The DOI matches its Crossref record; 60 of the 60 references it cites were found in a registry or an index; 34 of the 47 sources its claims cite were checked for support.
Real-Time Execution of Action Chunking Flow Policies
Open accessAbstract
Modern AI systems, especially those interacting with the physical world, increasingly require real-time performance. However, the high latency of state-of-the-art generalist models, including recent vision-language action models (VLAs), poses a significant challenge. While action chunking has enabled temporal consistency in high-frequency control tasks, it does not fully address the latency problem, leading to pauses or out-of-distribution jerky movements at chunk boundaries. This paper presents a novel inference-time algorithm that enables smooth asynchronous execution of action chunking policies. Our method, real-time chunking (RTC), is applicable to any diffusion- or flow-based VLA out of the box with no re-training. It generates the next action chunk while executing the current one, "freezing" actions guaranteed to execute and "inpainting" the rest. To test RTC, we introduce a new benchmark of 12 highly dynamic tasks in the Kinetix simulator, as well as evaluate 6 challenging real-world bimanual manipulation tasks. Results demonstrate that RTC is fast, performant, and uniquely robust to inference delay, significantly improving task throughput and enabling high success rates in precise tasks $\unicode{x2013}$ such as lighting a match $\unicode{x2013}$ even in the presence of significant latency. See https://pi.website/research/real_time_chunking for videos.
Overview
CoChat Summary
TL;DR
The paper introduces real-time chunking (RTC), an inference-time method that executes action-chunking diffusion/flow policies asynchronously by treating the next chunk as an inpainting problem. Across a 12-task Kinetix benchmark and 6 real-world bimanual tasks, RTC improves robustness to inference delay and task throughput without retraining.
Key findings
- In simulation, RTC is the most robust to inference delay among compared methods, outperforming bidirectional decoding (BID) and temporal ensembling (TE); soft masking improves performance especially at lower delays and execution horizons.
- In real-world experiments with π0.5, RTC achieves the best average task throughput at all inference delays, with statistically significant gains at +100 ms and +200 ms injected latency; synchronous inference degrades linearly, while both TE variants fail at those delays due to high oscillations triggering protective stops.
- RTC is completely robust to injected delay in the real-world setting, showing no degradation, and completes tasks faster than synchronous inference even when inference pauses are removed; in the precision-sensitive light candle task, RTC shows a large final-score advantage.
- Figure 1 reports RTC performs the same robot motion 20% faster than synchronous inference and smoother than all competing methods, including temporal ensembling.
Methods
- RTC is an inference-time algorithm for asynchronous execution of action-chunking policies: it generates the next action chunk while executing the current one, freezes actions guaranteed to execute due to inference delay, and inpaints the rest using a guidance-based flow-matching inpainting procedure with soft masking and guidance weight clipping (β = 5).
- Simulated benchmark: 12 highly dynamic tasks in the Kinetix simulator, with expert policies trained via RPO, action-chunking flow policies with prediction horizon H = 8 and a 4-layer MLP-Mixer, and 2048 rollouts per data point; delays simulated between 0 and 4.
- Real-world evaluation: π0.5 VLA with H = 50, Δt = 20 ms, n = 5 denoising steps, model latency 76 ms for baselines and 97 ms for RTC; 6 challenging bimanual manipulation tasks, 10 trials per task and method for 480 episodes and 28 hours of robot execution time, with injected latency of +100 ms and +200 ms.
Limitations
- The paper states RTC adds significant computational overhead compared to sampling directly from the base policy, is applicable only to diffusion- and flow-based policies, and its real-world experiments do not cover legged locomotion even though it is represented in simulation.
- The paper states that the data used to train π0.5 and the real robot runtime code are not released because they are proprietary; only the simulated benchmark code is provided.
- Evident from the text: real-world results use a single base policy (π0.5) and 10 trials per task/method, so generalization across policies and statistical precision are limited beyond what is reported.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures
![Figure 1 Top: Real-time chunking (RTC) enables the robot to perform highly dexterous and dynamic tasks, such as lighting a match-even in the presence of inference delays in excess of 300 milliseconds, corresponding to more than 30% of the model's prediction horizon. Bottom: RTC performs the same robot motion 20% faster than synchronous inference [5, 30, 8, 24, 31, 59], and smoother than all competing methods, including temporal ensembling [68]. The shown positions, velocities, and accelerations correspond to the shoulder joint of one arm, and are taken from the first 10 seconds of a real autonomous match-lighting rollout.](https://figures.cochat.ai/paper-figures/3cd4ba2056b24cc4b146393d27abcbea/real-time-execution-of-action-chunking-flow-policies-figure-1.png)
![Figure 2 An illustration of a typical bifurcation between consecutive chunks. Inference is started between timesteps 3 and 4. The original chunk that was executing, { a t } (black), had planned to go above the obstacle while the newly generated chunk { a ′ t } (red) goes below the obstacle. However, { a ′ t } is not available until d = 7 steps later. A naive asynchronous algorithm might jump from a 10 to a ′ 11 , inducing a very high, outof-distribution acceleration. Temporal ensembling [68], i.e., interpolating between chunks, reduces the acceleration but produces poor actions.](https://figures.cochat.ai/paper-figures/d3bac9cd3b5f4edc91d2704f2802ba8e/real-time-execution-of-action-chunking-flow-policies-figure-2.png)





![Figure 8 Left: Simulated ablation over different schedules for soft masking weights (Eq. 5). Exponential decay performs the best overall, although linear decay is very close behind. Right: Comparison with the inpainting algorithm from Diffuser [26], which overwrites a portion of the action chunk with the desired actions at each denoising step. While this simpler (and cheaper) inpainting method still provides some benefit, it is outperformed by our guidance-based approach.](https://figures.cochat.ai/paper-figures/536a75e2a3ad4b978b379cb9b7f0de58/real-time-execution-of-action-chunking-flow-policies-figure-8.png)
Verification
Claims the paper makes
14 Valid · 7 Not valid · 4 Not checked · 76 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“Though recent works combining learning methods with MPC have demonstrated real-time control capabilities in narrow domains [53, 21], they rely on explicit, hand-crafted dynamics models and cost functions.”
What the sources say
- Supportedcites [53] [53] Tim Salzmann, Elia Kaufmann, Jon Arrizabalaga, Marco Pavone, Davide Scaramuzza, and Markus Ryll. Real-time neural mpc: Deep learning model predictive control for quadrotors and agile robotic platforms. IEEE Robotics and Automation Letters , 8(4):2397-2404, 2023. - source
The Problem Setup (p. 3) says the MPC dynamics combine first-principles equations with a learned model. The Experimental Setup (pp. 5–6) specifies a nominal rigid-body dynamics model and a quadratic tracking cost. Thus the paper supports reliance on explicitly specified dynamics and cost functions, although its dynamics also include a learned residual. (LLM, conf=0.88)
- Disputedcites [21] [21] Nicklas Hansen, Xiaolong Wang, and Hao Su. Temporal difference learning for model predictive control, 2022. URL https://arxiv.org/abs/2203.04955 . - sourceExperimental — accuracy not yet measured
Section 3 (p. 3) describes TD-MPC as combining MPC with a learned latent dynamics model and learned reward model. The Introduction (p. 1) also says its terminal value function is learned. As an example of the cited recent work, it contradicts the claim that such methods rely on explicit, hand-crafted dynamics models and cost functions. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“While action chunking has enabled temporal consistency in high-frequency control tasks, it does not fully address the latency problem, leading to pauses or out-of-distribution jerky movements at chunk boundaries.”
What the sources say
- Partly supportedcites [68] [68] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705 , 2023. - sourceExperimental — accuracy not yet measured
In “Action Chunking and Temporal Ensemble” (page not given), the paper says chunking helps model temporally correlated behavior in high-frequency trajectories. It also says chunking alone can produce jerky motion at chunk boundaries and credits temporal ensembling with improving smoothness, so it does not establish that chunking by itself enables temporal consistency. (LLM, conf=0.84)
- Could not verifycites [33] [33] Lucy Lai, Ann Zixiang Huang, and Samuel J Gershman. Action chunking as policy compression. 2022. - sourceExperimental — accuracy not yet measured
The Introduction and §4.1 discuss action chunking in human sequence-learning tasks, including faster and more accurate responses to patterned stimuli. Our copy does not address temporal consistency in high-frequency control tasks; searches for that phrase and for high-frequency control, latency, and chunk-boundary movements found no supporting passage. (LLM, conf=0.76)
- Supportedcites [11] [11] 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 , page 02783649241273668, 2023. - source
Section 5.3 (p. 8) reports that an action horizon greater than one helps the policy predict consistent actions; Section 4.3 (p. 5) explains how sequence prediction avoids jitter between consecutive actions. The paper applies this approach to real-time robot control, with learned policies running at 10 Hz (D.1, p. 18). (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Kim et al. [31], who optimize the 7B OpenVLA model [30] specifically for inference speed, achieve no better than 321ms of latency on a server-grade A100 GPU.”
What the sources say
- Partly supportedcites [31] [31] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645 , 2025. - sourceExperimental — accuracy not yet measured
Table III (ALOHA Inference Efficiency Comparisons, p. 10) reports 0.321 seconds on an NVIDIA A100 for OpenVLA-OFT+, an optimized OpenVLA variant, in a bimanual task with three input images. The claim’s “no better than” wording extends that task-specific result to the model generally: Related Work (p. 2) reports lower latency for a single-arm task with one input image. (LLM, conf=0.88)
- Partly supportedcites [30] [30] 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. - sourceExperimental — accuracy not yet measured
Section 5.4 (p. 10) identifies OpenVLA as a 7B-parameter model and discusses inference speed on server-grade GPUs. But the text in our copy does not give a 321 ms A100 latency or establish it as a best achievable latency; the Figure 6 data are not present in the text. Section 3.5 (p. 6) describes its reported inference speed as measured without speed-up techniques, rather than for a model specifically optimized for inference speed. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“we devise a new benchmark based on the Kinetix simulator [43] consisting of 12 highly dynamic manipulation and locomotion tasks.”
What the sources say
- Supportedcites [43] [43] Michael Matthews, Michael Beukman, Chris Lu, and Jakob Foerster. Kinetix: Investigating the training of general agents through open-ended physics-based control tasks. arXiv preprint arXiv:2410.23208 , 2024. - source
The citation identifies Kinetix as the simulator on which the citing authors base their new benchmark; it does not attribute the 12 tasks to the cited paper. The Introduction describes Kinetix as a framework for 2D physics-based environments, and Section 3.1 identifies its underlying physics engine. (LLM, conf=0.91)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Inspecting the per-task results (Figure 5, top), we can conclude that RTC helps with more than just execution speed: it completes tasks faster than synchronous inference even when inference pauses are removed .”
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
42 more claims
Create a free account to see all 101 claims and to verify the 9 references still open.
- Probably not a substantial claim. 33
- It may state two separate claims as one. 25
- The quote only partly states this claim. 2
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Baselines. We compare to the following baselines: · p. 8
“Inspecting the per-task results (Figure 5, top)”
§ Baselines. We compare to the following baselines: · p. 8
“Figure 6: Top: Controller steps (equivalent to elapsed time with inference pauses removed multiplied by 50Hz) vs. cumulative progress for each task, aggregated across all delays.”
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 6
- [2][2] Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey. Journal of machine learning research , 18(153):1-43, 2018. -
- [8][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: Vision-languageaction models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 , 2023. - · arXiv
- [14][14] Donald R. Coughanowr and Steven E. LeBlanc. Process Systems Analysis and Control , chapter 18. McGraw-Hill, New York, 3rd edition, 2009. ISBN 978-0073397894. -
- [22][22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems , 33:6840-6851, 2020. -
- [44][44] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Selfsupervised skill abstractions for learning continuous control, 2024. URL https://arxiv. org/abs/2407.15840 . -
- [55][55] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations , 2023. -
Cited by
Causal World Modeling for Robot Control
Li, L., Zhang, Q., Luo, Y., Yang, S., Wang, R., Han, F., Yu, M., Gao, Z., Xue, N., Zhou, B., et al. (2026) · Robotics: Science and Systems · doi:10.15607/rss.2026.xxii.016 · Cited by 434
$π_{0.7}$: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesPreprint
Intelligence, P., Ai, B., Amin, A., Aniceto, R. J., Balakrishna, A., Balke, G., Black, K., Bokinsky, G., Cao, S., Charbonnier, T., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.15483 · Cited by 184
Self-citationVision-Language-Action Models for Robotics: A Review Towards Real-World Applications
Kawaharazuka, K., Oh, J., Yamada, J., Posner, I., & Zhu, Y. (2025) · IEEE Access · doi:10.1109/access.2025.3609980 · Cited by 250
A Survey on Vision-Language-Action Models: An Action Tokenization PerspectivePreprint
Zhong, Y., Bai, F., Cai, S., Huang, X., Chen, Z., Zhang, X., Wang, Y., Guo, S., Guan, T., Lui, K. N., et al. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.01925 · Cited by 142
MEM: Multi-Scale Embodied Memory for Vision Language Action ModelsPreprint
Torne, M., Pertsch, K., Walke, H., Vedder, K., Nair, S., Ichter, B., Ren, A. Z., Wang, H., Tang, J., Stachowicz, K., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.03596 · Cited by 145
Self-citation
Metadata
- Authors
- Black, K., Galliker, M., & Levine, S.
- Affiliations
- Physical Intelligence
- University of California, Berkeley (US)
- ORCID
- Levine, Sergey
- Venue
- Advances in Neural Information Processing Systems 38
- Published
- 2025
- Pages
- 37596-37620
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W4417097926
- Cited by
- 12
- References
- 0
- Citations by year
- 2026: 12
- FWCI
- 26.12
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Keywords
- action chunking, real-time execution, vision-language-action models, flow-based generative models, diffusion models, inference latency, bimanual manipulation, temporal coherence
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Real-Time Execution of Action Chunking Flow Policies. https://app.cochat.ai/papers/10.52202/085713-1122
- The paper
- Black, K., Galliker, M., & Levine, S. (2025). Real-Time Execution of Action Chunking Flow Policies. Advances in Neural Information Processing Systems 38. https://doi.org/10.52202/085713-1122
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.