pith. sign in

arxiv: 2606.03847 · v1 · pith:YFMDQZY4new · submitted 2026-06-02 · 💻 cs.RO

Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies

Xiangdong Feng , Yuxuan Cheng , Chen Shi , Boyao Han , Yuxuan Yan , Yitong Hong , Zhuotao Tian , Li Jiang This is my paper

Pith reviewed 2026-06-28 09:24 UTC · model grok-4.3

classification 💻 cs.RO
keywords flow-based policiesaction chunkingdenoising varianceadaptive replanningrobot manipulationtask phase detection
0
0 comments X

The pith

Denoising variance in flow-based robot policies indicates when to replan chunks of actions.

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

The paper shows that flow-based policies produce stable clean-action estimates during predictable free-space motions but higher variance around contact-rich or precision-sensitive phases. This variance acts as an intrinsic signal for deciding how many actions from a predicted chunk to execute before replanning. DVAC measures variance over the final denoising steps, commits only the stable low-variance prefix, and triggers replanning before high-variance segments, with rolling calibration to handle varying tasks. The approach raises success rates while lowering replan frequency across simulation benchmarks and real-world manipulation.

Core claim

The denoising process of flow-based policies contains an intrinsic signal of task phases: clean-action estimates remain stable during predictable motion phases, but fluctuate more strongly around contact-rich or precision-sensitive operations. DVAC measures the variance of clean-action estimates over the final denoising steps, executes the stable low-variance prefix, and replans before high-variance future actions are committed. A rolling estimate of local variance scale transfers the method across tasks and rollouts.

What carries the argument

Denoising-Variance Adaptive Chunking (DVAC), which uses variance of clean-action estimates in the final denoising steps to set variable execution horizons instead of a fixed chunk length.

Load-bearing premise

The variance of clean-action estimates over the final denoising steps reliably indicates task phases requiring different replanning frequencies and generalizes across tasks when the threshold is calibrated via a rolling estimate of the local variance scale.

What would settle it

On a new set of manipulation tasks, the variance signal shows no consistent correlation with phase type, so that DVAC yields success rates and replan counts indistinguishable from fixed-length chunking.

Figures

Figures reproduced from arXiv: 2606.03847 by Boyao Han, Chen Shi, Li Jiang, Xiangdong Feng, Yitong Hong, Yuxuan Cheng, Yuxuan Yan, Zhuotao Tian.

Figure 1
Figure 1. Figure 1: Denoising variance varies across manipulation phases. (a) Representative LIBERO rollout: denoising variance remains low during moving phases and rises around contact-rich opera￾tion phases. (b) Phase statistics across LIBERO tasks: operation phases exhibit higher mean denois￾ing variance than moving phases. Detailed experimental settings are provided in Appendix 7.3. A natural solution is adaptive executio… view at source ↗
Figure 2
Figure 2. Figure 2: Method overview of DVAC. The final denoising steps are monitored to computes the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scale-adaptive thresholding. Two episodes selected from Libero have different denoising-variance scales, making a fixed τ conservative in one and permissive in the other. DVAC instead sets τs adaptively from the local rolling distribution. where α ≥ 0 specifies the uncertainty tolerance in units of local standard deviation. Although DVAC still introduces a hyperparameter, unlike a fixed τ , α defines a rol… view at source ↗
Figure 4
Figure 4. Figure 4: Real-world rollout visualization of DVAC. Inf. idx denotes the policy inference index, and Nexec denotes the number of executed actions. The shaded regions separating adjacent action chunks. DVAC executes longer chunks in low-variance moving phases and shortens the horizon near high-variance, operation-sensitive phases. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Threshold sensitivity on LIBERO across different backbones. Success rates remain relatively stable as the fixed threshold τ varies from 10−4 to 10−1 , but the optimal value depends on both the suite and the backbone. pi0.5 τ=10 −4 τ=5 × 10 −4 τ=2 × 10 −3 τ=2 × 10 −2 DVAC 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Mean success rate 0.359 0.367 0.413 0.385 0.373 0.416 RoboTwin Result based on Pi0.5 Baseline Fixed τ DVAC (a… view at source ↗
Figure 6
Figure 6. Figure 6: Summary of variance threshold sweep and phase correlation analysis. (a) On RoboTwin, fixed-threshold variants improve over the fixed-prefix baseline, while DVAC achieves the best success rate. (b) On LIBERO, the adaptive α scan yields more stable phase correlations than the fixed-τ scan, showing more consistent phase-aware execution. most 3.716, while DVAC improves it to 4.040. This suggests that denoising… view at source ↗
Figure 7
Figure 7. Figure 7: Phase-wise distribution of denoising variance on LIBERO. Across all four suites, OPERATING phases have higher median log10(Vtotal) than MOVING phases, indicating less stable denoising predictions during contact-rich or precision-sensitive behavior. where µ0 = E[x | y = 0], µ1 = E[x | y = 1], σx is the standard deviation of x, and p0 = n0/n, p1 = n1/n. Since OPERATING is encoded as y = 1, a negative correla… view at source ↗
Figure 8
Figure 8. Figure 8: Per-episode correlation between total denoising variance and task phase on LIBERO. Each bar reports the correlation between Vtotal and the binary phase label for one episode. goal_t05 e00 goal_t01 e00 goal_t00 e00 object_t07 e00 Long_t09 e00 object_t04 e00 spatial_t06 e00 Long_t06 e00 spatial_t08 e00 Long_t02 e00 Long_t05 e01 goal_t08 e00 goal_t04 e00 object_t09 e00 goal_t07 e00 spatial_t03 e00 Long_t00 e0… view at source ↗
Figure 9
Figure 9. Figure 9: Per-episode mean total denoising variance on LIBERO. Each bar reports the episode￾level mean of Vtotal. The large variation across episodes shows that the absolute variance scale is task- and rollout-dependent, motivating the scale-adaptive threshold used in DVAC [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of executed chunk lengths under fixed execution and adaptive DVAC. 25 50 75 100 125 150 175 200 225 Environment Step 1 2 3 5 7 10 N exec 10−2 10−1 V τ tot al / LIBERO-Spatial · Task 004 · Episode 000 (failure) Nexec step = 58 Vtotal τ Long chunk is executed near contact Approaching grasp region Failing to grasp the object again Insufficient time to complete [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 11
Figure 11. Figure 11: Representative failure case of DVAC on LIBERO. The curve shows the executed chunk length Nexec, total denoising variance Vtotal, and adaptive threshold τ over environment steps. Near the pre-grasp stage, variance remains insufficiently elevated, causing DVAC to execute a long chunk when shorter execution and earlier replanning would be safer. This delayed correction leads to grasp failure and task incompl… view at source ↗
Figure 12
Figure 12. Figure 12: Per-task RoboTwin success rates under different variance thresholds. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Time-series visualization of log10(Vtotal) on LIBERO-Long episodes. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Time-series visualization of log10(Vtotal) on LIBERO-Goal episodes. 50 100 150 -3 -2 -1 log10 (Vtotal) LIBERO-Object, task00, episode00 50 100 LIBERO-Object, task01, episode00 50 100 LIBERO-Object, task02, episode00 50 100 LIBERO-Object, task03, episode00 50 100 LIBERO-Object, task04, episode00 50 100 Step -3 -2 -1 log10 (Vtotal) LIBERO-Object, task05, episode00 50 100 150 Step LIBERO-Object, task06, epis… view at source ↗
Figure 15
Figure 15. Figure 15: Time-series visualization of log10(Vtotal) on LIBERO-Object episodes. 20 40 60 80 -2 0 log10 (Vtotal) LIBERO-Spatial, task00, episode00 25 50 75 100 125 LIBERO-Spatial, task01, episode00 25 50 75 100 125 LIBERO-Spatial, task02, episode00 20 40 60 80 LIBERO-Spatial, task03, episode00 25 50 75 100 125 LIBERO-Spatial, task04, episode00 20 40 60 80 100 Step -2 0 log10 (Vtotal) LIBERO-Spatial, task05, episode0… view at source ↗
Figure 16
Figure 16. Figure 16: Time-series visualization of log10(Vtotal) on LIBERO-Spatial episodes. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
read the original abstract

Action chunking has become a common inference strategy for flow-based robot policies, improving action coherence by modeling multi-step temporal dependencies in demonstrations. However, the execution horizon is still typically set as an empirical fixed value, overlooking that predictable free-space motions and precision-critical interaction phases often require different replanning frequencies. In this work, we first show that the denoising process of flow-based policies contains an intrinsic signal of task phases: clean-action estimates remain stable during predictable motion phases, but fluctuate more strongly around contact-rich or precision-sensitive operations. Motivated by this observation, we propose DVAC (Denoising-Variance Adaptive Chunking), a test-time method that adaptively determines how many actions to execute from each predicted chunk. DVAC measures the variance of clean-action estimates over the final denoising steps, executes the stable low-variance prefix, and replans before high-variance future actions are committed. To transfer across tasks and rollouts, DVAC further calibrates the threshold with a rolling estimate of the local variance scale. Experiments on LIBERO, RoboTwin, CALVIN, and real-world manipulation show that DVAC improves task success while reducing replanning frequency. With a $\pi_{0.5}$-based policy, DVAC improves LIBERO success from 94.75% to 98.00% and reduces replanning by 43.0%, while also yielding aggregate gains on RoboTwin and CALVIN and improving real-world execution efficiency.

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 the denoising process in flow-based robot policies contains an intrinsic signal of task phases, with clean-action estimates showing higher variance during contact-rich or precision-sensitive operations than in predictable free-space motion. Motivated by this, it proposes DVAC, a test-time adaptive chunking method that measures variance over final denoising steps, executes stable low-variance prefixes, and uses rolling calibration of the local variance scale to decide replanning points. This yields higher task success and lower replanning frequency than fixed chunking on LIBERO (94.75% to 98.00% success, 43% replan reduction with a π0.5 policy), RoboTwin, CALVIN, and real hardware.

Significance. If the empirical results hold, the work offers a practical, architecture-agnostic, training-free improvement to flow-based policies by exploiting a property of the existing denoising process. The test-time nature, evaluation across multiple simulation benchmarks plus real hardware, and direct comparison to fixed chunking are strengths that could make the heuristic immediately useful for deployment.

major comments (2)
  1. [Experiments / abstract results] The experimental results (abstract and implied Experiments section) report specific gains such as the LIBERO success improvement and 43% replanning reduction without error bars, number of trials, or statistical significance tests. This leaves open whether the gains are robust or could arise from unexamined experimental choices in threshold calibration or rollout selection.
  2. [Method (DVAC description) and Experiments] The weakest assumption—that the variance signal over final denoising steps reliably indicates phases and generalizes via rolling calibration—is presented as an observation but lacks an ablation on the calibration window size or sensitivity of performance to the variance threshold. Without this, it is unclear if the method's benefits are tied to post-hoc tuning on the evaluated tasks.
minor comments (2)
  1. [Abstract] The abstract introduces π0.5 without definition; the main text should clarify what this base policy is (e.g., a specific flow-matching model) on first use.
  2. [Method] Notation for the variance measure and rolling estimate could be formalized with an equation to improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive recommendation. We address the two major comments below and will revise the manuscript accordingly to improve clarity and robustness of the reported results.

read point-by-point responses

  1. Referee: [Experiments / abstract results] The experimental results (abstract and implied Experiments section) report specific gains such as the LIBERO success improvement and 43% replanning reduction without error bars, number of trials, or statistical significance tests. This leaves open whether the gains are robust or could arise from unexamined experimental choices in threshold calibration or rollout selection.

    Authors: We agree that the current presentation lacks sufficient statistical detail. In the revised manuscript we will explicitly state the number of evaluation rollouts per task (50 on LIBERO, 30 on RoboTwin and CALVIN), report mean success rates with standard deviations across three random seeds where available, and include a statistical significance test (paired t-test) comparing DVAC against fixed chunking. If any benchmark was evaluated with a single seed, we will note this limitation and add multi-seed results for the camera-ready version. revision: yes

  2. Referee: [Method (DVAC description) and Experiments] The weakest assumption—that the variance signal over final denoising steps reliably indicates phases and generalizes via rolling calibration—is presented as an observation but lacks an ablation on the calibration window size or sensitivity of performance to the variance threshold. Without this, it is unclear if the method's benefits are tied to post-hoc tuning on the evaluated tasks.

    Authors: We accept that an explicit ablation would strengthen the claim of robustness. We will add a new subsection in Experiments that varies the rolling calibration window (5, 10, 20 steps) and the variance threshold multiplier (±10 % and ±20 % around the calibrated value) on LIBERO and RoboTwin. The results will show that performance remains above the fixed-chunking baseline across a reasonable range of these hyperparameters, confirming that the gains are not the result of narrow post-hoc tuning. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central contribution is an empirical observation (clean-action variance over final denoising steps is higher in contact/precision phases) followed by a purely test-time heuristic (DVAC) that uses a rolling variance-scale threshold to decide chunk execution length. No derivation chain reduces a claimed prediction or result to its own inputs by construction: there are no fitted parameters renamed as predictions, no self-definitional loops, and no load-bearing self-citations or uniqueness theorems. The method makes no architectural change to the underlying flow policy and is evaluated against a fixed-chunking baseline on multiple external benchmarks. The reported gains (e.g., LIBERO 94.75% → 98.00%) are therefore direct empirical outcomes of the heuristic rather than quantities forced by the paper's own equations or prior self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on one domain assumption about the meaning of denoising variance; no free parameters or invented entities are introduced beyond the rolling calibration procedure.

axioms (1)
  • domain assumption Variance of clean-action estimates over the final denoising steps indicates task phase predictability.
    This observation is stated as the intrinsic signal that motivates DVAC and is treated as given for the method to function.

pith-pipeline@v0.9.1-grok · 5822 in / 1462 out tokens · 35959 ms · 2026-06-28T09:24:33.540190+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

41 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, et al. RT-1: Robotics transformer for real-world control at scale. In Robotics: Science and Systems, 2023. URLhttps://arxiv.org/abs/2212.06817

    Pith/arXiv arXiv 2023

  2. [2]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. In Proceedings of The 7th Conference on Robot Learning, volume 229 ofProceedings of Machine Learning Research, pages 2165–2183. PMLR, 2023. URLhttps://proceedings.mlr. press/v229...

    2023

  3. [3]

    Driess, F

    D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y . Chebotar, P. Sermanet, D. Duckworth, S. Levine, V . Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence. PaLM-E: An embodied multimodal language model. InProceedings of the 40th International Confere...

  4. [4]

    URLhttps://proceedings.mlr.press/v202/driess23a.html

    PMLR, 2023. URLhttps://proceedings.mlr.press/v202/driess23a.html

    2023

  5. [5]

    Padalkar, A

    Open X-Embodiment Collaboration, A. Padalkar, A. Pooley, A. Jain, A. Bewley, A. Herzog, A. Irpan, A. Khazatsky, A. Rai, A. Singh, et al. Open x-embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864, 2023. URLhttps://arxiv.org/abs/ 2310.08864

    Pith/arXiv arXiv 2023

  6. [6]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024. URLhttps://arxiv.org/abs/2405.12213

    Pith/arXiv arXiv 2024

  7. [7]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. OpenVLA: An open-source vision-language-action model. InProceedings of The 8th Conference on Robot Learning, volume 270 ofProceedings of Machine Learni...

    2025

  8. [8]

    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:2410.07291, 2024. URLhttps://arxiv.org/abs/2410.07291

    arXiv 2024

  9. [9]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success. InRobotics: Science and Systems, 2025. URLhttps://arxiv.org/ abs/2502.19645

    Pith/arXiv arXiv 2025

  10. [10]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InRobotics: Science and Systems, Daegu, Republic of Korea, July

  11. [11]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

    doi:10.15607/RSS.2023.XIX.016. URLhttps://arxiv.org/abs/2304.13705

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.15607/rss.2023.xix.016 2023

  12. [12]

    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. InRobotics: Science and Systems,

  13. [13]

    URLhttps://arxiv.org/abs/2303.04137

    Pith/arXiv arXiv

  14. [15]

    D. Jing, G. Wang, J. Liu, W. Tang, Z. Sun, Y . Yao, Z. Wei, Y . Liu, Z. Lu, and M. Ding. Mixture of horizons in action chunking.arXiv preprint arXiv:2511.19433, 2025. URLhttps: //arxiv.org/abs/2511.19433. 10

    Pith/arXiv arXiv 2025

  15. [17]

    Y . Liu, J. I. Hamid, A. Xie, Y . Lee, M. Du, and C. Finn. Bidirectional decoding: Improv- ing action chunking via guided test-time sampling. InInternational Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=qZmn2hkuzw

    2025

  16. [18]

    J. So, C. Lee, S. Lee, J. Ok, and E. Park. Improving generative behavior cloning via self- guidance and adaptive chunking.arXiv preprint arXiv:2510.12392, 2025. URLhttps:// arxiv.org/abs/2510.12392

    arXiv 2025

  17. [19]

    R. Gou. Learning temporal action chunking for motor control. Master’s thesis, Univer- sity of British Columbia, 2024. URLhttps://open.library.ubc.ca/soa/cIRcle/ collections/ubctheses/24/items/1.0445184. M.Sc. thesis

    2024

  18. [20]

    Liang, X

    Y . Liang, X. Wang, K. Wang, S. Wang, X. Peng, H. Chen, D. K. H. Chua, and P. Vadakkepat. Adaptive action chunking at inference-time for vision-language-action models. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. URL https://arxiv.org/abs/2604.04161

    Pith/arXiv arXiv 2026

  19. [21]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. LIBERO: Benchmarking knowl- edge transfer for lifelong robot learning. InAdvances in Neural Information Processing Sys- tems Datasets and Benchmarks Track, 2023. URLhttps://arxiv.org/abs/2306.03310

    Pith/arXiv arXiv 2023

  20. [22]

    Y . Mu, T. Chen, Z. Chen, S. Peng, Z. Lan, Z. Gao, Z. Liang, Q. Yu, Y . Zou, M. Xu, L. Lin, Z. Xie, M. Ding, and P. Luo. RoboTwin: Dual-arm robot benchmark with generative digital twins. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2025. URLhttps://arxiv.org/abs/2504.13059

    arXiv 2025

  21. [23]

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard. CALVIN: A benchmark for language- conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022. doi:10.1109/LRA.2022.3180108. URLhttps: //arxiv.org/abs/2112.03227

    work page doi:10.1109/lra.2022.3180108 2022

  22. [24]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840– 6851, 2020. URLhttps://proceedings.neurips.cc/paper/2020/hash/ 4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html

    2020

  23. [25]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score- based generative modeling through stochastic differential equations. InInternational Con- ference on Learning Representations, 2021. URLhttps://openreview.net/forum?id= PxTIG12RRHS

    2021

  24. [26]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, 2023. URLhttps: //openreview.net/forum?id=PqvMRDCJT9t

    2023

  25. [27]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInternational Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2209.03003

    Pith/arXiv arXiv 2023

  26. [28]

    Black, N

    Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.π 0.5: A vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025. URLhttps://arxiv.org/abs/ 2504.16054. 11

    Pith/arXiv arXiv 2025

  27. [29]

    Black, M

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

    Pith/arXiv arXiv 2025

  28. [31]

    Lee and Y .-L

    S.-W. Lee and Y .-L. Kuo. Diff-DAgger: Uncertainty estimation with diffusion policy for robotic manipulation.arXiv preprint arXiv:2410.14868, 2024. URLhttps://arxiv.org/ abs/2410.14868

    arXiv 2024

  29. [32]

    S. Ye, J. Jang, B. Jeon, S. Joo, J. Yang, B. Peng, A. Mandlekar, R. Tan, Y .-W. Chao, B. Y . Lin, et al. Latent action pretraining from videos.arXiv preprint arXiv:2406.05173, 2024. URL https://arxiv.org/abs/2406.05173

    arXiv 2024

  30. [33]

    Q. Zhao, Y . Lu, M. J. Kim, Z. Fu, Z. Zhang, Y . Wu, Z. Li, Q. Ma, S. Han, C. Finn, et al. CoT- VLA: Visual chain-of-thought reasoning for vision-language-action models.arXiv preprint arXiv:2505.15324, 2025. URLhttps://arxiv.org/abs/2505.15324

    arXiv 2025

  31. [35]

    URLhttps://arxiv.org/abs/2506.02146

    arXiv

  32. [36]

    Pertsch, K

    K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine. FAST: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2504.01227, 2025. URLhttps://arxiv.org/abs/2504.01227

    arXiv 2025

  33. [37]

    X. Chen, H. Wei, P. Zhang, C. Zhang, K. Wang, Y . Guo, R. Yang, Y . Wang, X. Xiao, L. Zhao, et al. VILLA-X: Enhancing latent action modeling in vision-language-action models.arXiv preprint arXiv:2507.03867, 2025. URLhttps://arxiv.org/abs/2507.03867

    arXiv 2025

  34. [38]

    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:2508.12268, 2025. URLhttps://arxiv.org/abs/2508.12268

    Pith/arXiv arXiv 2025

  35. [39]

    Q. Bu, Y . Yang, J. Cai, S. Gao, G. Ren, M. Yao, P. Luo, and H. Li. UniVLA: Learning to act anywhere with task-centric latent actions.arXiv preprint arXiv:2503.10474, 2025. URL https://arxiv.org/abs/2503.10474

    arXiv 2025

  36. [40]

    Zhang, J

    K. Zhang, J. Zhang, R. Xu, Y . Sun, S. Xue, Y . Wen, X. Guo, M. Guo, W. Liufu, L. Zi- hou, K. Ji, Y . Zhang, J. Zhu, J. Liu, Z. Li, R. Chen, M. Cao, J. Zhang, S. Zhao, X. Chang, F. Zheng, I. Laptev, and X. Liang. A1: A fully transparent open-source, adaptive and effi- cient truncated vision-language-action model.arXiv preprint arXiv:2604.05672, 2026. URL ...

    Pith/arXiv arXiv 2026

  37. [41]

    Y . Fu, Z. Zhang, Y . Zhang, Z. Wang, Z. Huang, and Y . Luo. MergeVLA: Cross-skill model merging toward a generalist vision-language-action agent. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. URLhttps://arxiv.org/ abs/2511.18810

    arXiv 2026

  38. [42]

    H. Wang, G. Zhang, Y . Yan, R. R. Kompella, and G. Liu. VLA knows its limits.arXiv preprint arXiv:2602.21445, 2026. URLhttps://arxiv.org/abs/2602.21445

    Pith/arXiv arXiv 2026

  39. [43]

    T. A. Driscoll and R. J. Braun.Fundamentals of Numerical Computation. SIAM, 2017

    2017

  40. [44]

    C. Yu, Y . Wang, Z. Guo, H. Lin, S. Xu, H. Zang, Q. Zhang, Y . Wu, C. Zhu, J. Hu, Z. Huang, M. Wei, Y . Xie, K. Yang, B. Dai, Z. Xu, J. Du, X. Wang, X. Fu, L. Shi, Z. Liu, K. Chen, W. Liu, G. Liu, B. Li, J. Yang, Z. Yang, G. Dai, and Y . Wang. RLinf: Flexible and efficient large-scale reinforcement learning via macro-to-micro flow transformation.arXiv pre...

    arXiv 2025

  41. [45]

    StarVLA: A lego-like codebase for vision-language-action model de- veloping.arXiv preprint arXiv:2604.05014, 2026

    StarVLA Community. StarVLA: A lego-like codebase for vision-language-action model de- veloping.arXiv preprint arXiv:2604.05014, 2026. URLhttps://arxiv.org/abs/2604. 05014. 7 Appendix 7.1 Proof of the Tail-Variance Error Bound This appendix expands the derivation summarized in Section 3. We work at one fixed input state sand one fixed future chunk indexk, ...

    Pith/arXiv arXiv 2026