pith. machine review for the scientific record. sign in

arxiv: 2605.11381 · v1 · submitted 2026-05-12 · 💻 cs.RO · cs.DC

Recognition: no theorem link

Kairos: A Scalable Serving System for Physical AI

Yinwei Dai , Ganesh Ananthanarayanan , Landon Cox , Xenofon Foukas , Bozidar Radunovic , Ravi Netravali
Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:26 UTC · model grok-4.3

classification 💻 cs.RO cs.DC
keywords physical AIrobot servinggenerate-execute loopmulti-round inferenceend-to-end latencyscalable servingrobot fleets
0
0 comments X

The pith

Kairos cuts physical AI task latency by 32 to 66 percent by treating the generate-execute loop as first-class and staying active during robot execution.

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

Physical AI tasks require models to perform multiple rounds of inference, each producing a chunk of actions that the robot executes, with inference and execution interleaved asynchronously. Existing serving systems built for digital AI step away after inference and therefore leave these loops inefficient, especially when many robots run in parallel. Kairos redesigns the serving layer to remain involved through the execution phase rather than handing off control. This produces measured reductions in average end-to-end task latency of 31.8 to 66.5 percent across different models and robots. The size of the improvement grows as the robot fleet expands, which directly affects whether large physical AI models can be deployed at scale.

Core claim

The paper claims that physical AI inference consists of repeated rounds that generate action chunks and run asynchronously with execution, a structure that existing digital serving systems do not handle efficiently. By making the generate-execute loop a first-class citizen and keeping the serving system active during execution, Kairos reduces average end-to-end task latency by 31.8 to 66.5 percent relative to state-of-the-art digital practices, and these gains increase with fleet size.

What carries the argument

The generate-execute loop made first-class with active serving-system involvement during the execution phase, which coordinates multiple inference rounds, chunked actions, and asynchronous interleaving.

If this is right

  • Average end-to-end task latency falls by 31.8 to 66.5 percent for physical AI workloads.
  • The latency reduction grows larger as the number of robots in the fleet increases.
  • The approach works across a range of physical AI models and robot platforms.
  • It removes a key obstacle to deploying large models on sizable robot fleets.

Where Pith is reading between the lines

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

  • Serving designs that stay active during physical execution could be applied to other interleaved compute-and-act systems such as autonomous vehicles or industrial control.
  • Evaluation benchmarks for AI serving may need new multi-round, chunked-action test cases to reflect physical deployment realities.
  • Robot hardware and communication layers might be optimized specifically to reduce overhead in asynchronous interleaving, amplifying the latency gains.

Load-bearing premise

The multi-round inference, chunked actions, and asynchronous interleaving of physical AI are the dominant sources of latency, and keeping the serving system involved during execution reduces them without creating new bottlenecks or correctness problems.

What would settle it

Run the same physical AI task on a fleet of robots using both Kairos and a standard digital AI serving system, then measure whether the end-to-end completion time drops by 31.8 to 66.5 percent and whether the gap widens as the number of robots increases.

Figures

Figures reproduced from arXiv: 2605.11381 by Bozidar Radunovic, Ganesh Ananthanarayanan, Landon Cox, Ravi Netravali, Xenofon Foukas, Yinwei Dai.

Figure 1
Figure 1. Figure 1: The state of the art physical AI models. arms, mobile bases, and full-scale humanoids [1–3], phys￾ical AI is poised to shape many domains, including ware￾houses [9], factories [51], homes [1, 2], and healthcare [35]. 2.1 Physical AI Models Foundation models for physical AI map observational inputs to actions. Example observational inputs include camera im￾ages, robot states, and task instructions. Actions,… view at source ↗
Figure 3
Figure 3. Figure 3: The best per-task execution horizon varies widely across tasks and workloads. Each curve shows the CDF of the best per-task 𝐻 for a workload. 1.0 1.5 2.0 2.5 3.0 Optimal Per-Round Horizon (Normalized) 0.0 0.5 1.0 CDF LIBERO / SmolVLA LIBERO / Pi0.5 Bimanual / Pi0.5 Best static horizon [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The best per-task execution horizon varies within a task. CDF of the optimal per-round optimal execution hori￾zon, normalized by the best per-task horizon. how many actions are executed from each chunk. The set of actions in the chunk beyond the first 𝐻 are discarded.2 The execution horizon is a key parameter that allows trad￾ing off accuracy with resource demand. A shorter value of 𝐻 generates actions mor… view at source ↗
Figure 5
Figure 5. Figure 5: We use four physical AI tasks to illustrate that execution-unaware scheduling leads to suboptimal decisions. Task 2 (green) has longer total generation time but shorter end-to-end latency. FIFO ignores task-level information and only prioritizes request-level wait time, while Autellix [37] misclassifies it as a long-running task and prioritizes it. Execution-aware scheduling correctly prioritizes Task 2, r… view at source ↗
Figure 6
Figure 6. Figure 6: System architecture. its respective execution horizons, keeping with the observa￾tions in §2.2. Focusing on Task 2’s latency (green), we can see the benefit of execution awareness. FIFO and fairness schedulers prioritize tasks with longer execution times ahead of Task 2, do not focus on the end-to-end task latency, and end up starving shorter tasks. Task 2’s latency is 7 units with an execution-aware sched… view at source ↗
Figure 7
Figure 7. Figure 7: Diffusion confidence and dynamic horizon selec￾tion. Box sizes show per-action update magnitudes at each step. The horizon policy scans from 𝐴1, comparing each ac￾tion’s final update magnitude against (1+𝑡) times its mean over earlier steps. Here 𝐴5 exceeds the threshold, setting 𝐻=4 and discarding 𝐴5–𝐴6. that maps the intermediate diffusion updates or other inter￾mediate inference states to the execution … view at source ↗
Figure 9
Figure 9. Figure 9: Within-bucket ordering by estimated execution latency. Left: FIFO ordering can leave long-execution tasks to the end. Right: longest execution first reduces the tail latency while keeping the average latency. 0 50 100 150 200 Relative Difference (%) 0.0 0.5 1.0 CDF LIBERO / SmolVLA LIBERO / Pi0.5 Bimanual / Pi0.5 [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: CDF of the relative difference between consecutive-round execution latencies. Across workloads, 52.3–85.4% of rounds have less than 10% deviation. by estimated execution latency (descending). The reason is that scheduling execution-dominant tasks first prevents long execution phases from starving, which can reduce tail latency while maintaining the average latency ( [PITH_FULL_IMAGE:figures/full_fig_p008… view at source ↗
Figure 11
Figure 11. Figure 11: Accuracy–efficiency trade-off across six work￾loads ( [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Average end-to-end latency under increasing task arrival rates across workloads. Error bars show P25 and P95. 2.5 5.0 7.5 Rate (task/s) 0 50 Latency (s) LIBERO / SmolVLA 2 4 Rate (task/s) 0 50 100 MetaWorld / SmolVLA 2.5 5.0 7.5 Rate (task/s) 0 50 LIBERO / XVLA 2.5 5.0 7.5 Rate (task/s) 0 50 100 Isaac / GR00T N1.5 2.5 5.0 7.5 Rate (task/s) 0 100 LIBERO / Pi0.5 1 2 Rate (task/s) 0 100 Bimanual / Pi0.5 2.5 … view at source ↗
Figure 13
Figure 13. Figure 13: Effect of edge–cloud offloading on average end-to-end latency for increasing arrival rates. Error bars show P25-90. dynamic horizon (up to 26.9%) in the majority of workloads. This validates that Kairos’s execution-aware scheduler, with visibility into both generation and execution phases, cor￾rectly identifies and prioritizes the truly short requests— without coming at the expense of longer tasks. Finall… view at source ↗
Figure 14
Figure 14. Figure 14: Average end-to-end latency as the dedicated robot fleet scales from 10 to 100 concurrent robots. 1 5 10 20 ∞ Bucket Number 40 60 80 Avg Latency (s) LIBERO / Pi0.5 MetaWorld / SmolVLA LIBERO / SmolVLA [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Impact of wait ra￾tio bucket count on average latency. LIBERO / Pi0.5 MetaWorld / SmolVLA LIBERO / SmolVLA 0 50 100 Avg Latency (s) Longer Exec. First FIFO [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
read the original abstract

Physical AI is experiencing rapid growth with frontier foundation models increasing its capabilities across general environments. Physical AI tasks are characterized by inference properties that are markedly different from digital AI. They consist of multiple rounds of inference and action execution, generating a chunk of actions in each inference round, and asynchronously interleaving inference and execution. This makes existing digital AI serving systems unsuited for physical AI; a shortcoming that is critical for enabling their wide adoption, considering their size and the scale of the robot fleets they have to serve. To fill this gap, we design Kairos, the first multi-robot serving system that makes the generate-execute loop a first-class citizen, with active involvement in the execution phase. Across a wide range of physical AI models and robots, Kairos reduces the average end-to-end task latency by 31.8--66.5% over state-of-the-art digital AI serving practices, with gains scaling with the robot fleet size.

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 introduces Kairos, the first multi-robot serving system for physical AI that treats the generate-execute loop as a first-class citizen by actively participating in the execution phase. It identifies key differences in physical AI inference properties (multiple rounds, chunked actions, asynchronous interleaving) that make digital AI serving systems unsuitable, and claims that Kairos reduces average end-to-end task latency by 31.8-66.5% over state-of-the-art digital AI serving practices across a range of models and robots, with gains that scale with robot fleet size.

Significance. If the empirical results hold under scrutiny, this work addresses a timely and practically important gap in systems support for physical AI deployments. The focus on execution-phase involvement and fleet-size scaling provides a concrete path toward efficient operation of large foundation models on robot fleets, which could accelerate real-world adoption. The empirical nature of the central claim (latency measurements rather than derivations) is appropriately matched to the problem.

major comments (2)
  1. [§5] §5 (Evaluation): the latency reduction numbers (31.8-66.5%) are presented without an accompanying table or figure that breaks down the contribution of each Kairos mechanism (e.g., async interleaving vs. chunked action handling) versus baseline overheads; this makes it difficult to confirm that the gains are attributable to the claimed design rather than unstated experimental choices.
  2. [§4] §4 (System Design): the description of active execution-phase involvement does not quantify the additional communication or synchronization overhead introduced by Kairos itself; without this, it is unclear whether the net latency improvement remains positive under higher robot counts or network variability.
minor comments (2)
  1. The abstract would benefit from a one-sentence summary of the experimental scope (number of models/robots, fleet sizes tested) to give readers immediate context for the reported percentages.
  2. Figure captions and axis labels in the scaling plots should explicitly state whether error bars represent standard deviation across runs or across robot instances.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of Kairos and for the constructive major comments. We address each point below and have revised the manuscript to incorporate additional analysis and clarifications as suggested.

read point-by-point responses

  1. Referee: [§5] §5 (Evaluation): the latency reduction numbers (31.8-66.5%) are presented without an accompanying table or figure that breaks down the contribution of each Kairos mechanism (e.g., async interleaving vs. chunked action handling) versus baseline overheads; this makes it difficult to confirm that the gains are attributable to the claimed design rather than unstated experimental choices.

    Authors: We appreciate the referee's request for greater attribution of the reported gains. While the original evaluation focused on end-to-end results across models and robots, we agree that an explicit breakdown strengthens the claims. In the revised manuscript, we have added a new table and accompanying figure in §5 that decomposes the latency reductions into contributions from async interleaving, chunked action handling, and other mechanisms, while isolating baseline overheads from digital AI serving systems. The breakdown shows that async interleaving and active execution-phase participation account for the majority of the improvement (typically 45-60% of the total reduction), with the remainder from chunked handling; this confirms the gains are attributable to Kairos's design rather than experimental artifacts. We have also expanded the experimental setup description for clarity. revision: yes

  2. Referee: [§4] §4 (System Design): the description of active execution-phase involvement does not quantify the additional communication or synchronization overhead introduced by Kairos itself; without this, it is unclear whether the net latency improvement remains positive under higher robot counts or network variability.

    Authors: We thank the referee for noting this omission. To address it directly, the revised §4 now includes a dedicated overhead analysis subsection with empirical measurements of the communication and synchronization costs introduced by Kairos's active execution-phase involvement. These overheads were quantified under varying fleet sizes (up to 64 robots) and network conditions (including simulated variability). The results demonstrate that the added overhead remains low (typically 3-7% of end-to-end latency) and scales sublinearly, preserving net latency reductions of 28-62% even at higher robot counts and with 15-25% network jitter. This supports the scalability of the design. revision: yes

Circularity Check

0 steps flagged

No significant circularity in empirical systems evaluation

full rationale

The paper describes a serving system for physical AI and supports its latency-reduction claims through direct experimental measurements across multiple models, robots, and fleet sizes. No equations, derivations, fitted parameters, or predictions are presented that could reduce to inputs by construction. The argument relies on empirical results rather than self-citations, uniqueness theorems, or ansatzes, making the evaluation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review provides no visibility into implementation parameters, hyperparameters, or design constants; the core premise rests on a domain characterization of physical AI tasks.

axioms (1)
  • domain assumption Physical AI tasks consist of multiple rounds of inference and action execution that asynchronously interleave inference and execution.
    Stated directly in the abstract as the distinguishing property that makes existing systems unsuited.

pith-pipeline@v0.9.0 · 5479 in / 1227 out tokens · 47244 ms · 2026-05-13T02:26:46.855048+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

67 extracted references · 67 canonical work pages · 2 internal anchors

  1. [1]

    1X Technologies.https://www.1x.tech, 2024

    work page 2024

  2. [2]

    Figure AI.https://www.figure.ai, 2024

    work page 2024

  3. [3]

    Unitree Robotics.https://www.unitree.com, 2024

    work page 2024

  4. [4]

    Fourier gr-1.https://www.fftai.com/products-gr1, 2026

    work page 2026

  5. [5]

    Nvidia isaac lab.https://developer.nvidia.com/isaac/lab, 2026

    work page 2026

  6. [6]

    Nvidia isaac sim.https://developer.nvidia.com/isaac/sim, 2026

    work page 2026

  7. [7]

    So-101.https://huggingface.co/docs/lerobot/en/so101, 2026

    work page 2026

  8. [8]

    Abhyankar, Z

    R. Abhyankar, Z. He, V. Srivatsa, H. Zhang, and Y. Zhang. Infercept: Efficient intercept support for augmented large language model infer- ence, 2024

    work page 2024

  9. [9]

    Amazon robotics.https://www.aboutamazon.com/news/ operations/amazon-robotics-robots-fulfillment-center, 2024

    Amazon. Amazon robotics.https://www.aboutamazon.com/news/ operations/amazon-robotics-robots-fulfillment-center, 2024

    work page 2024

  10. [10]

    The Claude 3 model family: Opus, Sonnet, Haiku

    Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku. Technical report, 2024.https : / / www - cdn . anthropic . com / de8ba9b01c9ab7cbabf5c33b80b7bbc618857627 / Model _ Card _ Claude_3.pdf

    work page 2024

  11. [11]

    Black, M

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

    work page 2025

  12. [12]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, X. Chen, K. Choroman- ski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. G. Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y. Kuang, I. Leal, L. Lee, T.-W. E. Lee, S. Levine, Y. Lu, H. Michalewski, I. Mordatch, K. Per...

    work page 2023

  13. [13]

    Cadene, S

    R. Cadene, S. Aliberts, F. Capuano, M. Aractingi, A. Zouitine, P. Kooij- mans, J. Choghari, M. Russi, C. Pascal, S. Palma, M. Shukor, J. Moss, A. Soare, D. Aubakirova, Q. Lhoest, Q. Gallouédec, and T. Wolf. Lerobot: An open-source library for end-to-end robot learning, 2026

    work page 2026

  14. [14]

    T. Chen, Z. Chen, B. Chen, Z. Cai, Y. Liu, Z. Li, Q. Liang, X. Lin, Y. Ge, Z. Gu, W. Deng, Y. Guo, T. Nian, X. Xie, Q. Chen, K. Su, T. Xu, G. Liu, M. Hu, H. ang Gao, K. Wang, Z. Liang, Y. Qin, X. Yang, P. Luo, and Y. Mu. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation, 2025

    work page 2025

  15. [15]

    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, 2024

    work page 2024

  16. [16]

    Corporation

    N. Corporation. Tensorrt-llm: An open-source library to accelerate inference of large language models on nvidia gpus.https://github. com/NVIDIA/TensorRT-LLM, 2023. Library: TensorRT-LLM

    work page 2023

  17. [17]

    H. Fang, Y. Liu, Y. Du, L. Du, and H. Yang. Sqap-vla: A synergistic quantization-aware pruning framework for high-performance vision- language-action models, 2025

    work page 2025

  18. [18]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team, Google. Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    I. Gim, S. seob Lee, and L. Zhong. Asynchronous llm function calling, 2024

    work page 2024

  20. [20]

    J. Gu, M. Chowdhury, K. G. Shin, Y. Zhu, M. Jeon, J. Qian, H. Liu, and C. Guo. Tiresias: a gpu cluster manager for distributed deep learning. InProceedings of the 16th USENIX Conference on Networked Systems Design and Implementation, NSDI’19, page 485–500, USA, 2019. USENIX Association

    work page 2019

  21. [21]

    P. B. Hansen.Operating system principles. Prentice-Hall, Inc., USA, 1973

    work page 1973

  22. [22]

    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. Vu...

    work page 2025

  23. [23]

    Jiang, X

    T. Jiang, X. Jiang, Y. Ma, X. Wen, B. Li, K. Zhan, P. Jia, Y. Liu, S. Sun, and X. Lang. The better you learn, the smarter you prune: Towards efficient vision-language-action models via differentiable token pruning, 2025

    work page 2025

  24. [24]

    Jiang, J

    W. Jiang, J. Clemons, K. Sankaralingam, and C. Kozyrakis. How fast can i run my vla? demystifying vla inference performance with vla-perf, 2026

    work page 2026

  25. [25]

    M. J. Kim, Y. Gao, T.-Y. Lin, Y.-C. Lin, Y. Ge, G. Lam, P. Liang, S. Song, M.-Y. Liu, C. Finn, and J. Gu. Cosmos policy: Fine-tuning video models for visuomotor control and planning, 2026

    work page 2026

  26. [26]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burch- fiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. Openvla: An open-source vision-language-action model, 2024

    work page 2024

  27. [27]

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    work page 2023

  28. [28]

    S.-W. Lee, X. Kang, and Y.-L. Kuo. Diff-dagger: Uncertainty estimation with diffusion policy for robotic manipulation, 2025

    work page 2025

  29. [29]

    H. Li, Q. Mang, R. He, Q. Zhang, H. Mao, X. Chen, H. Zhou, A. Cheung, J. Gonzalez, and I. Stoica. Continuum: Efficient and robust multi-turn llm agent scheduling with kv cache time-to-live, 2026

    work page 2026

  30. [30]

    P. Li, Y. Zhou, D. Muhtar, L. Yin, S. Yan, L. Shen, S. Vosoughi, and S. Liu. Diffusion language models know the answer before decoding, 2026

    work page 2026

  31. [31]

    S. Li, Y. Gao, D. Sadigh, and S. Song. Unified video action model, 2025

    work page 2025

  32. [32]

    X. Li, K. Hsu, J. Gu, K. Pertsch, O. Mees, H. R. Walke, C. Fu, I. Lunawat, I. Sieh, S. Kirmani, S. Levine, J. Wu, C. Finn, H. Su, Q. Vuong, and T. Xiao. Evaluating real-world robot manipulation policies in simulation, 2024

    work page 2024

  33. [33]

    Lipman, R

    Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling, 2023. 13

    work page 2023

  34. [34]

    B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning, 2023

    work page 2023

  35. [35]

    Liu et al

    Y. Liu et al. A survey of embodied AI in healthcare: Techniques, applications, and opportunities.arXiv preprint arXiv:2501.07468, 2025

    work page arXiv 2025

  36. [36]

    Z. Liu, Y. Chen, H. Cai, T. Lin, S. Yang, Z. Liu, and B. Zhao. Vla- pruner: Temporal-aware dual-level visual token pruning for efficient vision-language-action inference, 2026

    work page 2026

  37. [37]

    M. Luo, X. Shi, C. Cai, T. Zhang, J. Wong, Y. Wang, C. Wang, Y. Huang, Z. Chen, J. E. Gonzalez, and I. Stoica. Autellix: An efficient serving engine for llm agents as general programs, 2025

    work page 2025

  38. [38]

    T. Ma, J. Zheng, Z. Wang, C. Jiang, A. Cui, J. Liang, and S. Yang. Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control, 2026

    work page 2026

  39. [39]

    Nuyens and A

    M. Nuyens and A. Wierman. The foreground–background queue: A survey.Performance Evaluation, 65(3):286–307, 2008

    work page 2008

  40. [40]

    Bjorck, F

    NVIDIA, :, J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. J. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. Llontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y. L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y. Xie, Y. Xu, Z. Xu, S. Ye, Z. Yu, ...

    work page 2025

  41. [41]

    GPT-4 Technical Report

    OpenAI. GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    J. Pai, L. Achenbach, V. Montesinos, B. Forrai, O. Mees, and E. Nava. mimic-video: Video-action models for generalizable robot control be- yond vlas, 2025

    work page 2025

  43. [43]

    S. G. Patil, T. Zhang, X. Wang, and J. E. Gonzalez. Gorilla: Large language model connected with massive apis, 2023

    work page 2023

  44. [44]

    Pohland, X

    S. Pohland, X. Foukas, G. Ananthanarayanan, A. Kolobov, S. Mehro- tra, B. Radunovic, and A. Verma. Offload or overload: A platform measurement study of mobile robotic manipulation workloads, 2026

    work page 2026

  45. [45]

    Prasad, K

    A. Prasad, K. Lin, J. Wu, L. Zhou, and J. Bohg. Consistency policy: Accelerated visuomotor policies via consistency distillation, 2024

    work page 2024

  46. [46]

    Sawyer robot.https://robotsguide.com/robots/sawyer, 2026

    Robots Guide. Sawyer robot.https://robotsguide.com/robots/sawyer, 2026

    work page 2026

  47. [47]

    H. Shi, B. Xie, Y. Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, and G. Huang. Memoryvla: Perceptual-cognitive memory in vision- language-action models for robotic manipulation, 2026

    work page 2026

  48. [48]

    Shukor, D

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zoui- tine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, S. Alibert, M. Cord, T. Wolf, and R. Cadene. Smolvla: A vision-language-action model for affordable and efficient robotics, 2025

    work page 2025

  49. [49]

    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, 2025

    work page 2025

  50. [50]

    G. R. Team, S. Abeyruwan, J. Ainslie, J.-B. Alayrac, M. G. Arenas, T. Armstrong, A. Balakrishna, R. Baruch, M. Bauza, M. Blokzijl, S. Bo- hez, K. Bousmalis, A. Brohan, T. Buschmann, A. Byravan, S. Cabi, K. Caluwaerts, F. Casarini, O. Chang, J. E. Chen, X. Chen, H.-T. L. Chi- ang, K. Choromanski, D. D’Ambrosio, S. Dasari, T. Davchev, C. Devin, N. D. Palo, ...

    work page 2025

  51. [51]

    Tesla optimus.https://www.tesla.com/optimus, 2024

    Tesla. Tesla optimus.https://www.tesla.com/optimus, 2024

    work page 2024

  52. [52]

    Z. Wang, Z. Li, A. Mandlekar, Z. Xu, J. Fan, Y. Narang, L. Fan, Y. Zhu, Y. Balaji, M. Zhou, M.-Y. Liu, and Y. Zeng. One-step diffusion policy: Fast visuomotor policies via diffusion distillation, 2024

    work page 2024

  53. [53]

    Y. Xu, X. Kong, T. Chen, and D. Zhuo. Conveyor: Efficient tool-aware llm serving with tool partial execution, 2024

    work page 2024

  54. [54]

    Y. Xu, Y. Yang, Z. Fan, Y. Liu, Y. Li, B. Li, and Z. Zhang. Qvla: Not all channels are equal in vision-language-action model’s quantization, 2026

    work page 2026

  55. [55]

    H. Yan, Z. Zhong, J. Zhu, J. He, W. Yuan, W. Song, X. Gong, Y. Cai, G. Zhao, X. Yan, B. Liu, Y.-C. Chen, and H. Li. S-vam: Shortcut video- action model by self-distilling geometric and semantic foresight, 2026

    work page 2026

  56. [56]

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. R. Narasimhan, and O. Press. SWE-agent: Agent-computer interfaces enable automated software engineering. InNeurIPS, 2024

    work page 2024

  57. [57]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. React: Synergizing reasoning and acting in language models, 2023

    work page 2023

  58. [58]

    A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, H. Li, J. Li, J. Lv, J. Liu, M. Cao, P. Li, Q. Deng, W. Mei, X. Wang, X. Chen, X. Zhou, Y. Wang, Y. Chang, Y. Li, Y. Zhou, Y. Ye, Z. Liu, and Z. Zhu. Gigaworld-policy: An efficient action-centered world–action model, 2026

    work page 2026

  59. [59]

    H. Ye, J. Yuan, R. Xia, X. Yan, T. Chen, J. Yan, B. Shi, and B. Zhang. Training-free adaptive diffusion with bounded difference approxima- tion strategy, 2024

    work page 2024

  60. [60]

    S. Ye, Y. Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y. L. Tan, C. Zhu, J. Xiang, A. Malik, K. Lee, W. Liang, N. Ranawaka, J. Gu, Y. Xu, G. Wang, F. Hu, A. Narayan, J. Bjorck, J. Wang, G. Kim, D. Niu, R. Zheng, Y. Xie, J. Wu, Q. Wang, R. Julian, D. Xu, Y. Du, Y. Chebotar, S. Reed, J. Kautz, Y. Zhu, L. J. Fan, and J. Jang. World action models are...

    work page 2026

  61. [61]

    T. Yu, D. Quillen, Z. He, R. Julian, A. Narayan, H. Shively, A. Bellathur, K. Hausman, C. Finn, and S. Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning, 2021

    work page 2021

  62. [62]

    T. Yuan, Z. Dong, Y. Liu, and H. Zhao. Fast-wam: Do world action models need test-time future imagination?, 2026

    work page 2026

  63. [63]

    Zhang, M

    K. Zhang, M. Sharma, J. Liang, and O. Kroemer. A modular robotic arm control stack for research: Franka-interface and frankapy, 2020

    work page 2020

  64. [64]

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

    work page 2023

  65. [65]

    Zheng, J

    J. Zheng, J. Li, Z. Wang, D. Liu, X. Kang, Y. Feng, Y. Zheng, J. Zou, Y. Chen, J. Zeng, Y.-Q. Zhang, J. Pang, J. Liu, T. Wang, and X. Zhan. X-vla: Soft-prompted transformer as scalable cross-embodiment vision- language-action model, 2025

    work page 2025

  66. [66]

    Zheng, L

    L. Zheng, L. Yin, Z. Xie, C. Sun, J. Huang, C. H. Yu, S. Cao, C. Kozyrakis, I. Stoica, J. E. Gonzalez, C. Barrett, and Y. Sheng. Sglang: Efficient execution of structured language model programs, 2024

    work page 2024

  67. [67]

    S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y. Bisk, D. Fried, U. Alon, and G. Neubig. WebArena: A realistic web environment for building autonomous agents. InICLR, 2024. 14

    work page 2024