Pith.

Robotics · arXiv:2610.07922 v1 · submitted 6 Oct 2026 · 109 references

OpenWAM: An Open Framework for Composable World-Action Models

Heng Yu, David D. Yuan, Juze Zhang, Changan Chen, Yao Feng, Michelle Baldonado, Steve Cousins, Li Fei-Fei

show 2 more authors

Jiajun Wu, Ehsan Adeli

The paper says

“World-action models (WAMs) couple future prediction with robot control, yet existing systems often vary the video backbone, interaction structure, supervision, and inference procedure simultaneously, making their design choices difficult to compare.”

more of the abstract

We introduce OPENWAM, an open world-action modeling framework built around a common causal robot-video foundation and configurable video-action interaction. Starting from Wan2.2-5B, we perform causal robot-video pretraining on over 10,000 hours of video, then integrate an action expert through a shared Mixture-of-Transformers architecture that supports joint, video-then-action, action-then-video, and decoupled generation. OPENWAM achieves high success rates on four LIBERO suites and real-world bimanual tasks; robot-video training with causal adaptation improves VTA success on LIBERO-Long from 68.4% to 97.8%. The same configurable architecture naturally extends to inverse and forward dynamics, allowing us to study how counterfactual transitions improve independently trained dynamics models beyond demonstrations alone. When only the video predictor is adapted to a new task, a frozen local-context inverse dynamics model trained on counterfactual data and demonstrations achieves 84.0% mean success across four held-out LIBERO-90 tasks, compared with 47.0% for a full-context inverse model and 21.5% for a local-context model trained only on demonstrations. For forward dynamics, counterfactual supervision reduces RGB prediction error by 34.5% and raises outcome identification from 21.1% to 71.3% among 16 same-state outcomes. OPENWAM provides a common testbed for comparing WAM interaction designs and for studying dynamics learning from video data beyond successful demonstrations.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 7 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A useful open WAM testbed with an honest, well-run evaluation; the frozen-IDM transfer claim is plausible but the counterfactual-content attribution needs a scale-matched demo-only control before it is causal. This line is wrong

4 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 109 references say.

The paper says, at Sec. 4.3 / Table 5; Sec. C.1 / Table 9

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The causal reading of the headline — that counterfactual transition content, rather than demonstration content, makes a frozen local-context IDM reusable — is not isolated by the reported comparison. Table 9 reports 138,090 demonstration controls versus 4,096,000 counterfactual controls (a 29.7x gap), and the demo-only local row of Table 5 (21.5%) is trained on the former while the CF-only and …

more of objection 1

… mixed rows (84.5% and 84.0%) are trained on roughly two orders of magnitude more records; the same gap applies to the FDM conditions in Table 7. Because the authors credit broader transition coverage to counterfactual supervision, a scale-matched demo-only leg is required, for example demonstration-derived transitions collected at comparable volume by replaying demonstrated action sequences from perturbed or revisited states. The full-context mixed row (47.0% versus 44.5% demo-only) shows that scale alone did not rescue the full-context model, which is suggestive, but that model differs in conditioning scope and in task-context mismatch, so it does not substitute for the missing local-context control. The paper deserves credit for reporting the 29.7x factor transparently; the missing leg is nevertheless precisely the comparison that would make the attribution causal rather than correlational.

The paper says, at Sec. 4.3 / Eq. (9); Sec. C.2 / Table 11

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The composition in Eq. (9) succeeds only if futures generated by the task-tuned LIBERO-90 video predictors fall inside the support of the transitions seen by the frozen LIBERO-Long IDM. The paper reports aggregate contact statistics for the counterfactual corpus (Table 11) but never measures this match directly: there is no open-loop IDM action error reported on (generated future, ground-truth action) pairs from the four targets, and no comparison of the adapted predictors' output distribution with the counterfactual training distribution.

more of objection 2

This is load-bearing because LIBERO-Long is built from LIBERO-90-style subtasks, so the four held-out targets may share local transitions with the source corpus, and because the video latents passed at inference are produced by the adapted predictor rather than recorded. A concrete test would be to run each task-tuned video predictor on the target demonstrations, feed its generated futures to the frozen IDM, and report per-task action error relative to the demonstrated actions, alongside the same numbers for the demo-only IDM; this would also probe the latent domain gap between generated and recorded conditioning. The authors' caveat that local-context sufficiency is task-dependent (Sec. 3.4) is acknowledged, but the distributional-support premise is a distinct, unacknowledged assumption.

The paper says, at Sec. 4.4 / Table 7

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The FDM headline numbers — a 34.5% MSE reduction and outcome identification rising from 21.1% to 71.3% at K=16 — are computed on the counterfactual evaluation set of 2,560 futures from 160 contexts with 16 same-state branches, which is the same distribution on which the CF-only and mixed FDMs were trained, while the demo-only FDM is evaluated off-distribution.

more of objection 3

The gain therefore reflects in-distribution specialization as much as, or instead of, better action-conditioned dynamics. Table 14 partly addresses this by reporting demo-side MSE, where mixed supervision is indeed superior and CF-only is not; the same complementary reporting should be applied to the outcome-identification metric, and the abstract should be reworded so the reduction is not stated as a general capability without the distribution qualifier.

The paper says, at Tables 5-7; Sec. A.4

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The main policy protocol (Sec. A.4) specifies three seeds with shared environment seeds and 50 episodes per task, but Tables 5 and 6 and Table 7 report single point values without standard deviations, and the text does not state whether the frozen-transfer and FDM conditions follow the same protocol.

more of objection 4

Since the central composability claim and the reading of CF-only (84.5%) and mixed (84.0%) as comparable rest on these numbers, the paper should report per-seed results and episode counts for the transfer conditions, or explicitly amend the protocol statement in Sec. A.4 to cover them.

5 minor objections

The paper says, at Secs. 3.4 and C

Pith says: minor objection 1 of 5

The term 'counterfactual' denotes scripted alternative rollouts branched from shared initial states, not counterfactual inference under an explicit structural model; a one-sentence clarification of this usage would prevent over-reading by causal-inference readers.

The paper says, at Table 3

Pith says: minor objection 2 of 5

Baseline numbers are quoted from prior papers without a shared protocol; a footnote stating that these values are as-reported and were not re-run under the OpenWAM evaluation protocol would improve transparency.

show 3 more

The paper says, at Figure 3

Pith says: minor objection 3 of 5

The caption is very terse; the meaning of the generation-pass numerals and of the 'no attention' legend entries is recoverable only by cross-referencing Sec. A.3, so an expanded caption would help.

The paper says, at Abstract and Sec. 4.3

Pith says: minor objection 4 of 5

The phrase 'when only the video predictor is adapted' invites the reading that adaptation uses no target actions, whereas the task-tuned VTA predictor is adapted with full action supervision on target demonstrations; the discussion acknowledges the correct reading ([b240]), but the abstract should state the supervised-adaptation setting explicitly.

The paper says, at Sec. C.1

Pith says: minor objection 5 of 5

The 29.7x factor conflates control records with distinct starting states (32,000 counterfactual segments versus 500 demonstration sequences); reporting unique-state counts per task alongside control counts would sharpen the scale discussion and directly support the scale-matched control proposed above.

What would settle it

Downsample LIBERO-Long-CF to the demo-only scale and retrain the Table 5 local-context IDM: randomly sample 138,090 transition records from the 4,096,000 CF controls, preserving per-task balance and start-state distribution, then train CF-only and mixed variants for the same 10,000 steps and 3 seeds with the same task-tuned LIBERO-90 video predictors, reporting per-task and mean transfer success. If the 138k CF-only model stays near 84.5%, the counterfactual-content attribution clears the scale confound; if it falls toward the 21.5% demo-only level, or if the mixed variant's local-vs-full gap shrinks substantially, the headline gain is a volume effect.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 7 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 37 s · 26,826 tokens read, 12,259 written
  3. Editor wrote the verdict above 18 s · 27,867 tokens read, 2,191 written
  4. Referee major revision: 4 major objections, 5 minor 1 min 57 s · 27,086 tokens read, 17,226 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Short version: this is a systems-and-empirics paper, and a useful one. What is new is not the ingredients—Wan2.2, Mixture-of-Transformers, flow matching, chunk-causal prediction, and the four interaction modes all exist elsewhere—but the controlled comparison on a shared causal backbone, plus the measurement that a frozen local-context IDM can transfer after visual adaptation. The main LIBERO protocol is careful: three seeds, shared environment seeds, 50 episodes per task, standard deviations reported. The 68.4% to 97.8% gain is honestly labeled as the combined effect of robot-video pretraining and causal adaptation (Table 12), and the MoT ablation isolates architecture. The chain-rule factorization in Eq. 3 is an identity, and the paper does not pretend it is a theorem; the value is the evaluation recipe, the open code, and the reusable-component study.

The soft spot is the one the stress-test flags, and it lands. Table 5 compares a demo-only local IDM at 21.5% against CF-only at 84.5% and mixed at 84.0%, but Table 9 reports 138,090 demonstration controls versus 4,096,000 counterfactual controls—a 29.7x gap. Raw gradient steps are matched, but unique transition diversity per batch is exactly the variable counterfactual supervision is supposed to supply, so the headline attribution to counterfactual content rather than data scale is not isolated. A scale-matched demo-only leg is missing. The authors report the 29.7x figure transparently, but reporting a confound does not remove it. This is fixable and should be the main referee ask.

Secondary issues, in proportion: Table 5 has no error bars or seed counts, unlike the main tables; full-context CF-only is not reported, so the conditioning-scope comparison is incomplete; and the real-robot results have no baselines. The distributional-match premise—that adapted video-predictor futures fall inside the frozen IDM's training support—is not directly measured; Table 11 gives aggregate contact statistics rather than a coverage test. I would call that a gap rather than a fatal flaw, because the downstream success itself is weak evidence of adequate support, but a direct measure would strengthen the paper.

Who this is for: anyone comparing WAM interaction designs, or working on inverse and forward dynamics from counterfactual transitions. The framework and code release make it a durable testbed regardless of the open questions. Verdict: conditional. I would send it to serious peer review and ask for the scale-matched control, error bars on the transfer table, and ideally a coverage statistic. If those come back, the central claim is well supported.

The strongest objection

The central claim is causal: counterfactual transition content, not demonstration content, makes a frozen local-context IDM reusable after visual adaptation. Table 5 shows demo-only local at 21.5%, CF-only at 84.5%, and mixed at 84.0%. But Table 9 reports 138,090 demonstration controls versus 4,096,000 counterfactual controls, a 29.7x gap. Under the 10,000-step downstream protocol of §A.2, raw gradient steps are matched; what is not matched is the diversity of unique transitions per batch, which is exactly the variable the paper credits to counterfactual supervision. A local-context IDM could improve from 21.5% simply because 4M distinct transitions give broader coverage and regularization than resampling 138k demonstration controls, independent of whether the added transitions are counterfactual. The CF-only versus mixed comparison keeps scale roughly fixed, but there is no scale-matched demo-only leg, so the comparison that motivates the paper—counterfactual rather than demonstration supervision as the decisive ingredient—is not isolated. The authors report the 29.7x figure transparently, and the reader's rationale already flags the missing control, but the missing leg is precisely what would make the attribution causal rather than correlational.

The paper's most interesting contribution is the claim that counterfactual transitions, not merely larger or more diverse demonstration data, enable a frozen action-grounding component to transfer. That claim is load-bearing because if the effect is only scale, the result reduces to 'more data helps,' which is far less novel. The most direct threat is therefore the missing quantity-matched control, not the distribution-overlap premise highlighted in the reader's weakest_assumption: Table 5's end-to-end success rates are themselves a first-order test that the adapted predictor's futures are usable by the frozen IDM, whereas the scale confound is untouched by any reported experiment. The reader's rationale already names this confound, so the conditional verdict is correct and my read does not change it. I also credit the paper's transparency: Table 9 states the 29.7x gap, §C describes the intervention distribution, and the shared backbone, evaluation protocol, and seed conventions for native policies are careful. The proposed downsampling test is feasible with the released code and dataset and would settle whether the counterfactual-content attribution survives.

The referee's report

If the composability result holds, the paper establishes a practically valuable modularity property: an action-grounding component trained once on local transitions can be frozen and reused across supervised visual adaptation to new tasks in the same domain, with counterfactual-style transition coverage as the reported driver. The framework contribution is genuine: by fixing a common causal backbone, objective family, and evaluation protocol across four interaction programs, the paper isolates generation order and cross-modal visibility, which most prior WAM systems vary simultaneously, and the concurrent name-twin OpenWAM [79] is explicitly positioned. I credit the paper for several strengths: the 68.4-to-97.8 LIBERO-Long gain is honestly labeled as the combined effect of pretraining and causal adaptation rather than attributed to one factor; the MoT ablation (Table 13) isolates the architecture; Eq. (3) is presented as an interface factorization rather than an identity that imposes content; the main policy evaluation uses three seeds, shared environment seeds, and 50 episodes per task; the 29.7x data-volume gap is reported transparently in Sec.

Pith's reading of the paper (a machine's summary, not the authors')

The paper sets out to show that a robot's action-grounding component can be trained once, frozen, and reused across new tasks that only require adapting the visual predictor, provided the grounding model was trained on counterfactual transitions (alternative action branches executed from the same simulator state) rather than on successful demonstrations alone. Its headline numbers: a frozen local-context inverse dynamics model, which sees only the current observation, proprioception, and a supplied future video trajectory, reaches 84.0% mean closed-loop success across four held-out LIBERO-90 tasks when trained on a mix of counterfactual and demonstration data, versus 21.5% for the demonstration-only version and 47.0% for a full-context inverse model. The paper also claims that causal robot-video pretraining without action labels is the chief contributor to native policy performance, lifting video-then-action success on LIBERO-Long from 68.4% to 97.8%. If both claims hold, adapting a robot to a new task could stop at swapping in a task-adapted video model and feeding its imagined futures to a frozen action model, skipping action retraining entirely.

What carries the argument. The load-bearing object is the local-context dynamics interface, given by $p_{\mathrm{IDM}}^{\mathrm{local}}(A_t^+ \mid O_t^+, O_t^0, q_t)$ and $p_{\mathrm{FDM}}^{\mathrm{local}}(O_t^+ \mid A_t^+, O_t^0, q_t)$: the inverse model maps a supplied future visual trajectory back to the actions that would produce it, and the forward model maps candidate actions to their visual consequences, with neither receiving language, past actions, or history before the current step. Around this interface the paper builds a shared causal video backbone (a Wan2.2-5B model continued pretraining on roughly 14.6k hours of robot and human-interaction video without action labels, using chunk-causal attention so future chunks can see only the present and past), a Mixture-of-Transformers dual-stream architecture whose four interaction programs (joint, video-then-action, action-then-video, decoupled) differ only in generation order and future cross-modal attention, and the counterfactual branch dataset LIBERO-Long-CF, which expands local transition supervision by roughly 29.7× over demonstrations. The inference identity carrying the transfer claim is composition: the task-adapted video predictor proposes $\hat{O}_t^+$, and the frozen local IDM consumes that trajectory together with the current observation and proprioception to produce $\hat{A}_t^+$.

Core claim. The central discovery is that counterfactual supervision, not architecture, not demonstrations alone, and not task context, is the decisive ingredient for reusing a frozen inverse dynamics component after visual adaptation. In the transfer study of Sec. 4.3, the video predictor is task-tuned to each of four held-out LIBERO-90 targets while the action component stays frozen: a local-context IDM trained on counterfactual-only branches reaches 84.5% mean success, a 60/40 counterfactual/demonstration mixture reaches 84.0%, the demonstration-only local-context IDM reaches 21.5%, and a full-context IDM with the same mixed supervision reaches 47.0%. The authors attribute the gap to transition coverage: demonstrations contain only the single action continuation selected by the demonstrating policy, so an inverse model trained on them cannot ground the alternative futures a task-adapted video predictor proposes, whereas counterfactual branches executed from the same initial states populate exactly that local state-action-outcome space. The same pattern appears in forward dynamics, where counterfactual supervision cuts RGB prediction error by 34.5% and raises identification of the matching action among 16 same-state futures from 21.1% to 71.3%.

The premise it rests on. The transfer result rests on an untested distributional-match premise: the futures proposed by the task-tuned video predictors on the four LIBERO-90 targets must fall inside the support of the transitions the frozen inverse model learned from the LIBERO-Long counterfactual branches, so the 84.0% mean could collapse if that overlap breaks on other tasks.

If the paper is right (Pith's inferences, not the authors')

Adapting to a new manipulation task can be reduced to adapting only the video predictor when a counterfactual-trained local-context inverse dynamics model is available and frozen.

Counterfactual-only supervision matches mixed supervision for transfer (84.5% vs 84.0%), so successful demonstrations are not the essential ingredient in reusable action grounding.

Full-context conditioning degrades reuse (47.0% vs 84.0% under matched mixed supervision), so task-level language and history tie the action model to its source tasks.

Causal robot-video pretraining without action labels is worth 29.4 points of LIBERO-Long video-then-action success (68.4% to 97.8%), so unlabeled robot video carries most of the task-relevant visual structure.

Counterfactual supervision also makes forward dynamics action-discriminative, lifting 16-way same-state outcome identification from 21.1% to 71.3%.

Editorial extension: the same recipe, a frozen local dynamics module plus a per-task adapted perception module, could transfer to physical robots if what-if transitions from failures, exploration, or teleoperation can be collected cheaply, because the local-transition formulation needs no task success label.

Testable extension: deliberately push the adapted video predictor outside the counterfactual training support, for example with far object or grasp shifts, and measure the success drop; the distributional-overlap explanation predicts a sharp collapse where a task-specific explanation would predict none.

If the pretraining claim holds at scale, the cost of new-task adaptation drops to visual-only adaptation, which uses no action labels at all, a consequence with direct implications for robot data collection.

The 16-way outcome identification metric is a narrow probe of causal understanding; extending it to long-horizon rollouts and real-robot deployment would test whether counterfactual-trained forward models can support planning rather than only identification.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

World-action models (WAMs) couple future prediction with robot control, yet existing systems often vary the video backbone, interaction structure, supervision, and inference procedure simultaneously, making their design choices difficult to compare. We introduce OPENWAM, an open world-action modeling framework built around a common causal robot-video foundation and configurable video-action interaction. Starting from Wan2.2-5B, we perform causal robot-video pretraining on over 10,000 hours of video, then integrate an action expert through a shared Mixture-of-Transformers architecture that supports joint, video-then-action, action-then-video, and decoupled generation. OPENWAM achieves high success rates on four LIBERO suites and real-world bimanual tasks; robot-video training with causal adaptation improves VTA success on LIBERO-Long from 68.4% to 97.8%. The same configurable architecture naturally extends to inverse and forward dynamics, allowing us to study how counterfactual transitions improve independently trained dynamics models beyond demonstrations alone. When only the video predictor is adapted to a new task, a frozen local-context inverse dynamics model trained on counterfactual data and demonstrations achieves 84.0% mean success across four held-out LIBERO-90 tasks, compared with 47.0% for a full-context inverse model and 21.5% for a local-context model trained only on demonstrations. For forward dynamics, counterfactual supervision reduces RGB prediction error by 34.5% and raises outcome identification from 21.1% to 71.3% among 16 same-state outcomes. OPENWAM provides a common testbed for comparing WAM interaction designs and for studying dynamics learning from video data beyond successful demonstrations.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-07

The authors' main claim, as the classifier picked it out 0.82

“OPENWAM provides a common testbed for comparing WAM interaction designs and for studying dynamics learning from video data beyond successful demonstrations.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.13 · human participants 0.07 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.94 · mentions code or data availability 0.46

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
World-action models (WAMs) couple future prediction with robot control, yet existing systems often vary the video backbone, interaction structure, supervision, and inference procedure simultaneously, making their design choices difficult to compare. background 0.58 0.08 0.03 asserted 0.95
We introduce OPENWAM, an open world-action modeling framework built around a common causal robot-video foundation and configurable video-action interaction. contribution 1.00 0.88 0.12 asserted 0.99
Starting from Wan2.2-5B, we perform causal robot-video pretraining on over 10,000 hours of video, then integrate an action expert through a shared Mixture-of-Transformers architecture that supports joint, video-then-action, action-then-video, and decoupled generation. method 0.99 0.95 0.06 asserted 1.00
OPENWAM achieves high success rates on four LIBERO suites and real-world bimanual tasks; robot-video training with causal adaptation improves VTA success on LIBERO-Long from 68.4% to 97.8%. result 1.00 0.96 0.31 asserted 1.00
The same configurable architecture naturally extends to inverse and forward dynamics, allowing us to study how counterfactual transitions improve independently trained dynamics models beyond demonstrations alone. contribution 0.84 0.85 0.10 asserted 0.96
When only the video predictor is adapted to a new task, a frozen local-context inverse dynamics model trained on counterfactual data and demonstrations achieves 84.0% mean success across four held-out LIBERO-90 tasks, compared with 47.0% for a full-context inverse model and 21.5% for a local-context model trained only on demonstrations. result 1.00 0.95 0.13 asserted 1.00
For forward dynamics, counterfactual supervision reduces RGB prediction error by 34.5% and raises outcome identification from 21.1% to 71.3% among 16 same-state outcomes. result 1.00 0.94 0.08 asserted 1.00
OPENWAM provides a common testbed for comparing WAM interaction designs and for studying dynamics learning from video data beyond successful demonstrations. contribution 0.85 0.61 0.50 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-07: 1 request, 6,626 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. Overview of OPENWAM. Left: Starting from Wan2.2 [75], causal pretraining on robot and human-interaction video builds a predictive backbone without action labels, followed by video–action training on simulated and real-robot demonstrations. Top right: A shared Mixture-of-Transformers couples temporally aligned video and action streams through joint attention while retaining modality￾specific experts. Middle right: Configurable cross-modal attention and generation order support joint, video-then-action, action-then-video, and decoupled prediction within the same architecture. Bottom…
Figure 2 of the paper
Figure 2. Shared MoT architecture. Two modality-specific
Figure 3 of the paper
Figure 3. Attention masks the shared architecture exposes: four
Figure 4 of the paper
Figure 4. Evaluation tasks. Left: the three real-world bimanual tasks of
Figure 5 of the paper
Figure 5. Real-world bimanual setup with two Franka FR3 arms,

Figures are the authors', reproduced from the paper.

References (109)
  1. [79] Yuran Wang, Siqiao Huang, Mingleyang Li, Chenhao Zhang, Jiaqi Liang, Weiyang Jin, Yue Chen, Xuemin Chi, Donghao Zhou, Qize Yu, et al. Openwam: An open, modular explo- ration towards systematic world-action model pretraining. arXiv preprint arXiv:2609.07398, 2026. 3 on PithPith's reviewarXiv
  2. [1] Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Ade- line Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yo- gesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai. arXiv preprint arXiv:2606.02800, on PithPith's reviewarXiv
  3. [2] Ezra Ameperosa, Jeremy A Collins, Mrinal Jain, and Ani- mesh Garg. Rocoda: Counterfactual data augmentation for data-efficient robot learning from demonstrations. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 13250–13256. IEEE, 2025. 3 on Pith
  4. [3] Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chen- dong Xiang, Yinze Rong, et al. Motus: A unified latent action world model. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 3... on Pith
  5. [4] Hongzhe Bi, Zihao Zhou, Yihang Tang, Jingrui Pang, Shuhe Huang, Haitian Liu, Runqing Wang, Shuai Huang, Yichen Wang, Yiming Cheng, et al. Motus2: A self-evolving general world model for dexterous manipulation. arXiv preprint arXiv:2608.30237, 2026. 3 on PithPith's reviewarXiv
  6. [5] Johan Bjorck, Fernando Casta ˜neda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025. 8 on PithPith's reviewarXiv
  7. [6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A vision-language- action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2, 8 on PithPith's reviewarXiv
  8. [7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022 on PithPith's reviewarXiv
  9. [8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: 9 Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2 on PithPith's reviewarXiv
  10. [9] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, et al. Agibot world colosseo: A large-scale ma- nipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 7 on PithPith's reviewarXiv
  11. [10] Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, et al. Worldvla: Towards autoregressive action world model. arXiv preprint arXiv:2506.21539, 2025. 3 on PithPith's reviewarXiv
  12. [11] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 3 on PithPith's reviewarXiv
  13. [12] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Ben- jamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shu- ran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. 7 on PithPith's reviewarXiv
  14. [13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 44(10-11):1684–1704, 2025. 2 on Pith
  15. [14] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video genera- tion. Advances in neural information processing systems , 36:9156–9172, 2023. 3 on Pith
  16. [15] Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter Fox, Ajay Mandlekar, and Yijie Guo. Aha: A vision- language-model for detecting and reasoning over failures in robotic manipulation. arXiv preprint arXiv:2410.00371, on PithPith's reviewarXiv
  17. [16] Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model- based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. 3 on PithPith's reviewarXiv
  18. [17] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In 2017 IEEE international conference on robotics and automation (ICRA), pages 2786– on Pith
  19. [18] Javier Alejandro Lopetegui Gonzalez, Paul Pacaud, and Cordelia Schmid. Spatially aware world action model via geometric latent diffusion. arXiv preprint arXiv:2609.02531, on PithPith's reviewarXiv
  20. [19] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings... on Pith
show 80 more
  1. [20] Jun Guo, Qiwei Li, Peiyan Li, Zilong Chen, Nan Sun, Yifei Su, Heyun Wang, Yuan Zhang, Xinghang Li, and Huaping Liu. Unified 4d world action modeling from video priors with asynchronous denoising. arXiv preprint arXiv:2604.26694, 2026. 3 on PithPith's reviewarXiv
  2. [21] Yanjiang Guo, Yucheng Hu, Jianke Zhang, Yen-Jen Wang, Xiaoyu Chen, Chaochao Lu, and Jianyu Chen. Prediction with action: Visual policy learning via joint denoising pro- cess. Advances in Neural Information Processing Systems, 37:112386–112410, 2024. 3 on Pith
  3. [22] Yanjiang Guo, Lucy Shi, Jianyu Chen, and Chelsea Finn. Ctrl-world: A controllable generative world model for robot manipulation. In International Conference on Learning Representations, pages 6121–6138, 2026. 3 on Pith
  4. [23] Huy Ha, Yihuai Gao, Zipeng Fu, Jie Tan, and Shuran Song. UMI on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers. In Proceed- ings of the 2024 Conference on Robot Learning, 2024. 3 on Pith
  5. [24] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Moham- mad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. 3 on PithPith's reviewarXiv
  6. [25] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3 on PithPith's reviewarXiv
  7. [26] Yucheng Hu, Yanjiang Guo, Pengchao Wang, Xiaoyu Chen, Yen-Jen Wang, Jianke Zhang, Koushil Sreenath, Chaochao Lu, and Jianyu Chen. Video prediction policy: A generalist robot policy with predictive visual representations. arXiv preprint arXiv:2412.14803, 2024. 3 on PithPith's reviewarXiv
  8. [27] Zheyuan Hu, Robyn Wu, Naveen Enock, Jasmine Li, Riya Kadakia, Zackory Erickson, and Aviral Kumar. Rac: Robot learning for long-horizon tasks by scaling recovery and cor- rection. IEEE Transactions on Robotics, 2026. 3 on Pith
  9. [28] Wen Huang, Haoran Sun, Yongjian Guo, Yunxuan Ma, Hao- ran Li, Jing Long, Zhouying Mo, Zhong Guan, Yucheng Guo, Shuai Di, et al. Noisegate: Learning per-latent timestep schedules as information gating in world action models. arXiv preprint arXiv:2605.07794, 2026. 3 on PithPith's reviewarXiv
  10. [29] Xuyao Huang, Yixuan Wang, Zengyao Ye, Boyuan Zhao, Chenyang Yu, Haoran Wen, and Zhijie Deng. Streaming-wam: Action-conditioned world-action model for asynchronous robot manipulation. arXiv preprint arXiv:2609.28927, 2026. 3 on PithPith's reviewarXiv
  11. [30] Yixuan Huang, Novella Alvina, Mohanraj Devendran Shan- thi, and Tucker Hermans. Fail2progress: Learning from real-world robot failures with stein variational inference. arXiv preprint arXiv:2509.01746, 2025. 3 on PithPith's reviewarXiv
  12. [31] Ze Huang, Jiahui Zhang, Hairuo Liu, Chenxi Zhang, Ran Cheng, and Li Zhang. Learning transferable dynamics priors from action to world modeling. arXiv preprint arXiv:2606.29501, 2026. 3 on PithPith's reviewarXiv
  13. [32] Physical Intelligence, Ali Amin, Raichelle Aniceto, Ash- win Balakrishna, Kevin Black, Ken Conley, Grace Con- nors, James Darpinian, Karan Dhabalia, Jared DiCarlo, et al. π∗ 0.6: a vla that learns from experience. arXiv preprint arXiv:2511.14759, 2025. 3 on PithPith's reviewarXiv
  14. [33] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, 10 Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π0.5: A vision-language-action model with open-world generaliza- tion. arXiv preprint arXiv:2504.16054, 2025. 2, 8 on PithPith's reviewarXiv
  15. [34] Joel Jang, Seonghyeon Ye, Zongyu Lin, Jiannan Xiang, Jo- han Bjorck, Yu Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Yen-Chen Lin, et al. Dreamgen: Unlocking gen- eralization in robot learning through video world models. arXiv preprint arXiv:2505.12705, 2025. 3 on PithPith's reviewarXiv
  16. [35] Haoyi Jiang, Liu Liu, Xinjiang Wang, Zhihao Sun, Zequn Chen, Sen Wang, Xinjie Wang, Xia Chen, Jingfeng Yao, Weiheng Zhao, et al. Rethinking representations for world- action modeling. arXiv preprint arXiv:2609.38163, 2026. 3 on PithPith's reviewarXiv
  17. [36] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 8 on PithPith's reviewarXiv
  18. [37] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and suc- cess. arXiv preprint arXiv:2502.19645, 2025. 8 on PithPith's reviewarXiv
  19. [38] Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming- Yu Liu, Chelsea Finn, et al. Cosmos policy: Fine-tuning video models for visuomotor control and planning. arXiv preprint arXiv:2601.16163, 2026. 3 on PithPith's reviewarXiv
  20. [39] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In International Conference on Learning Representations, pages 40938–40958, 2024. 3 on Pith
  21. [40] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on robot learning, pages 143–156. PMLR, 2017. 3 on Pith
  22. [41] Baoyu Li, Xinchen Yin, Mengying Lin, Yixin Zhang, and Danfei Xu. Egowam: World action models beyond pixels with in-the-wild egocentric human data. In Robot World Models, 2026. 3 on Pith
  23. [42] Dayou Li, Jiuzhou Lei, Hao Wang, Lulin Liu, Yunhao Yang, Zihan Wang, Bangya Liu, Minghui Zheng, and Zhiwen Fan. Learning actionable manipulation recovery via counterfac- tual failure synthesis. arXiv preprint arXiv:2603.13528 , on PitharXiv
  24. [43] Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control. arXiv preprint arXiv:2601.21998, 2026. 1, 3, 8 on PithPith's reviewarXiv
  25. [44] Shuang Li, Yihuai Gao, Dorsa Sadigh, and Shuran Song. Unified video action model, 2025. arXiv preprint arXiv:2503.00200. 3, 5 on PithPith's reviewarXiv
  26. [45] Shalfun Li, Victor Yao, Charles Yang, Truth Qu, Regis Cheng, Ryan Yu, Howard Lu, Newton V on, Vincent Chen, Yohann Tang, et al. Wall-wm: Carving world action mod- eling at the event joints. arXiv preprint arXiv:2606.01955, on PithPith's reviewarXiv
  27. [46] Sizhe Lester Li, Evan Kim, Xingjian Bai, Tong Zhao, Tao Pang, Max Simchowitz, and Vincent Sitzmann. Turning video models into generalist robot policies. 2026. 3 on Pith
  28. [47] Yaxuan Li, Yichen Zhu, Junjie Wen, Chaomin Shen, and Yi Xu. Worldeval: World model as real-world robot policies evaluator. arXiv preprint arXiv:2505.19017, 2025. 3 on PithPith's reviewarXiv
  29. [48] Ziang Li, Dongzhou Cheng, Yibin Wang, Shiyue Wang, Xiaoyang Xu, Lingxuan Weng, Juan Wang, and Jiaqi Wang. Light-wam: Efficient world action models with state-fusion action decoding. arXiv preprint arXiv:2606.08242, 2026. 3 on PithPith's reviewarXiv
  30. [49] Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv preprint arXiv:2411.04996, 2024. 1, 4 on PithPith's reviewarXiv
  31. [50] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 5 on PithPith's reviewarXiv
  32. [51] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36:44776–44791, on Pith
  33. [52] Renhang Liu, Wenzhi Zhao, Zhuo Yang, Liliang Chen, Pengfei Zhou, Shengcong Chen, Guanghui Ren, Youlun Peng, Rongjun Jin, Nan Wang, et al. Ge-act 2.0: Pretraining and scaling a world-action model for robotic manipulation. arXiv preprint arXiv:2609.05588, 2026. 3 on PithPith's reviewarXiv
  34. [53] Yushan Liu, Peibo Sun, Shoujie Li, Yifan Xie, Lingfeng Zhang, Xintao Chao, Shiyuan Dong, Fang Chen, Xiao-Ping Zhang, and Wenbo Ding. Oa-wam: Object-addressable world action model for robust robot manipulation. arXiv preprint arXiv:2605.06481, 2026. 3 on PithPith's reviewarXiv
  35. [54] Zuxing Lu, Hongjia Zhai, Guanzhi Wang, Huajian Zeng, Jiaqi Yang, Jingyu Liu, Lei Cheng, Yuantai Zhang, Yuheng Qiu, Zezhou Cheng, et al. World-action models for robot learning and control: A survey. arXiv preprint arXiv:2609.16074, 2026. 3 on PithPith's reviewarXiv
  36. [55] Teli Ma, Jia Zheng, Zifan Wang, Chunli Jiang, Andy Cui, Junwei Liang, and Shuo Yang. Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control. arXiv preprint arXiv:2603.10448, 2026. 3 on PitharXiv
  37. [56] Zipei Ma, Xiaofei Wei, Junzhe Jiang, Shunlin Lu, and Li Zhang. Tacpac: Tactile prediction and real-time action cor- rection in world-action models for contact-rich manipulation. arXiv preprint arXiv:2609.05266, 2026. 3 on PithPith's reviewarXiv
  38. [57] Ajay Mandlekar, Danfei Xu, Roberto Mart´ın-Mart´ın, Yuke Zhu, Li Fei-Fei, and Silvio Savarese. Human-in-the-loop imitation learning using remote teleoperation. arXiv preprint arXiv:2012.06733, 2020. 3 on PithPith's reviewarXiv
  39. [58] Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Ab- hishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Poo- ley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE Inte... on Pith
  40. [59] Bikang Pan, Fan Liu, Haotao Lu, Jingya Wang, and Ye Shi. Selfwam: A self-grounded unified world action model for fast robot control. arXiv preprint arXiv:2608.00725, 2026. 3 11 on PithPith's reviewarXiv
  41. [60] Quanquan Peng, Yutong Liang, Rui Yan, Nicklas Hansen, and Xiaolong Wang. Fact: Failure-aware causal training for world-action models. arXiv preprint arXiv:2608.10232, on PithPith's reviewarXiv
  42. [61] Han Qi, Haocheng Yin, Aris Zhu, Yilun Du, and Heng Yang. Inference-time enhancement of generative robot policies via predictive world modeling. IEEE Robotics and Automation Letters, 2026. 3 on Pith
  43. [62] Chenhao Qiu, Ruixiang Wang, Runyi Zhao, Sixu Lin, Songen Gu, Shufeng Nan, Guiliang Liu, Kui Jia, Yanwei Fu, and Simo Wu. Vid2wam: Distilling video diffusion priors into world action models. arXiv preprint arXiv:2608.08558, on PithPith's reviewarXiv
  44. [63] Lu Qiu, Yizhuo Li, Yi Chen, Yuying Ge, Yixiao Ge, and Xihui Liu. Making foresight actionable: Repurposing repre- sentation alignment in world action models. arXiv preprint arXiv:2606.12217, 2026. 3 on PithPith's reviewarXiv
  45. [64] Omar Rayyan, John Abanes, Mahmoud Hafez, Anthony Tzes, and Fares Abu-Dakka. Mv-umi: A scalable multi- view interface for cross-embodiment learning. IEEE Access, on Pith
  46. [65] St´ephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the four- teenth international conference on artificial intelligence and statistics, pages 627–635. JMLR Workshop and ... on Pith
  47. [66] Yichao Shen, Fangyun Wei, Zhiying Du, Yaobo Liang, Yan Lu, Jiaolong Yang, Nanning Zheng, and Baining Guo. Videovla: Video generators can be generalizable robot ma- nipulators. Advances in neural information processing sys- tems, 38:95597–95621, 2026. 3 on Pith
  48. [67] Zhenhao Shen, Jiaqi Liang, Jasper Lu, Feng Jiang, Yuran Wang, Chuanbo Wei, Jiayi Liu, Jianchun Yang, Qize Yu, Jiadi You, et al. Ld4wam: Learning latent dynamics from human videos for world action models. arXiv preprint arXiv:2608.22403, 2026. 3 on PithPith's reviewarXiv
  49. [68] Haofeng Sun, Jiangbo Pei, Fei Kang, Zexiang Liu, Yaokun Li, Boyi Jiang, Hua Xue, Cindy Zhou, Wei Li, Yichen Wei, et al. Riemann-1.0: An embodied world action model for physical ai. arXiv preprint arXiv:2608.27033, 2026. 3 on PithPith's reviewarXiv
  50. [69] Qu Tang, Benhui Zhuang, Bo Yuan, Xue Yu, Longteng Guo, and Junlan Feng. World tokens: Enhancing embodied policies with training-time world modeling. arXiv preprint arXiv:2608.09730, 2026. 3 on PithPith's reviewarXiv
  51. [70] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, on PithPith's reviewarXiv
  52. [71] Rhoda AI Team. Causal video models are data-efficient robot policy learners. Rhoda AI Blog, 2026. 1, 3 on Pith
  53. [72] Yang Tian, Sizhe Yang, Jia Zeng, Ping Wang, Dahua Lin, Hao Dong, and Jiangmiao Pang. Predictive inverse dynam- ics models are scalable learners for robotic manipulation. In International Conference on Learning Representations, pages 92033–92052, 2025. 3 on Pith
  54. [73] Yang Tian, Yuyin Yang, Yiman Xie, Zetao Cai, Xu Shi, Ning Gao, Hangxu Liu, Xuekun Jiang, Zherui Qiu, Feng Yuan, et al. Interndata-a1: Pioneering high-fidelity synthetic data for pre-training generalist policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pa... on Pith
  55. [74] An Dinh Vuong, Tuan Van V o, Abdullah Sohail, Haoran Ding, Liang Ma, Xiaodan Liang, Anqing Duan, Ivan Laptev, and Ian Reid. World2act: Latent action post-training from world model dynamics. arXiv preprint arXiv:2603.10422, on PithPith's reviewarXiv
  56. [75] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 2, 4, 7 on PithPith's reviewarXiv
  57. [76] Junke Wang, Qihang Zhang, Shuai Yang, Yiming Luo, Yu- jun Shen, Zuxuan Wu, Yu-Gang Jiang, and Yinghao Xu. Repwam: World action modeling with representation visual- action tokenizers. arXiv preprint arXiv:2606.13674, 2026. 3 on PithPith's reviewarXiv
  58. [77] Linhan Wang, Zijian An, Mingyuan Zhang, Chen Dai, Yi Xu, Can Cui, Zichong Yang, Yinlin Chen, Lifeng Zhou, and Chang-Tien Lu. Glancewam: Sparse test-time imagination for world-action models. arXiv preprint arXiv:2608.23927, on PithPith's reviewarXiv
  59. [78] Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, et al. World action models: The next frontier in embodied ai. arXiv preprint arXiv:2605.12090, on PithPith's reviewarXiv
  60. [80] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations, pages 10641–10662, 2024. 3 on Pith
  61. [81] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xi- aozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi- embodiment intelligence normative data for robot manipula- tion. arXiv preprint arXiv:2412.13877, 2024. 7 on PithPith's reviewarXiv
  62. [82] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226–2240. PMLR, 2023. 3 on Pith
  63. [83] Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, and Pieter Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163. IEEE, 2024. 2 on Pith
  64. [84] Shihan Wu, Xuecheng Liu, Shaoxuan Xie, Pengwei Wang, Xinghang Li, Bowen Yang, Zhe Li, Kai Zhu, Hongyu Wu, 12 Yiheng Liu, et al. Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation. arXiv preprint arXiv:2511.17441, 2025. 7 on PithPith's reviewarXiv
  65. [85] Xionghao Wu, Yijun Yang, Shiyang Zhou, Haoze Sun, Jian- hui Liu, Songsong Yu, Jiyao Zhang, Wenbo Li, Bo Wang, Guoqing Ma, et al. Zimablue: Evolving generalizable world action models through scalable video pre-training. arXiv preprint arXiv:2609.00188, 2026. 3 on PithPith's reviewarXiv
  66. [86] Mengda Xu, Han Zhang, Yifan Hou, Zhenjia Xu, Linxi Fan, Manuela Veloso, and Shuran Song. Dexumi: Using human hand as the universal manipulation interface for dexterous manipulation. In Conference on Robot Learning , pages 437–459. PMLR, 2025. 3 on Pith
  67. [87] Liudi Yang, Yang Bai, George Eskandar, Fengyi Shen, Mo- hammad Altillawi, Dong Chen, Ziyuan Liu, and Abhinav Valada. Covar: Co-generation of video and action for robotic manipulation via multi-modal diffusion. In 2026 IEEE In- ternational Conference on Robotics and Automation ... on Pith
  68. [88] Liuhaichen Yang, Zhuang Jiang, Chenchao Sheng, and Zezhi Tang. Wam-opd: On-policy distillation for world action models. arXiv preprint arXiv:2608.22364, 2026. 3 on PithPith's reviewarXiv
  69. [89] Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigaworld-policy: An efficient action- centered world–action model, 2026. URL https://arxiv. org/abs/2603.17240. 3 on PitharXiv
  70. [90] Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies. arXiv preprint arXiv:2602.15922, 2026. 1, 3 on PithPith's reviewarXiv
  71. [91] Tenny Yin, Zhiting Mei, Zhonghe Zheng, Miyu Yamane, David Wang, Jade Sceats, Samuel M Bateman, Lihan Zha, Apurva Badithela, Ola Shorinwa, et al. Playworld: Learning robot world models from autonomous play. arXiv preprint arXiv:2603.09030, 2026. 3 on PithPith's reviewarXiv
  72. [92] Jiadi You, Qize Yu, Yue Chen, Minghong Cai, Zhide Zhong, Yuran Wang, Bowen Ping, Jiaqi Liang, Zhenhao Shen, Haodong Yan, et al. Affordancewam: Affordance-aware joint world-action modeling for robot manipulation. arXiv preprint arXiv:2609.22332, 2026. 3 on PithPith's reviewarXiv
  73. [93] Yang Yu. On the capability separation between world-model policy learning and imitated world-action models. arXiv preprint arXiv:2608.22197, 2026. 3 on PithPith's reviewarXiv
  74. [94] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026. 3, 8 on PithPith's reviewarXiv
  75. [95] Chuhan Zhang, Seiji Ito, Kenta Hoshino, Satoshi Ikehata, and Ikuro Sato. World-coherent decoding: Self-verifying test-time planning for world action models. arXiv preprint arXiv:2609.02159, 2026. 3 on PithPith's reviewarXiv
  76. [96] Mengqi Zhang, Sahil Khose, Simar Kareer, Yuchen Song, Unnat Jain, and Judy Hoffman. Deva: Decoupled video- action model with physical guidance for robot policy learn- ing. arXiv preprint arXiv:2607.24159, 2026. 3 on PithPith's reviewarXiv
  77. [97] Qihang Zhang, Lin Li, Luyao Zhang, Shuai Yang, Yiming Luo, Shuaiting Li, Ruilin Wang, Junke Wang, Jiahao Shao, Gangwei Xu, et al. Native video-action pretraining for gen- eralizable robot control. arXiv preprint arXiv:2607.08639, on PithPith's reviewarXiv
  78. [98] Ganlong Zhao, Zijia Tang, Xingping Chen, Zhanghui Kuang, Ye Tian, and Guanbin Li. Flare: A failure-aware framework for autonomous correction and recovery in visual-language robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti... on Pith
  79. [99] Ruiteng Zhao, Zhengshen Zhang, Yue Su, Wenshuo Wang, Jiahui Li, Zhiyuan Yang, Francis EH Tay, Marcelo H Ang Jr, and Haiyue Zhu. Sg-wam: Self-guided world modeling in geometry-aware policy space. arXiv preprint arXiv:2608.01397, 2026. 3 on PithPith's reviewarXiv
  80. [100] Sizhe Zhao, Haozhe Xie, Weiyu Zhao, Chenchu Zhang, Huan Wang, Chenyang Wang, Qinglin Liu, and Sheng- ping Zhang. Memory as plans: World-action model- ing with memory-grounded planning. arXiv preprint arXiv:2609.11561, 2026. 3 on PithPith's reviewarXiv

The first 100 of 109 references are listed here. All of them are in this paper's citation record.

Cited by 3 Pith papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment