalization but not indicating circularity. The self-citations to prior group work (e.g., RoboFactory [46]) are present but are not load-bearing for the training and inference equations, so the score is a 2 under the minor-self-citation rubric rather than a 0.

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

The central contribution is empirical and architectural. It relies on four hand-selected hyperparameters and at least three domain assumptions. No new physical entities or forces are introduced; the only new objects are software mechanisms (causal mask, cyclic embedding, cache) that are not scored as invented entities.

free parameters (4)
  • Perturbed historical action noise scale σ = 1/6 × N(0,1) in simulation; 1/2 × N(0,1) in RoboFactory and real-world
    Chosen by ablation on Lift Barrier (Table 4) to balance tolerance to action errors against distortion of history; appears directly in Eq. 2.
  • Chunk size for causal attention = 8 for simpler tasks; larger for complex tasks
    Selected by ablation on Lift Barrier (Table 5); defines the local temporal window in the Causal Temporal Attention mask (Sec. 3.1).
  • Length of Historical Actions L = 20 in simulation; 8 in real-world
    Selected via ablation (Table 6) on Lift Barrier; determines how much past action context conditions each target block (Eq. 1).
  • Target action lengths (valid, redundant) = 8 + 4 in simulation; 16 + 8 in real-world
    Follows Diffusion Policy and is tuned with history length in Table 6; the redundant part is predicted and then discarded, as in DP.
assumptions (4)
  • standard math The L2 denoising objective (Eq. 1) with standard diffusion training yields a valid conditional action distribution.
    Background assumption taken from Diffusion Policy and used throughout Sec. 3 without re-derivation.
  • domain assumption Historical actions and current observation are jointly sufficient to predict the next action block.
    This is the premise of the causal conditioning design (Sec. 3.1) and the robustness story (Sec. 4.2); it is not proven and is tested only partially.
  • ad hoc to paper Adding small noise to historical actions during training prevents autoregressive error accumulation at inference.
    Introduced as 'Historical Actions Re-Denoising' (Eq. 2); its effectiveness is asserted via ablation on one task, not derived.
  • domain assumption The chunked causal attention mask used in training matches the cache-sharing inference procedure exactly.
    Required for the KV-cache speedup in Sec. 3.2 to be lossless; no proof or ablation isolating prediction change is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CDP: Towards Robust Autoregressive Visuomotor Policy Learning via Causal Diffusion." pith.science (2026). https://pith.science/paper/VWIMUOFH

@misc{pith2026250614769,
  author       = {Pith},
  title        = {Pith review of: CDP: Towards Robust Autoregressive Visuomotor Policy Learning via Causal Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWIMUOFH}},
  note         = {Machine review of arXiv:2506.14769}
}
read the original abstract

Diffusion Policy (DP) enables robots to learn complex behaviors by imitating expert demonstrations through action diffusion. However, in practical applications, hardware limitations often degrade data quality, while real-time constraints restrict model inference to instantaneous state and scene observations. These limitations seriously reduce the efficacy of learning from expert demonstrations, resulting in failures in object localization, grasp planning, and long-horizon task execution. To address these challenges, we propose Causal Diffusion Policy (CDP), a novel transformer-based diffusion model that enhances action prediction by conditioning on historical action sequences, thereby enabling more coherent and context-aware visuomotor policy learning. To further mitigate the computational cost associated with autoregressive inference, a caching mechanism is also introduced to store attention key-value pairs from previous timesteps, substantially reducing redundant computations during execution. Extensive experiments in both simulated and real-world environments, spanning diverse 2D and 3D manipulation tasks, demonstrate that CDP uniquely leverages historical action sequences to achieve significantly higher accuracy than existing methods. Moreover, even when faced with degraded input observation quality, CDP maintains remarkable precision by reasoning through temporal continuity, which highlights its practical robustness for robotic control under realistic, imperfect conditions.

Figures

Figures reproduced from arXiv: 2506.14769 by the authors.

Figure 1
Figure 1. Causal Diffusion Policy: a transformer-based diffusion model that enhances action pre￾diction by conditioning on historical action sequences. A: When performing the task of “grabbing the barrier” in practice, B: the quality of observations is degraded by factors such as sensor noise, occlusions, and hardware limitations. In fact, this degraded but high-dimensional observation data not only fails to provide sufficien… view at source ↗
Figure 2
Figure 2. Causal Action Generation of Our CDP. (a) During training, the Historical Actions A˜ are combined with the Denoising Targets N. This combined input is then fed into the Causal Action Generation module, which contains P blocks, for denoising. The Target Actions A are used for training supervision. Before denoising, A˜ is perturbed by a small-scale noise, which helps to reduce the accumulation of action prediction erro… view at source ↗
Figure 3
Figure 3. Chunk-wise Autoregressive Inference of Our CDP. (a) The orange and purple blocks denote actions whose Key and Value representations have been cached and not cached until the current step, respectively. The yellow block denotes the Gaussian noises. During the AR-k step, we perform denoising while simultaneously computing and storing the Key and Value representations for the Uncached Historical Actions. After denoisin… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of the Effectiveness and Efficiency of Our CDP. (a) The performance of DP3 collapses as observation degrades, whereas our CDP remains nearly unaffected. (b) Our Cache Sharing Mechanism yields inference-time speed-ups that scale approximately linearly with…
Figure 5
Figure 5. Figure 5: visualize the results of our CDP across various tasks on the Adroit, Dexart, MetaWorld, and RoboFactory benchmarks, which demonstrate the effectiveness of our model in multiple simulation environments [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Real-world Experimental Workspace B.2 Experiment Setup Demonstrations. The demonstrations utilized in our real-world experiments were meticulously collected via human teleoperation, facilitated by advanced vision-based retargeting techniques. For each task, we collecte…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. TFP: Temporally Conditioned Memory-Fusion Policies for Visuomotor Learning

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Injecting an LTC task-progress belief into a flow-matching VLA action head improves stage-dependent and occluded manipulation over reactive and step-indexed memory baselines.

Reference graph

Works this paper leans on

49 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [46]

    Y . Qin, L. Kang, X. Song, Z. Yin, X. Liu, X. Liu, R. Zhang, and L. Bai. Robofac- tory: Exploring embodied agent collaboration with compositional constraints. arXiv preprint arXiv:2503.16408, 2025

    arXiv 2025

  2. [1]

    Shridhar, L

    M. Shridhar, L. Manuelli, and D. Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023

    2023

  3. [2]

    C. Wang, L. Fan, J. Sun, R. Zhang, L. Fei-Fei, D. Xu, Y . Zhu, and A. Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422 , 2023

    arXiv 2023

  4. [3]

    Y . Ze, G. Yan, Y .-H. Wu, A. Macaluso, Y . Ge, J. Ye, N. Hansen, L. E. Li, and X. Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. InConference on Robot Learning, pages 284–301. PMLR, 2023

    work page 2023

  5. [4]

    X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020

    arXiv 2004

  6. [5]

    Agarwal, S

    A. Agarwal, S. Uppal, K. Shaw, and D. Pathak. Dexterous functional grasping. arXiv preprint arXiv:2312.02975, 2023

    arXiv 2023

  7. [6]

    Haldar, J

    S. Haldar, J. Pari, A. Rai, and L. Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations. arXiv preprint arXiv:2303.01497, 2023

    arXiv 2023

  8. [7]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023

    arXiv 2023

Show all 49 references
  1. [8]

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

    2020

  2. [9]

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

    2010 arXiv

  3. [10]

    Janner, Y

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991, 2022

    2022 arXiv

  4. [11]

    Y . Luo, C. Sun, J. B. Tenenbaum, and Y . Du. Potential based diffusion motion planning.arXiv preprint arXiv:2407.06169, 2024

    2024 arXiv

  5. [12]

    Carvalho, A

    J. Carvalho, A. T. Le, M. Baierl, D. Koert, and J. Peters. Motion planning diffusion: Learn- ing and planning of robot motions with diffusion models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 1916–1923. IEEE, 2023

    2023

  6. [13]

    K. Saha, V . Mandadi, J. Reddy, A. Srikanth, A. Agarwal, B. Sen, A. Singh, and M. Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 10351–10358. IEEE, 2024

    2024

  7. [14]

    E. Zhou, Y . Qin, Z. Yin, Y . Huang, R. Zhang, L. Sheng, Y . Qiao, and J. Shao. Minedreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024

    2024 arXiv

  8. [15]

    Y . Qin, A. Sun, Y . Hong, B. Wang, and R. Zhang. Navigatediff: Visual predictors are zero-shot navigation assistants. arXiv preprint arXiv:2502.13894, 2025

    2025 arXiv

  9. [16]

    Huang, Z

    S. Huang, Z. Wang, P. Li, B. Jia, T. Liu, Y . Zhu, W. Liang, and S.-C. Zhu. Diffusion-based gen- eration, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750–16761, 2023

    2023

  10. [17]

    Pearce, T

    T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V . Macua, S. Z. Tan, I. Momennejad, K. Hofmann, et al. Imitating human behaviour with diffusion models. arXiv preprint arXiv:2301.10677, 2023. 9

    2023 arXiv

  11. [18]

    H. Ha, P. Florence, and S. Song. Scaling up and distilling down: Language-guided robot skill acquisition. In Conference on Robot Learning, pages 3766–3777. PMLR, 2023

    2023

  12. [19]

    Z. Xian, N. Gkanatsios, T. Gervet, T.-W. Ke, and K. Fragkiadaki. Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation. In7th Annual Conference on Robot Learning, 2023

    2023

  13. [20]

    X. Li, V . Belagali, J. Shang, and M. S. Ryoo. Crossway diffusion: Improving diffusion-based visuomotor policy via self-supervised learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 16841–16849. IEEE, 2024

    2024

  14. [21]

    L. Wang, J. Zhao, Y . Du, E. H. Adelson, and R. Tedrake. Poco: Policy composition from and for heterogeneous robot learning. arXiv preprint arXiv:2402.02511, 2024

    2024 arXiv

  15. [22]

    K. Chen, E. Lim, K. Lin, Y . Chen, and H. Soh. Don’t start from scratch: Behavioral refinement via interpolant-based policy diffusion. arXiv preprint arXiv:2402.16075, 2024

    2024 arXiv

  16. [23]

    Sridhar, S

    K. Sridhar, S. Dutta, D. Jayaraman, J. Weimer, and I. Lee. Memory-consistent neural networks for imitation learning. arXiv preprint arXiv:2310.06171, 2023

    2023 arXiv

  17. [24]

    T. Z. Zhao, J. Tompson, D. Driess, P. Florence, K. Ghasemipour, C. Finn, and A. Wahid. Aloha unleashed: A simple recipe for robot dexterity. arXiv preprint arXiv:2410.13126, 2024

    2024 arXiv

  18. [25]

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024

    2024 arXiv

  19. [26]

    L. Kang, X. Song, H. Zhou, Y . Qin, J. Yang, X. Liu, P. Torr, L. Bai, and Z. Yin. Viki-r: Coordinating embodied multi-agent cooperation via reinforcement learning. arXiv preprint arXiv:2506.09049, 2025

    2025

  20. [27]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024

    2024

  21. [28]

    L. Fu, H. Huang, G. Datta, L. Y . Chen, W. C.-H. Panitch, F. Liu, H. Li, and K. Goldberg. In- context imitation learning via next-token prediction. arXiv preprint arXiv:2408.15980, 2024

    2024 arXiv

  22. [29]

    Zhang, Y

    X. Zhang, Y . Liu, H. Chang, L. Schramm, and A. Boularias. Autoregressive action sequence learning for robotic manipulation. IEEE Robotics and Automation Letters , 2025

    2025

  23. [30]

    Z. Gong, P. Ding, S. Lyu, S. Huang, M. Sun, W. Zhao, Z. Fan, and D. Wang. Carp: Visuomotor policy learning via coarse-to-fine autoregressive prediction. arXiv preprint arXiv:2412.06782, 2024

    2024 arXiv

  24. [31]

    J. Chen, J. Yu, C. Ge, L. Yao, E. Xie, Y . Wu, Z. Wang, J. Kwok, P. Luo, H. Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    2023 arXiv

  25. [32]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4195–4205, 2023

    2023

  26. [33]

    Y . Qin, Z. Shi, J. Yu, X. Wang, E. Zhou, L. Li, Z. Yin, X. Liu, L. Sheng, J. Shao, et al. Worldsimbench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024

    2024 arXiv

  27. [34]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image syn- thesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 10

    2022

  28. [35]

    B. Chen, D. Mart ´ı Mons ´o, Y . Du, M. Simchowitz, R. Tedrake, and V . Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems, 37:24081–24125, 2024

    2024

  29. [36]

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shechtman, and X. Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024

    2024

  30. [37]

    Y . Guo, C. Yang, A. Rao, Z. Liang, Y . Wang, Y . Qiao, M. Agrawala, D. Lin, and B. Dai. An- imatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023

    2023 arXiv

  31. [38]

    H. Lu, G. Yang, N. Fei, Y . Huo, Z. Lu, P. Luo, and M. Ding. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311, 2023

    2023 arXiv

  32. [39]

    X. Ma, Y . Wang, G. Jia, X. Chen, Z. Liu, Y .-F. Li, C. Chen, and Y . Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024

    2024 arXiv

  33. [40]

    W. Ren, H. Yang, G. Zhang, C. Wei, X. Du, W. Huang, and W. Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024

    2024 arXiv

  34. [41]

    J. Yu, Y . Qin, H. Che, Q. Liu, X. Wang, P. Wan, D. Zhang, and X. Liu. Position: Interactive generative video as next-generation game engine. arXiv preprint arXiv:2503.17359, 2025

    2025 arXiv

  35. [42]

    J. Yu, Y . Qin, H. Che, Q. Liu, X. Wang, P. Wan, D. Zhang, K. Gai, H. Chen, and X. Liu. A survey of interactive generative video. arXiv preprint arXiv:2504.21853, 2025

    2025 arXiv

  36. [43]

    J. Yu, Y . Qin, X. Wang, P. Wan, D. Zhang, and X. Liu. Gamefactory: Creating new games with generative interactive videos. arXiv preprint arXiv:2501.08325, 2025

    2025

  37. [44]

    J. Yu, J. Bai, Y . Qin, Q. Liu, X. Wang, P. Wan, D. Zhang, and X. Liu. Context as mem- ory: Scene-consistent interactive long video generation with memory retrieval. arXiv preprint arXiv:2506.03141, 2025

    2025 arXiv

  38. [45]

    K. Gao, J. Shi, H. Zhang, C. Wang, J. Xiao, and L. Chen. Ca2-vdm: Efficient autore- gressive video diffusion model with causal generation and cache sharing. arXiv preprint arXiv:2411.16375, 2024

    2024 arXiv

  39. [47]

    Rajeswaran, V

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstra- tions. arXiv preprint arXiv:1709.10087, 2017

    2017 arXiv

  40. [48]

    C. Bao, H. Xu, Y . Qin, and X. Wang. Dexart: Benchmarking generalizable dexterous manip- ulation with articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21190–21200, 2023

    2023

  41. [49]

    ∗1 - ∗2 - ∗3

    T. Yu, D. Quillen, Z. He, R. Julian, K. Hausman, C. Finn, and S. Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094–1100. PMLR, 2020. 11 A Simulation Experiments Details A.1 Experiment Setu...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.