Pith. sign in

REVIEW 4 major objections 3 minor 88 references

World Action Planner claims that letting a vision-language model propose and refine actions inside an action-conditioned world model beats end-to-end imitation policies on compositional, novel-layout, and zero-shot manipulation tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 04:54 UTC pith:HYVZR6BG

load-bearing objection Genuinely new integration of VLM reasoning with a pose-image world model; strong simulation results, but the known-inaccurate pose conditioning under contact is never stress-tested. the 4 major comments →

arxiv 2607.27599 v1 pith:HYVZR6BG submitted 2026-07-30 cs.AI cs.RO

World Action Planner: Generalizable Decision-Making with Action-Conditioned World Models

Xiangcheng Zhang , Yilun Du This is my paper
classification cs.AI cs.RO
keywords robot planningworld modelsvision-language modelsgeneralizationimitation learningmodel-based planningaction conditioningmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the bottleneck for robot generalization is not data or policy architecture but the absence of test-time reasoning about physical consequences. It introduces World Action Planner, which lets a vision-language model propose primitive actions, then imagines those actions in a learned, action-conditioned world model and iteratively optimizes and searches over candidate actions using feedback and ranking. The central claim, backed by experiments on long-horizon composed tasks, shifted object layouts, and zero-shot manipulation, is that this planning loop massively outperforms end-to-end imitation policies, which often fail entirely once the test scene or task departs from training. The paper also gives theoretical arguments that model-based planning shares dynamics data across tasks and therefore has a sample-complexity advantage over imitation learning in multi-task settings.

Core claim

The paper claims that a robot can solve novel manipulation problems by reasoning over imagined futures rather than regurgitating training motions. Its core discovery is that an action-conditioned world model, conditioned not on low-dimensional action vectors but on pose skeleton images computed from forward dynamics of the proposed actions, produces rollouts that a vision-language model can inspect, critique, and rank. On four compositional long-horizon tasks composed from seen primitives, the planner achieves 72-78% success where end-to-end baselines score near zero; on six tasks with shifted object layouts it achieves 66-90% using only five demonstrations per task; and on two zero-shot man

What carries the argument

The central object is the pose-image conditioned world model: a multi-view video diffusion model that takes as conditioning clean pose skeleton images of the robot, rendered by executing the proposed actions through forward dynamics, and predicts future video frames. The pose skeletons are computed from the robot's own kinematics, not learned from pixels, so the model can imagine the physical consequences of arbitrary, even out-of-distribution, actions. This imagined rollout is then fed back to a VLM agent in three ways: to propose initial primitive moves, to provide high-level corrective feedback on safety and clearance via 'global optimization', and to rank a grid of candidate actions via

Load-bearing premise

The whole pipeline depends on the world model's imagined rollouts being faithful enough to real physics—especially contact events—that a VLM's feedback and ranking over imagined futures selects actions that also succeed in the real environment; the paper itself notes that forward-dynamics-computed pose images misrepresent contact (e.g., a gripper shown as closed while actually holding an object), which could make imagined outcomes misleading exactly in the fine-grained graspi

What would settle it

Run the planner on a set of contact-rich manipulation tasks (inserting, pressing, stacking with tight tolerances) while logging both the world model's imagined future frames and the true environment rollouts. If the VLM's chosen candidate correlates no better than chance with ground-truth success, or if replacing the learned world model with ground-truth simulator rollouts dramatically improves success while the learned model does not, then the claim's load-bearing premise—faithful imagination—is refuted.

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

If this is right

  • End-to-end policies that memorize training motions will continue to fail when scenes or task compositions change; a planning loop that reasons over imagined consequences is a viable alternative for out-of-distribution manipulation.
  • The VLM does not need to output precise coordinates; it only needs to rank candidate outcomes, a capability that combines with the world model to provide metric grounding.
  • Policy learning becomes far more sample-efficient: with a good world model and VLM planner, a handful of demonstrations per task suffices, because the policy only handles local, in-distribution atomic skills.
  • Model-based exploration across tasks amortizes data collection: a shared dynamics model makes each new task's reward function cheap to learn, whereas imitation learning must re-learn a distinct policy per task.

Where Pith is reading between the lines

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

  • The 'policies as tools' framing suggests a modular robot-learning stack where local policies handle in-distribution atomic skills and the VLM-plus-world-model planner orchestrates them; this could combine the robustness of specialized policies with the flexibility of planning.
  • The pose-image conditioning trick is transferable: any controlled generative video model could be steered by rendering the agent's articulated state as a skeleton or keypoint overlay, without retraining the conditioning mechanism.
  • A natural stress test follows from the paper's own caveat: tasks with heavy contact forces (deformable objects, tight fits, regrasping under load) are where the forward-dynamics pose approximation should break down most; measuring planning success across a contact-richness axis would reveal the boundary of the claim.
  • The theoretical separation suggests that scaling the VLM's ability to judge task success (as a reward/evaluator model) may be a more efficient investment than scaling demonstration data or policy networks.

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

4 major / 3 minor

Summary. The paper proposes World Action Planner (WAP), a robot planning system that combines a Vision-Language Model (VLM) agent with a pose-image-conditioned video diffusion world model. The VLM proposes primitive actions, which are refined through (i) global optimization based on VLM feedback on imagined rollouts, and (ii) local grid search with VLM candidate ranking, optionally including imagined rollouts of a learned policy. The authors report large success-rate gains over end-to-end imitation baselines (π0.5, cosmos-policy) and over prior world-model-based policy-enhancement methods on LIBERO compositional tasks, LIBERO layout-shifted tasks, and Robosuite zero-shot tasks. The paper also provides theoretical results claiming model-based planning outperforms imitation learning in multi-task generalization in tabular and linear-function-approximation settings.

Significance. If the empirical claims hold, WAP would be a meaningful step toward test-time planning with world models and VLM reasoning, especially for compositional and layout-shifted manipulation where end-to-end policies fail. Strengths include the comparatively extensive world-model evaluation across four simulation suites and multiple embodiments, the ablations isolating global optimization, local search, and policy-rollout imagination, and a best-of-N comparison that uses ground-truth rewards as an upper bound. The theoretical appendix is detailed and builds on standard reward-free exploration and imitation-learning lower-bound techniques. However, as detailed below, several load-bearing issues remain: the physically inaccurate pose conditioning is not stress-tested, the theoretical comparison is asymmetric in information, the empirical headline lacks statistical support, and the 'zero-shot' claim is weakened by task-specific world-model fine-tuning.

major comments (4)
  1. [C.1.1, Fig. 6, Alg. 1] The paper explicitly acknowledges that forward-dynamics pose skeletons are inaccurate under contact: the gripper is rendered closed while an object between the jaws keeps it open, and reaction forces are unknown a priori. Since the planning loop in Alg. 1 uses only the world model's imagined rollouts for VLM feedback and ranking, and since the largest gains are in contact-rich tasks (Tables 4, 5, 9), the absence of any ablation using ground-truth or contact-aware poses leaves the physical-grounding mechanism unverified. Please add an oracle experiment with simulator ground-truth joint poses (or contact-aware forward simulation) to demonstrate that the performance does not rely on systematically wrong imagined outcomes.
  2. [B.2.1, Theorems 1-3] The theoretical comparison is not apples-to-apples. In the protocol, the imitation learning agent receives only expert action labels and never observes rewards, while the model-based agent collects rewards during exploration and, in the tabular theorem, is given the reward functions for all contexts. The Ω(|C|/K) lower bound for imitation learning therefore may reflect an information-structure advantage rather than an intrinsic superiority of model-based planning. Please either provide the IL agent with the same reward information, or explicitly reframe the results as a comparison between 'IL without rewards' and 'MB with rewards/known rewards.' This is important because contribution (ii) is stated as a general separation.
  3. [Tables 3-5 and §5.2] Success rates are reported as single 50-trial counts with no error bars, no multiple seeds, and no statistical tests. The abstract's 'significantly outperforming' is therefore not statistically supported, especially for the smaller gaps in Table 5 (PickPlaceCan 80 vs 58). Please provide multiple seeds or confidence intervals, state the number of runs, and use a significance test where appropriate. Also, the baseline protocol varies across experiments: in §5.2.1 baselines are fine-tuned by the authors, while in §5.2.2 official checkpoints are used; details for adapting SAILOR and GPC-RANK are sparse.
  4. [§5.2.3 and D.2.3] The label 'zero-shot generalization' is overstated. Although no expert demonstrations are used, the world model is fine-tuned on 50 task-specific exploratory trajectories for PickPlaceCan and StackCube. This is not zero-shot with respect to the target task. Please either rename the setting (e.g., 'demonstration-free' or 'no-policy planning') or include a truly zero-shot evaluation in which the world model is not fine-tuned on the target task, to support the abstract's claim of 'zero-shot generalization scenarios.'
minor comments (3)
  1. [Throughout] Typos and spacing issues: 'a action-conditioned', 'W AMs', 'receices', 'sufficies', 'the the' in places. A careful proofread is needed.
  2. [Table 1] The presentation of PSNR/LPIPS pairs is dense; a separate table or clearer column separation would improve readability. It is also worth stating explicitly which metrics are higher-better versus lower-better in the caption.
  3. [Appendix B] Some notation is introduced inconsistently (e.g., 'E[h...]' in a few displayed equations). These do not affect the logic but should be cleaned for publication.

Circularity Check

0 steps flagged

No significant circularity: the planner's reported success rates are measured by external simulators, and the world-model/planning claims do not reduce to fitted constants or self-citations.

full rationale

The paper's central claim is empirical: WAP outperforms π0.5, cosmos-policy, SAILOR, GPC-RANK, and a vision-language planner on LIBERO and Robosuite tasks, with success rates evaluated by the simulators (Tables 3–5). Those numbers are not derived from the paper's equations or from fitted parameters renamed as predictions; the world model is independently evaluated with PSNR/LPIPS against held-out trajectories (Table 1), and the planner is ablated against environment-grounded best-of-N sampling (Tables 10–11). The theoretical results (Theorems 1–3) are standard sample-complexity separations with proofs in Appendix B that do not assume the empirical conclusion. The authors cite their own prior work ([45], [56], [83]) for primitive-action prompting, a baseline protocol, and a global-local search heuristic, but these are design borrowings, not load-bearing justifications of the performance claims; no uniqueness theorem or fitted ansatz is imported from self-citations. The acknowledged limitation in C.1.1 and Fig. 6 — pose skeletons are inaccurate under contact because forward dynamics omit reaction forces — is a genuine correctness risk for the claimed mechanism, not a circularity: it concerns whether imagined rollouts are physically faithful, not whether the output reduces to the input by construction. The VLM both proposes and ranks candidates, which can bias selection, but the final metric is simulator success, so this is a robustness concern rather than a definitional equivalence. Overall, the derivation chain is self-contained against external benchmarks; any circularity burden is limited to minor, non-load-bearing self-citations, hence score 1.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The system's empirical success rests on training-data hyperparameters (noise scales, data budgets) and on two unproven-in-the-paper assumptions: forward-dynamics pose skeletons remain useful during contact, and a closed-source VLM can act as a reliable proposer/critic/ranker. The theoretical section adds known-reward, shared-dynamics, and linear-representation assumptions that are standard but stronger than the experimental setting.

free parameters (4)
  • global optimization modification scale = 0.06
    Prompted directional Delta changes are multiplied by a hand-set step size; not fitted to data, but chosen by authors and affects whether feedback corrects or overshoots.
  • grid search scale delta = 0.02 (StackCube), 0.05 (other tasks)
    Hand-set to roughly half object size; defines the candidate neighborhood in local search and directly affects fine-grained grasping and placement success.
  • exploratory action noise std = sigma = 0.08 and 0.16 (40/40/20 split)
    Gaussian action perturbations used to generate world-model training trajectories; noise levels and split are chosen manually and affect OOD imagination quality.
  • demonstration budgets = 30 per task (policy), 100 per task (world model); 5 per task for new-layout policy
    Hand-selected data budgets; the sample-efficiency comparison against baselines depends on these numbers, especially the 5-demonstration new-layout setup.
axioms (6)
  • domain assumption Known rewards r_c for each context (Theorem 1)
    The model-based tabular result assumes the reward function for every task is given; in practice the VLM supplies rewards/feedback, which is not part of the formal proof.
  • domain assumption Shared universal dynamics P across tasks
    The contextual MDP setup assumes all tasks share the same transition kernel, which the theory requires but real manipulation tasks may violate.
  • domain assumption Linear reward representation E[r_c(s,a)] = phi(s,a)^T Theta psi(c) with known task features psi(c)
    Assumption 1 in Appendix B.4; needed for reward generalization in Theorem 2 and not verified in the experiments.
  • standard math Exploratory assumption kappa > 0 in linear MDPs
    Assumption 2, standard in linear RL literature [37,39]; ensures feature-space exploration for the model-based bound.
  • ad hoc to paper Forward-dynamics pose skeletons are adequate conditioning despite contact inaccuracy
    C.1.1: the method's central representation is intentionally not exact physics, and the paper does not quantify how much error the world model can tolerate before planning degrades.
  • domain assumption VLM can propose, critique, and rank actions from images reliably enough
    The whole planner depends on Gemini or similar closed-source VLM capabilities; the provided prompt templates are detailed but fragile to model version and prompt wording.

pith-pipeline@v1.3.0-daily-deepseek · 34513 in / 13926 out tokens · 161633 ms · 2026-08-01T04:54:10.376709+00:00 · methodology

0 comments
read the original abstract

Building generalizable agents for diverse applications remains a fundamental challenge. While imitation learning-based policies succeed in specific training environments, they often fail to generalize to novel scenes and tasks. In this work, we propose World Action Planner, a robot planning system that leverages the reasoning capabilities of Vision-Language Models (VLMs) and the physical grounding of a multi-task pose-image conditioned world model. Our system enables an agent to propose initial action plans and iteratively refine them via optimization and search, reasoning over imagined world model rollouts. We demonstrate that our approach achieves superior performance across compositional tasks, new layouts, and zero-shot generalization scenarios, significantly outperforming state-of-the-art end-to-end policy models such as VLAs and WAMs. Project website at worldactionplanner.github.io

Figures

Figures reproduced from arXiv: 2607.27599 by Xiangcheng Zhang, Yilun Du.

Figure 1
Figure 1. Figure 1: We introduce World Action Planner, a planning system in which the agent reasons about and refines its action plans via world model imagination, generalizing to novel scenarios and solving new tasks at test time. 1 Introduction Developing generalizable robotic agents capable of mastering diverse tasks in heterogeneous environments remains a fundamental challenge in AI. Recent progress has been driven by end… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of our pose-image conditioned world model Prior robot action-conditioned video diffusion models typically condition on low-dimensional action vectors through AdaLN-Zero modulation [57, 86] or cross-attention [14]. However, such conditioning often fails to generalize to novel actions. In our approach, we pre-compute future robot joint positions via forward dynamics of the actions and render the… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of world action planner in compositional task generalization Method Task PnP alphabet soup & tomato sauce PnP white mug & yellow and white mug PnP white mug & chocolate pudding PnP alphabet soup & cream cheese box π0.5 4 0 0 0 cosmos-policy 0 0 0 0 SAILOR 18 0 8 2 GPC-RANK 10 0 0 0 Vision-language planner 56 28 46 32 World action planner 72 68 78 70 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of new layout generalization A generalizable robot must adapt to novel object configurations at test time, despite being trained on a limited distribution of layout samples. To evaluate this capability, we utilize six tasks from the LIBERO-Object suite involving pick-and-place operations, where we introduce significant layout shifts by varying the positions of both target and distractor object… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of world action planner with zero-shot planning in StackCube. Method Task PickPlaceCan StackCube Vision-language planner 58 22 World action planner 80 76 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of inaccurate pose image. In the image, the first row is the actual environment rollout and the second row is the computed pose image using forward dynamics. Because the gripper action is close for every step, in the computed robot joint pose, the grippers jaws are closed immediately. In actual execution, due to the presence of the cube between the jaws, the opening between the jaws remain unc… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results for OOD actions that knocks the mug over, with the first row being ground truth [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results cross-embodiment modeling, with the first row being ground truth frames with red [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Original object layout and our modified object layout in LIBERO-Object. Yellow arrow points to the [PITH_FULL_IMAGE:figures/full_fig_p032_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative evaluation of π0.5 on the modified layout of PnP chocolate pudding, down-sampled from 20 FPS to 1 FPS for clarity. The VLA exhibits low success rates, as the gripper rarely drifts toward the correct target and instead frequently grasps distractor objects placed at the original position during re-grasp attempts [PITH_FULL_IMAGE:figures/full_fig_p033_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative evaluation of the cosmos-policy on the modified layout of the PnP chocolate pudding [PITH_FULL_IMAGE:figures/full_fig_p033_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustration of local search for gripper rotation. The VLM successfully identifies the right rotation [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Illustration imagining consecutive policy rollout. While in the first row, it is hard to identify which [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Global Action Optimization. While VLMs often ignore physical constraints, our framework leverages world model imaginations to identify potential collisions and optimize action sequences for safe execution. We also compare our local search with GridSearch exclusively against the BoN baseline in the StackCube task, where global optimization for physical clearance is not essential and pure local search alone… view at source ↗
Figure 15
Figure 15. Figure 15: Example prompt template for agent action proposal. [PITH_FULL_IMAGE:figures/full_fig_p037_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Example prompt template for agent feedback for action optimization. [PITH_FULL_IMAGE:figures/full_fig_p038_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Example prompt template for agent candidate ranking for local search. [PITH_FULL_IMAGE:figures/full_fig_p039_17.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

88 extracted references · 45 linked inside Pith

  1. [1]

    Abbasi-Yadkori, D

    Y . Abbasi-Yadkori, D. Pál, and C. Szepesvári. Improved algorithms for linear stochastic bandits. Advances in neural information processing systems, 24, 2011. 21, 26

    2011

  2. [2]

    Agarwal, A

    N. Agarwal, A. Ali, M. Bala, Y . Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y . Chen, Y . Cui, Y . Ding, et al. Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575, 2025. 2, 16

    Pith/arXiv arXiv 2025

  3. [3]

    A. Ali, J. Bai, M. Bala, Y . Balaji, A. Blakeman, T. Cai, J. Cao, T. Cao, E. Cha, Y .-W. Chao, et al. World simulation with video foundation models for physical ai.arXiv preprint arXiv:2511.00062, 2025. 2, 16

    Pith/arXiv arXiv 2025

  4. [4]

    Alonso, A

    E. Alonso, A. Jelley, V . Micheli, A. Kanervisto, A. Storkey, T. Pearce, and F. Fleuret. Diffusion for world modeling: Visual details matter in atari.Advances in Neural Information Processing Systems, 37:58757–58791, 2024. 16

    2024

  5. [5]

    Assran, A

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025. 5, 16

    Pith/arXiv arXiv 2025

  6. [6]

    Bjorck, F

    J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025. 6

    Pith/arXiv arXiv 2025

  7. [7]

    Bonnen, J

    T. Bonnen, J. Malik, and A. Kanazawa. Human-level 3d shape perception emerges from multi-view learning.arXiv preprint arXiv:2602.17650, 2026. 3

    arXiv 2026

  8. [8]

    Bruce, M

    J. Bruce, M. D. Dennis, A. Edwards, J. Parker-Holder, Y . Shi, E. Hughes, M. Lai, A. Mavalankar, R. Steigerwald, C. Apps, et al. Genie: Generative interactive environments. InForty-first International Conference on Machine Learning, 2024. 16

    2024

  9. [9]

    B. Chen, T. Zhang, H. Geng, K. Song, C. Zhang, P. Li, W. T. Freeman, J. Malik, P. Abbeel, R. Tedrake, et al. Large video planner enables generalizable robot control.arXiv preprint arXiv:2512.15840, 2025. 3, 16, 17

    Pith/arXiv arXiv 2025

  10. [10]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025. 4, 16, 28, 30 10

    2025

  11. [11]

    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. 3, 16

    Pith/arXiv arXiv 2023

  12. [12]

    J. Duan, W. Yuan, W. Pumacay, Y . R. Wang, K. Ehsani, D. Fox, and R. Krishna. Manipulate- anything: Automating real-world robots using vision-language models.arXiv preprint arXiv:2406.18915, 2024. 16

    Pith/arXiv arXiv 2024

  13. [13]

    Ghosh, J

    D. Ghosh, J. Rahme, A. Kumar, A. Zhang, R. P. Adams, and S. Levine. Why generalization in rl is difficult: Epistemic pomdps and implicit partial observability.Advances in neural information processing systems, 34:25502–25515, 2021. 17

    2021

  14. [14]

    Y . Guo, L. X. Shi, J. Chen, and C. Finn. Ctrl-world: A controllable generative world model for robot manipulation.arXiv preprint arXiv:2510.10125, 2025. 2, 3, 6, 16, 29

    Pith/arXiv arXiv 2025

  15. [15]

    Hafner, T

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi. Dream to control: Learning behaviors by latent imagination.arXiv preprint arXiv:1912.01603, 2019. 3, 16

    Pith/arXiv arXiv 1912

  16. [16]

    Hafner, T

    D. Hafner, T. Lillicrap, M. Norouzi, and J. Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 16

    Pith/arXiv arXiv 2010

  17. [17]

    Hafner, J

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap. Mastering diverse domains through world models.arXiv preprint arXiv:2301.04104, 2023. 16

    Pith/arXiv arXiv 2023

  18. [18]

    Hallak, D

    A. Hallak, D. Di Castro, and S. Mannor. Contextual markov decision processes.arXiv preprint arXiv:1502.02259, 2015. 5, 17

    Pith/arXiv arXiv 2015

  19. [19]

    Hansen, H

    N. Hansen, H. Su, and X. Wang. Td-mpc2: Scalable, robust world models for continuous control.arXiv preprint arXiv:2310.16828, 2023. 3, 6, 16

    Pith/arXiv arXiv 2023

  20. [20]

    Hansen, X

    N. Hansen, X. Wang, and H. Su. Temporal difference learning for model predictive control. arXiv preprint arXiv:2203.04955, 2022. 16

    Pith/arXiv arXiv 2022

  21. [21]

    X. He, H. Yu, S. Peng, D. Tan, Z. Shen, H. Bao, and X. Zhou. Matchanything: Universal cross-modality image matching with large-scale pre-training.arXiv preprint arXiv:2501.07556,

    Pith/arXiv arXiv

  22. [22]

    Hore and D

    A. Hore and D. Ziou. Image quality metrics: Psnr vs. ssim. In2010 20th international conference on pattern recognition, pages 2366–2369. IEEE, 2010. 6

    2010

  23. [23]

    J. Hu, X. Chen, C. Jin, L. Li, and L. Wang. Near-optimal representation learning for linear bandits and linear rl. InInternational Conference on Machine Learning, pages 4349–4358. PMLR, 2021. 17

    2021

  24. [24]

    P. Hu, Y . Chen, and L. Huang. Nearly minimax optimal reinforcement learning with linear function approximation. InInternational Conference on Machine Learning, pages 8971–9019. PMLR, 2022. 18

    2022

  25. [25]

    Y . Hu, F. Lin, T. Zhang, L. Yi, and Y . Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning.arXiv preprint arXiv:2311.17842, 2023. 2, 16

    Pith/arXiv arXiv 2023

  26. [26]

    Huang, C

    W. Huang, C. Wang, R. Zhang, Y . Li, J. Wu, and L. Fei-Fei. V oxposer: Composable 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023. 2, 16

    Pith/arXiv arXiv 2023

  27. [27]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al. π0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025. 1, 3, 4, 7, 16, 17

    Pith/arXiv arXiv 2025

  28. [28]

    Jaech, A

    A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 36

    Pith/arXiv arXiv 2024

  29. [29]

    A. K. Jain, V . Mohta, S. Kim, A. Bhardwaj, J. Ren, Y . Feng, S. Choudhury, and G. Swamy. A smooth sea never made a skilled sailor: Robust imitation via learning to search.arXiv preprint arXiv:2506.05294, 2025. 3, 6, 7, 9, 16, 29, 32, 33, 34 11

    arXiv 2025

  30. [30]

    J. Jang, S. Ye, Z. Lin, J. Xiang, J. Bjorck, Y . Fang, F. Hu, S. Huang, K. Kundalia, Y .-C. Lin, et al. Dreamgen: Unlocking generalization in robot learning through video world models.arXiv preprint arXiv:2505.12705, 2025. 16

    Pith/arXiv arXiv 2025

  31. [31]

    E. Y .-T. Jia, W. Yuan, T. Shi, V . Guizilini, J. Mao, and Y . Wang. Dreamplan: Efficient reinforcement fine-tuning of vision-language planners via video world models.arXiv preprint arXiv:2603.16860, 2026. 2, 3, 16

    arXiv 2026

  32. [32]

    Jiang, Y

    Z. Jiang, Y . Xie, K. Lin, Z. Xu, W. Wan, A. Mandlekar, L. J. Fan, and Y . Zhu. Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 16923–16930. IEEE, 2025. 6

    2025

  33. [33]

    C. Jin, A. Krishnamurthy, M. Simchowitz, and T. Yu. Reward-free exploration for reinforcement learning. InInternational Conference on Machine Learning, pages 4870–4879. PMLR, 2020. 18, 19

    2020

  34. [34]

    C. Jin, Z. Yang, Z. Wang, and M. I. Jordan. Provably efficient reinforcement learning with linear function approximation.Mathematics of Operations Research, 48(3):1496–1521, 2023. 5, 20, 21, 26, 27

    2023

  35. [35]

    M. J. Kim, Y . Gao, T.-Y . Lin, Y .-C. Lin, Y . Ge, G. Lam, P. Liang, S. Song, M.-Y . Liu, C. Finn, et al. Cosmos policy: Fine-tuning video models for visuomotor control and planning.arXiv preprint arXiv:2601.16163, 2026. 3, 7, 16, 30

    Pith/arXiv arXiv 2026

  36. [36]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024. 3, 16, 30, 31

    Pith/arXiv arXiv 2024

  37. [37]

    F. Kong, X. Zhang, B. Wang, and S. Li. Improved regret bounds for linear adversarial mdps via linear optimization.arXiv preprint arXiv:2302.06834, 2023. 5, 21

    Pith/arXiv arXiv 2023

  38. [38]

    J. Kwon, Y . Efroni, C. Caramanis, and S. Mannor. Reward-mixing mdps with few latent contexts are learnable. InInternational Conference on Machine Learning, pages 18057–18082. PMLR,

  39. [39]

    G. Li, Y . Wei, Y . Chi, Y . Gu, and Y . Chen. Sample complexity of asynchronous q-learning: Sharper analysis and variance reduction.Advances in neural information processing systems, 33:7031–7043, 2020. 5, 21

    2020

  40. [40]

    L. Li, Q. Zhang, Y . Luo, S. Yang, R. Wang, F. Han, M. Yu, Z. Gao, N. Xue, X. Zhu, et al. Causal world modeling for robot control.arXiv preprint arXiv:2601.21998, 2026. 3, 16

    Pith/arXiv arXiv 2026

  41. [41]

    Z. Li, H. Liu, D. Zhou, and T. Ma. Chain of thought empowers transformers to solve inherently serial problems. InThe Twelfth International Conference on Learning Representations, 2024. 36

    2024

  42. [42]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022. 3

    Pith/arXiv arXiv 2022

  43. [43]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023. 6, 7

    2023

  44. [44]

    F. Liu, K. Fang, P. Abbeel, and S. Levine. Moka: Open-world robotic manipulation through mark-based visual prompting.arXiv preprint arXiv:2403.03174, 2024. 2, 3, 16

    Pith/arXiv arXiv 2024

  45. [45]

    H. Liu, S. Yao, H. Chen, J. Gao, J. Mao, J.-B. Huang, and Y . Du. Simpact: Simulation-enabled action planning using vision-language models.arXiv preprint arXiv:2512.05955, 2025. 3

    arXiv 2025

  46. [46]

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024. 6

    Pith/arXiv arXiv 2024

  47. [47]

    Lozano-Perez

    T. Lozano-Perez. Robot programming.Proceedings of the IEEE, 71(7):821–841, 2005. 2 12

    2005

  48. [48]

    R. Lu, G. Huang, and S. S. Du. On the power of multitask representation learning in linear mdp. arXiv preprint arXiv:2106.08053, 2021. 17

    Pith/arXiv arXiv 2021

  49. [49]

    Malik, Y

    D. Malik, Y . Li, and P. Ravikumar. When is generalizable reinforcement learning tractable? Advances in Neural Information Processing Systems, 34:8032–8045, 2021. 6, 17

    2021

  50. [50]

    Mandlekar, S

    A. Mandlekar, S. Nasiriany, B. Wen, I. Akinola, Y . Narang, L. Fan, Y . Zhu, and D. Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. 6

    Pith/arXiv arXiv 2023

  51. [51]

    Nasiriany, A

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y . Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523, 2024. 6

    Pith/arXiv arXiv 2024

  52. [52]

    Nasiriany, F

    S. Nasiriany, F. Xia, W. Yu, T. Xiao, J. Liang, I. Dasgupta, A. Xie, D. Driess, A. Wahid, Z. Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms.arXiv preprint arXiv:2402.07872, 2024. 16

    Pith/arXiv arXiv 2024

  53. [53]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5

    Pith/arXiv arXiv 2023

  54. [54]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. 6

    2023

  55. [55]

    Pichai, D

    S. Pichai, D. Hassabis, and K. Kavukcuoglu. A new era of intelligence with gemini 3.Google Blog, November, 18, 2025. 7

    2025

  56. [56]

    H. Qi, H. Yin, Y . Du, and H. Yang. Strengthening generative robot policies through predictive world modeling.arXiv e-prints, pages arXiv–2502, 2025. 3, 7, 9, 32, 33, 34

    2025

  57. [57]

    Quevedo, A

    J. Quevedo, A. K. Sharma, Y . Sun, V . Suryavanshi, P. Liang, and S. Yang. Worldgym: World model as an environment for policy evaluation.arXiv preprint arXiv:2506.00613, 2025. 2, 3, 6, 16

    arXiv 2025

  58. [58]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 5

    2021

  59. [59]

    Rajaraman, Y

    N. Rajaraman, Y . Han, L. Yang, J. Liu, J. Jiao, and K. Ramchandran. On the value of interaction and function approximation in imitation learning.Advances in Neural Information Processing Systems, 34:1325–1336, 2021. 6, 17, 25

    2021

  60. [60]

    Rajaraman, L

    N. Rajaraman, L. Yang, J. Jiao, and K. Ramchandran. Toward the fundamental limits of imitation learning.Advances in Neural Information Processing Systems, 33:2914–2924, 2020. 17, 19, 20

    2020

  61. [61]

    V . Sanh, L. Debut, J. Chaumond, and T. Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108, 2019. 30

    Pith/arXiv arXiv 1910

  62. [62]

    A. K. Sharma, Y . Sun, N. Lu, Y . Zhang, J. Liu, and S. Yang. World-gymnast: Training robots with reinforcement learning in a world model.arXiv preprint arXiv:2602.02454, 2026. 16

    arXiv 2026

  63. [63]

    K. Song, B. Chen, M. Simchowitz, Y . Du, R. Tedrake, and V . Sitzmann. History-guided video diffusion.arXiv preprint arXiv:2502.06764, 2025. 3, 28

    Pith/arXiv arXiv 2025

  64. [64]

    W. Sun, N. Jiang, A. Krishnamurthy, A. Agarwal, and J. Langford. Model-based rl in contextual decision processes: Pac bounds and exponential improvements over model-free approaches. In Conference on learning theory, pages 2898–2933. PMLR, 2019. 5, 17

    2019

  65. [65]

    Z. Tan, S. Liu, X. Yang, Q. Xue, and X. Wang. Ominicontrol: Minimal and universal control for diffusion transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14940–14950, 2025. 3 13

    2025

  66. [66]

    G. R. Team, K. Choromanski, C. Devin, Y . Du, D. Dwibedi, R. Gao, A. Jindal, T. Kipf, S. Kirmani, I. Leal, et al. Evaluating gemini robotics policies in a veo world simulator.arXiv preprint arXiv:2512.10675, 2025. 2, 16

    arXiv 2025

  67. [67]

    R. A. Team. Causal video models are data-efficient robot policy learners.Rhoda AI Blog, 2026. 3

    2026

  68. [68]

    A. J. Wagenmaker, Y . Chen, M. Simchowitz, S. Du, and K. Jamieson. Reward-free rl is no harder than reward-aware rl in linear markov decision processes. InInternational Conference on Machine Learning, pages 22430–22456. PMLR, 2022. 21

    2022

  69. [69]

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Pith/arXiv arXiv

  70. [70]

    Y . Wang, C. Wen, H. Guo, S. Peng, M. Qin, H. Bao, X. Zhou, and R. Hu. Precise action-to- video generation through visual action prompts. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12713–12724, 2025. 2, 16

    2025

  71. [71]

    Weltevrede, C

    M. Weltevrede, C. Horsch, M. T. Spaan, and W. Böhmer. Exploration implies data augmentation: Reachability and generalisation in contextual mdps.arXiv preprint arXiv:2410.03565, 2024. 17

    Pith/arXiv arXiv 2024

  72. [72]

    Wiedemer, Y

    T. Wiedemer, Y . Li, P. Vicol, S. S. Gu, N. Matarese, K. Swersky, B. Kim, P. Jaini, and R. Geirhos. Video models are zero-shot learners and reasoners.arXiv preprint arXiv:2509.20328, 2025. 28

    Pith/arXiv arXiv 2025

  73. [73]

    Williams, A

    G. Williams, A. Aldrich, and E. A. Theodorou. Model predictive path integral control: From theory to parallel computation.Journal of Guidance, Control, and Dynamics, 40(2):344–357,

  74. [74]

    J. Wu, S. Yin, N. Feng, X. He, D. Li, J. Hao, and M. Long. ivideogpt: Interactive videogpts are scalable world models.Advances in Neural Information Processing Systems, 37:68082–68119,

  75. [75]

    S. Yang, Y . Du, K. Ghasemipour, J. Tompson, L. Kaelbling, D. Schuurmans, and P. Abbeel. Learning interactive real-world simulators.arXiv preprint arXiv:2310.06114, 2023. 2

    Pith/arXiv arXiv 2023

  76. [76]

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 2

    Pith/arXiv arXiv 2024

  77. [77]

    H. Ye, X. Chen, L. Wang, and S. S. Du. On the power of pre-training for generalization in rl: provable benefits and hardness. InInternational Conference on Machine Learning, pages 39770–39800. PMLR, 2023. 17

    2023

  78. [78]

    S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y . L. Tan, C. Zhu, J. Xiang, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, 2026. 1, 3, 4, 16

    Pith/arXiv arXiv 2026

  79. [79]

    Young, A

    K. Young, A. Ramesh, L. Kirsch, and J. Schmidhuber. The benefits of model-based generaliza- tion in reinforcement learning.arXiv preprint arXiv:2211.02222, 2022. 17

    Pith/arXiv arXiv 2022

  80. [80]

    differentiation

    M. Yuksekgonul, F. Bianchi, J. Boen, S. Liu, Z. Huang, C. Guestrin, and J. Zou. Textgrad: Automatic" differentiation" via text.arXiv preprint arXiv:2406.07496, 2024. 4

    Pith/arXiv arXiv 2024

Showing first 80 references.