Pith. sign in

REVIEW 4 major objections 4 minor 62 references

This paper claims that a frozen flow-matching world-action model can be re-grounded at individual time steps inside its active chunk by imposing newly observed latent states and committed actions as masked pseudoinverse constraints on the v

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

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

T0 review · deepseek-v4-flash

2026-08-03 11:02 UTC pith:DF55J7BR

load-bearing objection A clean, well-formulated inference-time feedback mechanism for world-action models whose headline task-success gains are not yet attributable to the new state-feedback path, because the action-only ablation is never reported. the 4 major comments →

arxiv 2607.29235 v1 pith:DF55J7BR submitted 2026-07-31 cs.RO cs.AIcs.SYeess.SY

FBFM: A Training-Free Asynchronous Feedback Mechanism for Flow-Matching in World-Action Models Execution

Peize Li , Ruimeng Zhang , Ru Zhang , Cong Huang , Kai Chen , Shanghang Zhang This is my paper
classification cs.RO cs.AIcs.SYeess.SY
keywords world-action modelsflow matchingtraining-free guidancepseudoinverse guidanceclosed-loop robot controllatent state feedbackasynchronous executionmanipulation benchmarks
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 show that a world-action model—a system that jointly predicts future images and actions—can be corrected during generation, not just between chunks. The authors propose FBFM, a training-free inference mechanism that treats the flow-matching solver as a dynamically partially observed generation problem: newly encoded real observations and actions committed by the previous chunk become time-aligned constraints on the active chunk. A masked pseudoinverse guidance term steers the velocity field toward those measurements while the pretrained model stays frozen. On a stage-wise model the correction raises equal-weight task-configuration success by about three percentage points; on a joint-generation model the pooled gain is modest and suite-dependent. The paper also records an unresolved limitation: when state feedback covers only a finite prefix of a generated video, large visual artifacts appear after the measured prefix ends, which it attributes tentatively to the aligned-coordinate approximation or a codec/backbone mismatch.

Core claim

The central claim is that re-grounding can be pushed inside the actively generated chunk. While a world-action model is still solving its flow, the newly encoded latent state z_{t+i} and the preceding committed action a^{prev}_{t+i} can be imposed as masked targets on the corresponding temporal slots of the chunk. At every solver evaluation, the clean-endpoint estimate is compared with the lifted target on the masked subspace, and the discrepancy is propagated through the endpoint Jacobian as a vector–Jacobian product that corrects the velocity field. No parameters are updated. The paper argues that this time-aligned masked pseudoinverse correction is a common interface for both stage-wise W

What carries the argument

The central object is the masked pseudoinverse guidance update (Eq. 19): v^Q_FBFM = v̄^Q + λ (J^Q)^T P^Q W^Q (Y^Q − Q̂^1), where Q̂^1 is a clean-endpoint estimate from the current noisy sample, J^Q is its Jacobian, W^Q is the time-aligned support or confidence mask, P^Q is a modality scale preconditioner, Y^Q is the lifted target (committed previous action or encoded real latent state), and λ is a flow-time-dependent guidance strength. The mask converts chunked inference into a dynamically partially observed generation problem; the transpose Jacobian propagates residuals across state and action coordinates in joint models; the guidance schedule controls how strongly the frozen prior is pulle

Load-bearing premise

The load-bearing premise is the aligned-coordinate approximation h†(h(ˆX^1)) ≈ ˆX^1 on the masked subspace: if the frozen encoder's latent coordinates do not match the model's predicted clean-latent coordinates, injected measurements can push generation off the data manifold and produce the large visual artifacts the paper observes after a finite measured prefix ends.

What would settle it

On the auxiliary ball-collision sequence, compare the onset of large-area image corruption under finite 10-, 20-, and 30-slot feedback coverage. If the corruption onset does not track the end of the measured prefix, or if the same artifact appears when the exact decoded-and-reencoded residual h†(h(ˆX^1)) is substituted for the aligned-coordinate residual, then the paper's mechanism-level explanation—codec/backbone mismatch as distinct from the aligned-coordinate approximation—would be falsified.

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

If this is right

  • FBFM applies to any frozen flow-matching WAM without retraining, preserving the native chunking, solver schedule, KV cache, and classifier-free guidance settings.
  • On LingBot-VA over 42 RoboTwin2.0 tasks, the equal-weight task-configuration success rate rises from 80.1% to 83.1% overall, with +2.86 points in the clean setting and +3.10 points in the randomized setting.
  • On DreamZero over 800 LIBERO episodes, the pooled success rate rises from 70.125% to 70.75%, with a 2.5-point gain on LIBERO-Goal and LIBERO-10 and small decreases on LIBERO-Spatial and LIBERO-Object.
  • On real-world robot observation prediction, causally injecting 30 measured latent slots lowers MAE from 9.63 to 9.27 and raises PSNR from 20.06 to 23.10 dB, improving visual tracking of the task-relevant ball position.
  • In joint-generation WAMs, the cross-modal Jacobian block J^T_ZA lets a state residual directly correct action coordinates in the same solver step, without waiting for a separate context handoff.

Where Pith is reading between the lines

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

  • The same masked pseudoinverse interface could be lifted out of the WAM setting and applied to other latent-conditional flow or diffusion generators whenever real measurements arrive during sampling, turning this paper's recipe into a general online-guidance primitive.
  • The paper's aligned-coordinate approximation suggests a concrete testable refinement: decoding predicted endpoints and re-encoding them before forming the residual might reduce the post-prefix artifacts that the paper observes in Appendix F, at the cost of extra compute.
  • For joint models, the gain search shows that reciprocal scaling inside one solver linearization does not guarantee closed-loop stability after actions are executed; a natural next step, which the paper explicitly leaves open, is to close the loop with a learned or actual environment-transition model and use PID-style feedback instead of a fixed proportional gain.
  • Because FBFM only requires feedback to be visible before a solver evaluation, it could be combined with a learned predictor that anticipates future observations, allowing corrections to be applied before the physical measurement arrives.

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 / 4 minor

Summary. The paper proposes Feedback Flow Matching (FBFM), a training-free inference-time mechanism that applies masked pseudoinverse guidance to the conditional velocity field of a frozen Flow-Matching world-action model (WAM). FBFM injects two kinds of time-aligned constraints into the active generation chunk: dynamically observed latent states (state feedback) and previously committed actions from the preceding chunk (action-prefix consistency). The method is instantiated on a stage-wise WAM (LingBot-VA on RoboTwin2.0) and a joint-generation WAM (DreamZero on LIBERO), reporting equal-weight task-success gains of +2.98 percentage points on RoboTwin2.0 and +0.625 percentage points pooled on LIBERO, plus a real-world video-prediction diagnostic showing improved state tracking. The central claim is that state feedback, injected through Eq. (19), improves task success and visual re-grounding without modifying or retraining the WAM.

Significance. If established, FBFM would be a useful, simple, and architecture-agnostic interface for closed-loop correction in flow-matching WAMs, complementing action-only chunking methods like RTC. The paper is unusually transparent about implementation details: it provides tensor layouts, pseudo-clock schedules, commit hashes, a full parameter-search ledger, and an explicit statement and bound for the aligned-coordinate approximation in Appendix B.3. The real-world observation-prediction diagnostic is also a valuable step beyond pure simulation. However, the manuscript's headline empirical claims are not yet conclusive: the task-success tables never isolate the state-feedback component from the fixed action-prefix constraint, and the DreamZero gains are obtained with hyperparameters selected on the same benchmark tasks used in the final evaluation.

major comments (4)
  1. [§C.1, Tables 1 and 2] The central attribution of task-success gains to state feedback is not supported, because the task-success tables report only Base vs. FBFM. Appendix C.1 defines three mask settings—NONE, RTC (action mask only), and FBFM (action + state mask)—and states that they share checkpoint, noise, solver budget, and pseudo-clock. Since FBFM always includes the RTC-style fixed previous-action prefix, the reported +2.98 pp (RoboTwin) and +0.625 pp (LIBERO pooled) could in principle be produced entirely by the action-prefix constraint alone, which is not the paper's novel contribution. The §5.2 mechanism diagnostic shows only that state feedback changes computed action velocities, not that it affects task success. Please add task-success results for the RTC-only mode on the same tasks and configurations; this is a necessary control for the paper's central claim.
  2. [Appendix E, Table 2] The DreamZero configuration is not independently evaluated. The state preconditioner PZ and proportional gain kp were selected by screening on LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10 subsets (E.2–E.4), and the same suites appear in the final pooled estimate in Table 2. The selected gain kp=0.0486968 is the best point estimate on the screened tasks ('neither paired comparison establishes statistical superiority'), and the final table therefore reflects a configuration chosen on the evaluation data. Please provide a held-out evaluation on tasks or suites not used in tuning, or report a cross-validated estimate with appropriate uncertainty. At minimum, quantify how much of the Table 2 gain is specific to the screening tasks.
  3. [§C.2, Table 1] The RoboTwin timing schedule undermines the interpretation of the LingBot-VA gains as evidence for fine-grained state feedback. The first complete dynamic state target is formed only after action 16 of the 16-action suffix, when only 2 of the 25 numerical video updates remain. Thus state feedback can correct at most the final numerical video update; the remainder of the state chunk is generated open-loop. The task-success gain could stem from the fixed action-prefix constraint, from the corrected-context refresh of the action stage, or from a combination—but not from direct visual correction of the active chunk, as the paper's phrasing suggests. The RTC-only control requested above would disentangle this, but the timing itself should also be discussed explicitly in the interpretation of Table 1.
  4. [§D.1, Abstract] The RoboTwin2.0 evaluation uses a 'selected' set of 42 tasks, with longer tasks excluded. The abstract and introduction motivate FBFM by 'long-horizon reliability' and 'drift in long-horizon tasks,' but the main success-rate experiment omits the longer tasks where drift would be most severe. Please justify the selection criterion or include at least a sample of longer tasks; as presented, the evidence for the long-horizon motivation is indirect.
minor comments (4)
  1. [Table 5] The 'State preconditioner' row renders as '1P Z = 56/9600', which is ambiguous. Please separate the LingBot-VA value (presumably 1) and the DreamZero value (PZ = 56/9600) into distinct columns or cells.
  2. [§5.2/Figure 5] The mechanism diagnostic reports MSE and RMS changes on four paired task–trial units. No error bars or test statistics are given; given the small sample, please state whether the reported difference (e.g., 0.6828 vs. 0.6751) is within run-to-run variability.
  3. [§3.2] The sentence 'This formulation follows the Flow-Matching inpainting construction of Black et al. [47]' conflates pseudoinverse guidance (Song et al. [9]) with the RTC action-chunking paper. Please cite both appropriately and distinguish the general inverse-problem guidance from the action-prefix specialization.
  4. [§5.4/Appendix F] The real-world diagnostic uses all 30 measured latent slots (full-oracle coverage), while the paper's own Appendix F shows that finite-prefix coverage produces large artifacts. Please clarify the deployment relevance of the full-coverage condition and discuss how the finite-coverage failure mode would be handled in actual closed-loop execution.

Circularity Check

1 steps flagged

The FBFM guidance equation is a disclosed reuse of existing pseudoinverse-flow methods, so the mechanism itself is not circular; the main circularity is that the DreamZero state-gain and preconditioner were selected on tasks that later appear in the reported aggregate, making that headline improvement partly a fitted result rather than an independent prediction.

specific steps
  1. fitted input called prediction [Appendix E.2–E.4; §5.3 Table 2]
    "We first compared three values of PZ ... The controlled sweep used the ten LIBERO-Spatial and ten LIBERO-Object tasks ... We therefore fix PZ = 56/9600 for the subsequent gain search. ... Each new screening point used LIBERO-Spatial tasks 1 and 9, LIBERO-Object tasks 0 and 6 ... It is therefore selected as an operational candidate rather than a universal optimum."

    The DreamZero result in Table 2 is not an evaluation of a fixed FBFM: the PZ preconditioner and kp gain were chosen by running the FBFM-vs-Base comparison on LIBERO-Spatial and LIBERO-Object, and then on Spatial 1/9, Object 0/6, expanded to Goal 0/6 and LIBERO-10 0/6, with the selected kp winning on those same task IDs. Reporting aggregate success over those suites as FBFM's improvement therefore partly reports the selection criterion. The paper discloses the tuning and calls the choice an 'operational candidate', so this is partial fitted-input-called-prediction circularity, not a hidden derivation.

full rationale

FBFM's core update (Eq. 19) is explicitly an adaptation of published pseudoinverse-guided flow matching and real-time chunking: the paper says 'This formulation follows the Flow-Matching inpainting construction of Black et al. [47]' and cites ΠGDM [9]. It therefore reuses externally validated machinery rather than defining the target in terms of the output. The aligned-coordinate approximation h†(h(X))≈X is flagged in Appendix B.3 and Appendix F as an assumption with a stated error bound, not concealed. The paper contains no load-bearing self-citation: [9] and [47] are not by the present authors, and no uniqueness theorem is imported. The absence of an RTC-only task-success table (C.1 defines NONE/RTC/FBFM but Tables 1–2 report only Base vs FBFM) is a real experimental-control gap: the action-prefix constraint alone could explain part of the gains. That is an ablation/correctness issue, not circularity. The one reduction-by-construction risk is the DreamZero configuration: PZ and kp were selected on tasks that later appear in the reported aggregate (Appendix E, Tables 10–12 vs Table 2). Accordingly the pooled +0.625 pp DreamZero figure is partly a tuned-selection score rather than an independent prediction of a fixed method. Because the paper discloses the search, the effect is small, and the RoboTwin track and mechanism diagnostics are not similarly tuned, the overall score is 4.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

FBFM introduces no new physical entities or learned parameters. It relies on the pretrained WAM's encoder-latent space being compatible with its clean-endpoint space, on a pseudo-asynchronous clock standing in for real runtime feedback, and on two tuned DreamZero gains (kp, PZ) plus a guidance clip. The RoboTwin task subset is also a manually selected evaluation set.

free parameters (3)
  • kp = 0.0486968
    DreamZero proportional state-feedback gain; selected by logarithmic screening on LIBERO-Spatial/Goal/Object/L-10 tasks (Appendix E.3) and then used in the final reported LIBERO results.
  • PZ = 56/9600
    DreamZero state modality preconditioner; selected from three candidates on LIBERO-Spatial/Objective subsets (Appendix E.2) and used in the final reported LIBERO results.
  • guidance clip beta = 10
    Flow-time-dependent guidance strength clip in Eq. (46); taken from prior work [47] with value set in Table 5, not derived.
axioms (4)
  • domain assumption Aligned-coordinate approximation: h†(h(X^1)) ≈ X^1 on the mask-selected subspace
    Appendix B.3 and Section 3.2: FBFM replaces the exact pseudoinverse discrepancy with direct coordinate comparison in latent/action space. The paper states this is not guaranteed for arbitrary nonlinear encoder-decoder pairs.
  • domain assumption Observed latent z_{t+i} encoded by the frozen VAE is a valid clean-endpoint coordinate for the flow-matching prior
    Feedback targets are compared directly against predicted clean endpoints; Appendix F suggests independently encoded observations may not remain semantically meaningful clean-video states under guided dynamics, producing artifacts after the measured prefix.
  • domain assumption Pseudo-asynchronous clock can stand in for real wall-clock feedback timing
    Section 3.5 and Section 4.2: a deterministic pseudo-asynchronous clock couples environment transitions to solver evaluations; no real-time robot control loop is demonstrated, so real-time feasibility is assumed.
  • domain assumption Pretrained WAM endpoint Jacobian provides a useful correction direction
    Eq. (19) relies on VJPs through frozen endpoint predictors. Section 5.3 notes DreamZero's reuse of stale DiT velocities/Jacobians can damp or delay corrections.

pith-pipeline@v1.3.0-daily-deepseek · 23661 in / 14022 out tokens · 142644 ms · 2026-08-03T11:02:19.298552+00:00 · methodology

0 comments
read the original abstract

Although world-action models (WAMs) enhance long-horizon robot control by predicting visual evolution before acting, long-horizon reliability demands repeated re-grounding in real observations--not recursive rollout. Existing WAMs address this by refreshing history or KV cache with ground-truth data between chunks. However, such chunk-wise feedback operates at a coarse temporal granularity and thus fails to correct prediction errors at the individual time-step level. To address this, we propose Feedback Flow Matching (FBFM), a training-free inference mechanism that pushes re-grounding inside the actively generated chunk. During flow matching, FBFM applies a masked pseudoinverse correction to the conditional velocity field: it leverages the preceding action chunk to guide generation of the next action chunk, and uses the image observed after executing that preceding chunk to guide the next frame prediction. This cross-chunk pairing--where feedback from one chunk arrives in time to shape the next--creates an asynchronous loop that corrects errors without waiting for chunk boundaries. Being training-free, the mechanism improves responsiveness to unexpected events and suppresses drift in long-horizon tasks. We evaluate FBFM on both a joint-generation WAM (DreamZero) and a stage-wise WAM (LingBot-VA). On selected LIBERO and RoboTwin2.0 tasks, it improves success rates by over 5% in favorable settings, and real-world robot observation-prediction diagnostics show notably better tracking. We argue that FBFM offers a new paradigm for fine-grained online correction, bridging open-loop flow generation with closed-loop real-world dynamics.

Figures

Figures reproduced from arXiv: 2607.29235 by Cong Huang, Kai Chen, Peize Li, Ruimeng Zhang, Ru Zhang, Shanghang Zhang.

Figure 1
Figure 1. Figure 1: FBFM for a stage-wise WAM. While the preceding chunk is executed, encoded real observations progressively [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FBFM for a joint-generation WAM. Encoded transitions observed while the preceding chunk is executed and committed [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Masked FBFM guidance in stage-wise and joint-generation WAMs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world robot-arm ball-stopping observation prediction. The RGB sequence was recorded from a physical robot [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LingBot-VA mechanism diagnostic. (a) Paired [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Image-space error under finite feedback coverage. Curves report frame-wise table-region MAE; vertical lines mark [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: State-feedback coverage on the auxiliary ball-collision sequence. Rows show 0, 0.25, 0.5, 1, 2, 3, 4, and 5 s; columns [PITH_FULL_IMAGE:figures/full_fig_p029_7.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

62 extracted references · 20 linked inside Pith

  1. [1]

    Weiss, Niru Mah- eswaranathan,andSuryaGanguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah- eswaranathan,andSuryaGanguli. DeepUnsupervised Learning using Nonequilibrium Thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, 2015

    2015

  2. [2]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, 2020

    2020

  3. [3]

    De- noising Diffusion Implicit Models

    JiamingSong,ChenlinMeng,andStefanoErmon. De- noising Diffusion Implicit Models. InInternational Conference on Learning Representations, 2021

    2021

  4. [4]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. InInternational Conference on Learning Representations, 2021

    2021

  5. [5]

    High-Resolution Image Synthesis with Latent Diffusion Models

    RobinRombach,AndreasBlattmann,DominikLorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2022

    2022

  6. [6]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling. InInternational Conference on Learning Representations, 2023

    2023

  7. [7]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. InInternational Conference on Learning Representations, 2023

    2023

  8. [8]

    Re- Paint: Inpainting using Denoising Diffusion Proba- bilisticModels.InIEEE/CVFConferenceonComputer Vision and Pattern Recognition, 2022

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Re- Paint: Inpainting using Denoising Diffusion Proba- bilisticModels.InIEEE/CVFConferenceonComputer Vision and Pattern Recognition, 2022

    2022

  9. [9]

    Pseudoinverse-Guided Diffusion Models for Inverse Problems

    Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-Guided Diffusion Models for Inverse Problems. InInternational Conference on Learning Representations, 2023

    2023

  10. [10]

    Training-FreeLinearImageInverses via Flows

    AshwiniPokle,MatthewJ.Muckley,RickyT.Q.Chen, andBrianKarrer. Training-FreeLinearImageInverses via Flows. InInternational Conference on Learning Representations, 2024

    2024

  11. [11]

    World Models

    David Ha and Jürgen Schmidhuber. World Models. arXiv:1803.10122, 2018

    Pith/arXiv arXiv 2018

  12. [12]

    Model-Based Reinforcement Learning for Atari

    Lukasz Kaiser et al. Model-Based Reinforcement Learning for Atari. InInternational Conference on Learning Representations, 2020

    2020

  13. [13]

    LearningLatentDynamicsforPlanningfromPixels

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas,DavidHa,HonglakLee,andJamesDavidson. LearningLatentDynamicsforPlanningfromPixels. In Proceedings of the 36th International Conference on Machine Learning, 2019

    2019

  14. [14]

    DreamtoControl: LearningBehav- iorsbyLatentImagination.InInternationalConference on Learning Representations, 2020

    DanijarHafner,TimothyLillicrap,JimmyBa,andMo- hammadNorouzi. DreamtoControl: LearningBehav- iorsbyLatentImagination.InInternationalConference on Learning Representations, 2020

    2020

  15. [15]

    Mastering Atari with Dis- crete World Models

    Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering Atari with Dis- crete World Models. InInternational Conference on Learning Representations, 2021

    2021

  16. [16]

    Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model.Nature, 588:604–609, 2020

    Julian Schrittwieser et al. Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model.Nature, 588:604–609, 2020

    2020

  17. [17]

    Tempo- ral Difference Learning for Model Predictive Control

    NicklasHansen,XiaolongWang,andHaoSu. Tempo- ral Difference Learning for Model Predictive Control. InProceedingsofthe39thInternationalConferenceon Machine Learning, 2022

    2022

  18. [18]

    MasteringDiverseControlTasksthrough World Models.Nature, 640:647–653, 2025

    DanijarHafner,JurgisPasukonis,JimmyBa,andTimo- thyLillicrap. MasteringDiverseControlTasksthrough World Models.Nature, 640:647–653, 2025

    2025

  19. [19]

    MCVD: Masked Conditional Video DiffusionforPrediction,Generation,andInterpolation

    Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video DiffusionforPrediction,Generation,andInterpolation. InAdvancesinNeuralInformationProcessingSystems, 2022

    2022

  20. [20]

    Learning Interactive Real- World Simulators

    Sherry Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Tompson, Leslie Kaelbling, Dale Schuur- mans, and Pieter Abbeel. Learning Interactive Real- World Simulators. InInternational Conference on Learning Representations, 2024

    2024

  21. [21]

    Genie: Generative Interactive En- vironments

    Jake Bruce et al. Genie: Generative Interactive En- vironments. InProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  22. [22]

    Diffusion for World Modeling: Visual Details Matter in Atari

    Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, and Francois Fleuret. Diffusion for World Modeling: Visual Details Matter in Atari. InAdvances in Neural Information Processing Systems, 2024

    2024

  23. [23]

    IRASim: A Fine- Grained World Model for Robot Manipulation

    Fangqi Zhu, Hongtao Wu, Song Guo, Yuxiao Liu, Chilam Cheang, and Tao Kong. IRASim: A Fine- Grained World Model for Robot Manipulation. InIn- ternational Conference on Learning Representations, 2025

    2025

  24. [24]

    UnleashingLarge-ScaleVideoGen- erative Pre-Training for Visual Robot Manipulation

    HongtaoWuetal. UnleashingLarge-ScaleVideoGen- erative Pre-Training for Visual Robot Manipulation. arXiv:2312.13139, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    GR-2: A Generative Video- Language-Action Model with Web-Scale Knowledge for Robot Manipulation

    Chi-Lam Cheang et al. GR-2: A Generative Video- Language-Action Model with Web-Scale Knowledge for Robot Manipulation. arXiv:2410.06158, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    CausalWorldModelingforRobotControl

    LinLietal. CausalWorldModelingforRobotControl. arXiv:2601.21998, 2026

    Pith/arXiv arXiv 2026

  27. [27]

    World Action Models are Zero- Shot Policies

    Seonghyeon Ye et al. World Action Models are Zero- Shot Policies. arXiv:2602.15922, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Ran- domization for Robust Bimanual Robotic Manipula- tion

    Tianxing Chen et al. RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Ran- domization for Robust Bimanual Robotic Manipula- tion. arXiv:2506.18088, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    arXiv:2509.15965, 2025

    ChaoYuetal.RLinf: FlexibleandEfficientLarge-scale Reinforcement Learning via Macro-to-Micro Flow Transformation. arXiv:2509.15965, 2025

    arXiv 2025

  30. [30]

    LIBERO:Benchmark- ing Knowledge Transfer for Lifelong Robot Learning

    BoLiu,YifengZhu,ChongkaiGao,YihaoFeng,Qiang Liu,YukeZhu,andPeterStone. LIBERO:Benchmark- ing Knowledge Transfer for Lifelong Robot Learning. InAdvancesinNeuralInformationProcessingSystems, 2023

    2023

  31. [31]

    Fast-WAM:DoWorldActionModelsNeedTest- Time Future Imagination? arXiv:2603.16666, 2026

    Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-WAM:DoWorldActionModelsNeedTest- Time Future Imagination? arXiv:2603.16666, 2026

    Pith/arXiv arXiv 2026

  32. [32]

    Tenenbaum, and Sergey Levine

    Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with Diffusion for Flexible Behavior Synthesis. InProceedings of the 39th Inter- national Conference on Machine Learning, 2022

    2022

  33. [33]

    IsCondi- tionalGenerativeModelingAllYouNeedforDecision- Making? arXiv:2211.15657, 2023

    Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenen- baum, TommiJaakkola, andPulkitAgrawal. IsCondi- tionalGenerativeModelingAllYouNeedforDecision- Making? arXiv:2211.15657, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning Fine-Grained Bimanual Ma- nipulation with Low-Cost Hardware. InRobotics: Sci- ence and Systems, 2023

    2023

  35. [35]

    Diffusion Policy: Visuomotor Policy Learn- ing via Action Diffusion

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, YilunDu,BenjaminBurchfiel,RussTedrake,andShu- ran Song. Diffusion Policy: Visuomotor Policy Learn- ing via Action Diffusion. InRobotics: Science and Systems, 2023

    2023

  36. [36]

    3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3DRepresentations

    YanjieZe,GuZhang,KangningZhang,ChenyuanHu, Muhan Wang, and Huazhe Xu. 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3DRepresentations. InRobotics: ScienceandSystems, 2024

    2024

  37. [37]

    RDT-1B: A Diffusion Foundation ModelforBimanualManipulation

    Songming Liu et al. RDT-1B: A Diffusion Foundation ModelforBimanualManipulation. arXiv:2410.07864, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    RT-1: RoboticsTransformerfor Real-WorldControlatScale

    AnthonyBrohanetal. RT-1: RoboticsTransformerfor Real-WorldControlatScale. InRobotics: Scienceand Systems, 2023

    2023

  39. [39]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

    Anthony Brohan et al. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Octo: AnOpen-SourceGener- alist Robot Policy

    OctoModelTeametal. Octo: AnOpen-SourceGener- alist Robot Policy. InRobotics: Science and Systems, 2024

    2024

  41. [41]

    InConference on Robot Learning, 2024

    MooJinKimetal.OpenVLA:AnOpen-SourceVision- Language-Action Model. InConference on Robot Learning, 2024

    2024

  42. [42]

    Riemannian Flow Matching Policy for Robot Motion Learning

    MaxBraun,NoemieJaquier, LeonelRozo, andTamim Asfour. Riemannian Flow Matching Policy for Robot Motion Learning. InIEEE/RSJ International Confer- ence on Intelligent Robots and Systems, 2024

    2024

  43. [43]

    FlowMatchingImitationLearn- ing for Multi-Support Manipulation

    Quentin Rouxel, Andrea Ferrari, Serena Ivaldi, and Jean-BaptisteMouret. FlowMatchingImitationLearn- ing for Multi-Support Manipulation. InIEEE-RAS In- ternational Conference on Humanoid Robots, 2024

    2024

  44. [44]

    Affordance- Based Robot Manipulation with Flow Matching

    Fan Zhang and Michael Gienger. Affordance- Based Robot Manipulation with Flow Matching. arXiv:2409.01083, 2025

    arXiv 2025

  45. [45]

    arXiv:2410.24164, 2024

    KevinBlacketal.π 0: AVision-Language-ActionFlow Model for General Robot Control. arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  46. [46]

    arXiv:2504.16054, 2025

    Physical Intelligence et al.π 0.5: A Vision- Language-ActionModelwithOpen-WorldGeneraliza- tion. arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Galliker, and Sergey Levine

    Kevin Black, Manuel Y. Galliker, and Sergey Levine. Real-Time Execution of Action Chunking Flow Poli- cies. InAdvances in Neural Information Processing Systems, 2025

    2025

  48. [48]

    Ren, Michael Equi, and Sergey Levine

    Kevin Black, Allen Z. Ren, Michael Equi, and Sergey Levine. Training-Time Action Conditioning for Effi- cient Real-Time Chunking. arXiv:2512.05964, 2025

    arXiv 2025

  49. [49]

    Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling

    Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie, Yoonho Lee, Maximilian Du, and Chelsea Finn. Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. InInternational Conference on Learning Representations, 2025

    2025

  50. [50]

    Learning Native Continuation for Action Chunking Flow Policies

    Yufeng Liu et al. Learning Native Continuation for Action Chunking Flow Policies. InRobotics: Science and Systems, 2026

    2026

  51. [51]

    Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection

    Trong-Bao Ho et al. Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection. arXiv:2606.19774, 2026

    Pith/arXiv arXiv 2026

  52. [52]

    RA-DP: Rapid Adaptive Diffusion Policy for Training-Free High-Frequency Robotics Replanning

    XiYe,RuiHengYang,JunJin,YinchuanLi,andAmir Rasouli. RA-DP: Rapid Adaptive Diffusion Policy for Training-Free High-Frequency Robotics Replanning. InIEEE/RSJ International Conference on Intelligent Robots and Systems, 2025

    2025

  53. [53]

    VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference

    Jiaming Tang et al. VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference. arXiv:2512.01031, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    LeaveNoObser- vation Behind: Real-Time Correction for VLA Action Chunks

    Kohei Sendai, Maxime Alvarez, Tatsuya Matsushima, YutakaMatsuo,andYusukeIwasawa. LeaveNoObser- vation Behind: Real-Time Correction for VLA Action Chunks. arXiv:2509.23224, 2025

    arXiv 2025

  55. [55]

    Closed-LoopAction Chunks with Dynamic Corrections for Training-Free DiffusionPolicy

    Pengyuan Wu, Pingrui Zhang, Zhigang Wang, Dong Wang,BinZhao,andXuelongLi. Closed-LoopAction Chunks with Dynamic Corrections for Training-Free DiffusionPolicy. InIEEEInternationalConferenceon Robotics and Automation, 2026

    2026

  56. [56]

    AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models

    Yuhua Jiang, Shuang Cheng, Yan Ding, Feifei Gao, and Biqing Qi. AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models. arXiv:2511.14148, 2026

    Pith/arXiv arXiv 2026

  57. [57]

    TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control

    Yuteng Sun, Haoran Wang, Ruofei Bai, Zhengguo Li, Jun Li, Meng Yee Michael Chuah, and Wei Yun Yau. TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control. arXiv:2601.14945, 2026

    Pith/arXiv arXiv 2026

  58. [58]

    DynaGuide: Steer- ing Diffusion Polices with Active Dynamic Guidance

    Maximilian Du and Shuran Song. DynaGuide: Steer- ing Diffusion Polices with Active Dynamic Guidance. InAdvancesinNeuralInformationProcessingSystems, 2025

    2025

  59. [59]

    Feedback World Model Enables Precise GuidanceofDiffusionPolicy.arXiv:2605.15705,2026

    Tuo An et al. Feedback World Model Enables Precise GuidanceofDiffusionPolicy.arXiv:2605.15705,2026

    Pith/arXiv arXiv 2026

  60. [60]

    AHA-WAM: Asynchronous Horizon- Adaptive World-Action Modeling with Observation- Guided Context Routing

    Jisong Cai et al. AHA-WAM: Asynchronous Horizon- Adaptive World-Action Modeling with Observation- Guided Context Routing. arXiv:2606.09811, 2026

    Pith/arXiv arXiv 2026

  61. [61]

    SteeringRobustnessintoWorldAction Models via Mechanistic Interpretability and Optimal Control

    Jihoon Hong, Julian Skifstad, Qiyue Dai, Alice Chan, andGlenChou. SteeringRobustnessintoWorldAction Models via Mechanistic Interpretability and Optimal Control. arXiv:2607.14943, 2026

    Pith/arXiv arXiv 2026

  62. [62]

    arXiv:2503.20314, 2025

    TeamWanetal.Wan: OpenandAdvancedLarge-Scale Video Generative Models. arXiv:2503.20314, 2025. A Notation Index For quick reference, Table 3 indexes the core notation introduced in Preliminaries, and Table 4 lists the extensions used by the stage-wise and joint-generation FBFM formulations. Symbol Definition T,T, cT , ρT A task, the task family, the conditi...

    Pith/arXiv arXiv 2025