pith. sign in

arxiv: 2606.19774 · v1 · pith:M53SEZNDnew · submitted 2026-06-18 · 💻 cs.RO

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

Trong-Bao Ho , Quang-Tan Nguyen , Thien-Loc Ha , Gia-Binh Nguyen , Viet-Thanh Nguyen , Long Dinh , Minh N. Vu , Duy M. H. Nguyen
show 2 more authors
An Thai Le Ngo Anh Vien
This is my paper

Pith reviewed 2026-06-26 17:29 UTC · model grok-4.3

classification 💻 cs.RO
keywords action chunkingflow-based policiesasynchronous executionrobot manipulationnoise selectionbackward Euler inversion
0
0 comments X

The pith

Selecting the right initial noise lets unmodified flow ODEs generate consistent action chunks for asynchronous robot control.

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

The paper establishes that prefix consistency under asynchronous execution can be achieved simply by choosing an appropriate initial noise before generation starts, so the standard flow ODE produces a coherent next chunk without any steering or policy changes. This reframes the latency problem as one of noise selection rather than trajectory correction. A sympathetic reader would care because the approach requires no gradients, retraining, or model modification yet still raises consistency and task success on many simulated and real robot setups.

Core claim

Prefix consistency can be achieved by selecting an appropriate initial noise before generation begins, allowing the unmodified flow ODE to produce a coherent next chunk. This reframes asynchronous inference as a noise selection problem rather than a trajectory steering problem. PAINT finds this noise via backward Euler inversion and constructs the final chunk through a repainting rule; it requires no gradients, retraining, or policy modification yet improves execution consistency and task performance across 12 simulated benchmarks and 6 real-world manipulation tasks.

What carries the argument

Initial noise selection via backward Euler inversion, which enables the unmodified flow ODE to maintain prefix consistency when producing the next action chunk.

If this is right

  • Asynchronous execution becomes possible while keeping the original flow policy unchanged.
  • Consistency and task performance improve across 12 simulated benchmarks without added computation for steering.
  • The same gains appear on 6 real-world tasks spanning single-arm, bimanual, and humanoid robots.
  • The method eliminates the need for gradients or policy retraining at deployment time.

Where Pith is reading between the lines

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

  • The noise-selection view may transfer to other continuous generative models used for robot actions.
  • Avoiding steering iterations could reduce control-loop latency beyond what the paper measures.
  • The repainting rule might combine with other inference accelerations while preserving the core guarantee.

Load-bearing premise

An appropriate initial noise exists and can be reliably recovered via backward Euler inversion so that the unmodified flow ODE produces a coherent next chunk without steering.

What would settle it

If applying backward Euler inversion on the evaluated benchmarks yields no measurable gain in boundary consistency or task success relative to existing steering methods, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.19774 by An Thai Le, Duy M. H. Nguyen, Gia-Binh Nguyen, Long Dinh, Minh N. Vu, Ngo Anh Vien, Quang-Tan Nguyen, Thien-Loc Ha, Trong-Bao Ho, Viet-Thanh Nguyen.

Figure 1
Figure 1. Figure 1: The prefix constraint: the first d actions of chunk At must approximate the last d actions of chunk At−1. A carefully chosen initial noise sat￾isfies this constraint without modifying the policy. Flow matching [1] and diffusion [2] policies have achieved remarkable dexterity by predict￾ing action chunks, sequences of future actions generated in a single forward pass [3, 4, 5]. Action chunking improves temp… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of standard diffusion policy [4, 5, 12] (a), RTC [6] (b), and our proposed method PAINT (c), which leverages modified noise. Given a frozen, pretrained diffusion or flow-matching policy πθ, instead of modifying the ODE at each denoising step, we modify the initial noise x0 to get x ∗ 0 , which satisfies the prefix constraint (Equation (1)). We show that x ∗ 0 can improve policy performance while s… view at source ↗
Figure 3
Figure 3. Figure 3: Top left: Inference delay (d) vs. prefix consistency (CON ↓) across all environments. Bottom left: Execution horizon vs. success rate (SR ↑) at d=1. Right: Inference delay (d) vs. success rate (SR ↑) across simulated d ∈ {0, 1, 2, 3, 4}. PAINT-Euler achieves the strongest overall performance across all delay values. Each data point aggregates 2048 trials. 5.1 Simulated Benchmark Setup. We follow the exact … view at source ↗
Figure 4
Figure 4. Figure 4: Average consistency scores and success rates over environments. Among various inver￾sion methods, our chosen method (Euler) offers the best balance between quality and complexity. Results [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The environments for real-world evaluation. Each sub-figure’s left image shows the initial [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The visualizer of 12 environments in Kinetix [37]. A.1 Baseline Details We compare PAINT against representative baselines for sync/asynchronous action-chunk execution. Unless otherwise stated, all methods use the same pretrained policy checkpoint and differ only in how action chunks are executed, filtered, or modified at chunk boundaries. Naive Async. The robot executes the current chunk while the next chu… view at source ↗
Figure 7
Figure 7. Figure 7: PAINT v.s. training-time delay-aware methods on the Kinetix benchmark. We evaluate PAINT on top of A2C2 [18] and training-time RTC (TT-RTC) [7] , measuring both consistency score and task success rate under varying execution horizons and inference delays. Hardware Setup. For ALOHA [39] experiments, the robot consists of two ViperX 6-DoF arms, each equipped with a parallel-jaw gripper, for a total of 14 con… view at source ↗
Figure 8
Figure 8. Figure 8: Examples of completed sequences in real-world tasks. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
read the original abstract

Action chunking enables robot policies to produce temporally coherent behavior, but generating multi-step action sequences with flow-based policies incurs latency that is incompatible with real-time control. Under asynchronous execution, the robot continues executing the current chunk while the next one is generated, causing even minor delays to create inconsistencies at chunk boundaries. Existing methods address this problem by steering generation toward the already executed action prefix. We instead show that prefix consistency can be achieved by selecting an appropriate initial noise before generation begins, allowing the unmodified flow ODE to produce a coherent next chunk. This reframes asynchronous inference as a noise selection problem rather than a trajectory steering problem. We introduce \textbf{PAINT}, a training-free method that finds this noise via backward Euler inversion and constructs the final chunk through a repainting rule. In summary, \texttt{PAINT} requires no gradients, retraining, or policy modification; yet it improves execution consistency and task performance across \textit{12 simulated benchmarks} and \textit{6 real-world manipulation tasks} spanning single-arm, bimanual, and humanoid embodiments. Website: ~\href{https://paint-action-chunking.github.io}{\texttt{https://paint-action-chunking.github.io}}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that asynchronous execution of action chunks from flow-based robot policies can achieve prefix consistency at chunk boundaries by selecting an appropriate initial noise via backward Euler inversion, rather than steering the generation process. This allows the unmodified flow ODE to produce a coherent next chunk. The authors introduce the training-free PAINT method, which uses this inversion and a repainting rule to construct the chunk. They report improved execution consistency and task performance on 12 simulated benchmarks and 6 real-world manipulation tasks across single-arm, bimanual, and humanoid robots.

Significance. If the central claim holds, the work reframes asynchronous inference as an initial-noise selection problem instead of trajectory steering. This could simplify real-time control implementations by avoiding policy modifications, gradients, or retraining. The training-free aspect and evaluation spanning simulation and real hardware across multiple embodiments would indicate practical utility for flow-based policies in robotics.

major comments (2)
  1. [Abstract / Method description] The central claim (abstract) that backward Euler inversion recovers an initial noise such that the unmodified flow ODE produces a next chunk whose initial segment is consistent with the executed prefix within control tolerances is load-bearing. Backward Euler is first-order; without a quantitative error analysis or bounds on truncation error accumulation in high-dimensional action spaces (as raised by the stress-test note), it is unclear whether the recovered noise guarantees coherence without implicit steering.
  2. [Abstract] The abstract states that the repainting rule constructs the final chunk and removes dependence on accurate inversion, but provides no derivation or pseudocode showing how repainting interacts with the inverted noise to enforce prefix consistency. This interaction is required to support the claim that the method avoids steering entirely.
minor comments (2)
  1. [Abstract] The abstract mentions improvements across 12 simulated benchmarks and 6 real tasks but does not specify the metrics, baselines, or statistical significance (e.g., error bars or p-values) used to support the performance claims.
  2. No mention of code or data release; adding a reproducibility statement would strengthen the training-free claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Method description] The central claim (abstract) that backward Euler inversion recovers an initial noise such that the unmodified flow ODE produces a next chunk whose initial segment is consistent with the executed prefix within control tolerances is load-bearing. Backward Euler is first-order; without a quantitative error analysis or bounds on truncation error accumulation in high-dimensional action spaces (as raised by the stress-test note), it is unclear whether the recovered noise guarantees coherence without implicit steering.

    Authors: We agree that the first-order accuracy of backward Euler warrants explicit discussion of truncation error, especially in high-dimensional action spaces. The manuscript provides extensive empirical validation across 12 simulated and 6 real-world tasks showing that the recovered noise yields prefix-consistent chunks within control tolerances when paired with the repainting rule. To address the concern directly, we will add a new subsection in the Methods that derives a practical error bound based on the Lipschitz continuity of the learned velocity field and includes numerical stress tests on error accumulation over typical chunk lengths. revision: yes

  2. Referee: [Abstract] The abstract states that the repainting rule constructs the final chunk and removes dependence on accurate inversion, but provides no derivation or pseudocode showing how repainting interacts with the inverted noise to enforce prefix consistency. This interaction is required to support the claim that the method avoids steering entirely.

    Authors: The full derivation of the repainting rule, its interaction with the inverted noise, and the supporting pseudocode appear in Section 3.2 and Algorithm 1. Briefly, after integrating the unmodified flow ODE from the inverted noise, the rule overwrites the initial segment of the resulting chunk with the executed prefix; this step enforces consistency without altering the ODE itself. We will revise the abstract to include one sentence summarizing this interaction so that the high-level claim is self-contained. revision: yes

Circularity Check

0 steps flagged

No circularity: method is a training-free inversion procedure validated empirically on external benchmarks

full rationale

The paper reframes asynchronous chunking as an initial-noise selection problem solved by backward Euler inversion plus a repainting rule (PAINT). The central claim is that this unmodified-ODE procedure yields prefix-consistent chunks; this is presented as an algorithmic construction whose correctness is assessed via task success rates on 12 simulated and 6 real-world benchmarks rather than by any definitional equivalence, fitted-parameter renaming, or self-citation chain. No equations reduce the output to the input by construction, and the method is explicitly training-free with no learned parameters tuned to the target consistency metric.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review is based solely on the abstract; no specific free parameters, axioms, or invented entities are detailed in the provided text.

pith-pipeline@v0.9.1-grok · 5782 in / 1029 out tokens · 26411 ms · 2026-06-26T17:29:49.351757+00:00 · methodology

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

42 extracted references · 17 linked inside Pith

  1. [1]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. In11th International Conference on Learning Representations, ICLR 2023, 2023

    2023

  2. [2]

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

    2020

  3. [3]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  4. [4]

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

    2025

  5. [5]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Black, M

    K. Black, M. Y . Galliker, and S. Levine. Real-time execution of action chunking flow policies. arXiv preprint arXiv:2506.07339, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Black, A

    K. Black, A. Z. Ren, M. Equi, and S. Levine. Training-time action conditioning for efficient real-time chunking.arXiv preprint arXiv:2512.05964, 2025

    arXiv 2025

  8. [8]

    J. Tang, Y . Sun, Y . Zhao, S. Yang, Y . Lin, Z. Zhang, J. Hou, Y . Lu, Z. Liu, and S. Han. Vlash: Real-time vlas via future-state-aware asynchronous inference.arXiv preprint arXiv:2512.01031, 2025

    arXiv 2025

  9. [9]

    Y . Liu, J. I. Hamid, A. Xie, Y . Lee, M. Du, and C. Finn. Bidirectional decoding: Improving action chunking via closed-loop resampling.arXiv preprint arXiv:2408.17355, 2024

    arXiv 2024

  10. [10]

    J. Mao, X. Wang, and K. Aizawa. Guided image synthesis via initial image editing in diffusion model. InProceedings of the 31st ACM International Conference on Multimedia, pages 5321– 5329, 2023

    2023

  11. [11]

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y . Zhang, J. Rector-Brooks, G. Wolf, and Y . Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Bjorck, F

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

    Pith/arXiv arXiv 2025

  13. [13]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, M. Y . Galliker, D. Ghosh, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. V...

    Pith/arXiv arXiv 2025

  14. [14]

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

    Pith/arXiv arXiv 2024

  15. [15]

    Shukor, D

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. Smolvla: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844, 2025. 9

    Pith/arXiv arXiv 2025

  16. [16]

    S. H. Høeg, Y . Du, and O. Egeland. Streaming diffusion policy: Fast policy synthesis with variable noise diffusion models.arXiv preprint arXiv:2406.04806, 2024

    arXiv 2024

  17. [17]

    Jiang, X

    S. Jiang, X. Fang, N. Roy, T. Lozano-P ´erez, L. P. Kaelbling, and S. Ancha. Streaming flow policy: Simplifying diffusion/flow policies by treating action trajectories as flow trajectories. InICRA 2025 Workshop: Beyond Pick and Place, 2025

    2025

  18. [18]

    Sendai, M

    K. Sendai, M. Alvarez, T. Matsushima, Y . Matsuo, and Y . Iwasawa. Leave no observation behind: Real-time correction for vla action chunks.arXiv preprint arXiv:2509.23224, 2025

    arXiv 2025

  19. [19]

    N. R. Arachchige, Z. Chen, W. Jung, W. C. Shin, R. Bansal, P. Barroso, Y . H. He, Y . C. Lin, B. Joffe, S. Kousik, et al. Sail: Faster-than-demonstration execution of imitation learning policies.arXiv preprint arXiv:2506.11948, 2025

    arXiv 2025

  20. [20]

    Y . Lu, Z. Liu, X. Fan, Z. Yang, J. Hou, J. Li, K. Ding, and H. Zhao. Faster: Rethinking real-time flow vlas.arXiv preprint arXiv:2603.19199, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    F. Yang, P. Jing, K. Qu, N. Zhao, and Y . Su. Abpolicy: Asynchronous b-spline flow policy for real-time and smooth robotic manipulation.International Conference on Robotics and Automation (ICRA), 2026

    2026

  22. [22]

    P. Wang, K. Hong, C. Peng, K. Driggs-Campbell, M. Tomizuka, C. Xu, and C. Tang. Dis- cretertc: Discrete diffusion policies are natural asynchronous executors.arXiv preprint arXiv:2604.25050, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Agouzoul

    A. Agouzoul. Understanding asynchronous inference methods for vision-language-action models.arXiv preprint arXiv:2605.08168, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    J. Mao, X. Wang, and K. Aizawa. The lottery ticket hypothesis in denoising: Towards semantic-driven initialization. InEuropean Conference on Computer Vision, pages 93–109. Springer, 2024

    2024

  25. [25]

    Patil, O

    O. Patil, O. Biza, T. Weng, K. Schmeckpeper, W. Thomason, X. Zhang, R. Walters, N. Gopalan, S. Castro, and E. Rosen. You’ve got a golden ticket: Improving generative robot policies with a single noise vector.arXiv preprint arXiv:2603.15757, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    Wagenmaker, Y

    A. Wagenmaker, Y . Zhang, M. Nakamoto, S. Park, W. Yagoub, A. Nagabandi, A. Gupta, and S. Levine. Steering your diffusion policy with latent space reinforcement learning. In9th Annual Conference on Robot Learning, 2025

    2025

  27. [27]

    Y . Duan, H. Yin, and D. Kragic. Real-time iteration scheme for diffusion policy.arXiv preprint arXiv:2508.05396, 2025

    arXiv 2025

  28. [28]

    J. Jia, G. Li, X. Chen, T. An, Y . Hu, J. Li, X. Guo, and J. Yang. Action-to-action flow matching. arXiv preprint arXiv:2602.07322, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    J. Lu, X. Qin, Y . Jiang, K. Wang, C. Zhang, B. Liang, J. Yang, M. Xu, and L. Zhao. Unified noise steering for efficient human-guided VLA adaptation.arXiv preprint arXiv:2605.10821, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. InInternational Con- ference on Learning Representations, 2021

    2021

  31. [31]

    Meiri, D

    B. Meiri, D. Samuel, N. Darshan, G. Chechik, S. Avidan, and R. Ben-Ari. Fixed-point inver- sion for text-to-image diffusion models.arXiv e-prints, pages arXiv–2312, 2023

    2023

  32. [32]

    Z. Pan, R. Gherardi, X. Xie, and S. Huang. Effective real image editing with accelerated iterative diffusion inversion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15912–15921, 2023. 10

    2023

  33. [33]

    Zhang, J

    G. Zhang, J. P. Lewis, and W. B. Kleijn. Exact diffusion inversion via bidirectional integration approximation. InEuropean Conference on Computer Vision, pages 19–36. Springer, 2024

    2024

  34. [34]

    S. Hong, K. Lee, S. Y . Jeon, H. Bae, and S. Y . Chun. On exact inversion of dpm-solvers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7069–7078, 2024

    2024

  35. [35]

    Z. Li, S. Tang, and N. Azizan. Reverse flow matching: A unified framework for online rein- forcement learning with diffusion and flow policies.arXiv preprint arXiv:2601.08136, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Lipman, M

    Y . Lipman, M. Havasi, P. Holderrieth, N. Shaul, M. Le, B. Karrer, R. T. Chen, D. Lopez-Paz, H. Ben-Hamu, and I. Gat. Flow matching guide and code.arXiv preprint arXiv:2412.06264, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Matthews, M

    M. Matthews, M. Beukman, C. Lu, and J. N. Foerster. Kinetix: Investigating the training of general agents through open-ended physics-based control tasks. InThe Thirteenth Interna- tional Conference on Learning Representations, 2025

    2025

  38. [38]

    I. O. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Unterthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021

    2021

  39. [39]

    Z. Fu, T. Z. Zhao, and C. Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation.arXiv preprint arXiv:2401.02117, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    J. Song, A. Vahdat, M. Mardani, and J. Kautz. Pseudoinverse-guided diffusion models for inverse problems. InInternational conference on learning representations, 2023

    2023

  41. [41]

    C. Pan, G. Anantharaman, N.-C. Huang, C. Jin, D. Pfrommer, C. Yuan, F. Permenter, G. Qu, N. Boffi, G. Shi, et al. Much ado about noising: Dispelling the myths of generative robotic control.arXiv preprint arXiv:2512.01809, 2025

    arXiv 2025

  42. [42]

    X. Ye, R. H. Yang, J. Jin, Y . Li, and A. Rasouli. Ra-dp: Rapid adaptive diffusion policy for training-free high-frequency robotics replanning. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6227–6234. IEEE, 2025. 11 Appendix Contents A Kinetix Benchmark 13 A.1 Baseline Details . . . . . . . . . . . . . . . . . . ...

    2025