Pith. sign in

REVIEW 2 major objections 5 minor 32 references

SEAM steers action-chunked robot policies at inference so consecutive chunks stay continuous, cutting boundary jerk about 28% without retraining or backprop.

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

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

T0 review · grok-4.5

2026-07-11 16:28 UTC pith:IRIGL66Z

load-bearing objection A clean, cheap inference-time fix for chunk-boundary jerk in flow VLAs; the smoothness gains look real, the “preserves success” claim is under-specified but not obviously false. the 2 major comments →

arxiv 2607.04609 v1 pith:IRIGL66Z submitted 2026-07-06 cs.RO

SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies

Dijia Zhan , Xuemiao Xu , Jinyi Li , Jie Tang This is my paper
classification cs.RO
keywords Vision-Language-Actionaction chunkingflow matchinginference-time steeringcross-chunk consistencymultimodal bifurcationrobot manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Vision-language-action policies that predict fixed-length action chunks often break at chunk boundaries: each chunk is drawn from independent noise, so neighboring chunks can pick incompatible local strategies and produce sharp reversals or gaps. SEAM is a training-free fix for flow-matching policies that uses a simple fact of how robots actually run these models—only a prefix is executed, so the previous chunk’s unused tail is already sitting there as a free consistency reference. Its core step, Velocity-guided Loss Steering, builds a time-dependent target from that tail and applies a closed-form correction after each Euler step, without backpropagating through the policy. On LIBERO-10 with π0.5 this cuts boundary jerk by about 28% and chunk-transition discontinuity by about 27%, keeps task success at baseline level, and holds denoising-loop cost near the unguided policy. A reader who cares about reliable contact-rich manipulation gets smoother motion without the latency of gradient guidance, the cost of rejection sampling, or a full retrain.

Core claim

In synchronous action-chunked flow-matching VLAs, cross-chunk multimodal bifurcation can be reduced at inference by treating the previous chunk’s unexecuted tail as an aligned prior and applying Velocity-guided Loss Steering: after each reverse Euler step, a closed-form correction nudges only the guided overlap window toward the time-interpolated target r_i = (1−t_i)·a_al without policy-network backpropagation. On LIBERO-10 with π0.5 this yields roughly 28% lower boundary jerk and 27% lower chunk discontinuity while preserving baseline task success at about 1.01× denoising-loop cost.

What carries the argument

Velocity-guided Loss Steering (VLS): a closed-form, post-Euler correction that steers the guided overlap prefix of the reverse flow-matching ODE toward a time-dependent target built from the previous chunk’s unexecuted tail, scaled by λ(1−t), without backpropagating through the policy.

Load-bearing premise

The previous chunk’s unused tail is treated as a trustworthy guide for the next chunk’s overlap, so lightly steering toward it smooths motion without locking the robot into a wrong plan.

What would settle it

Re-run the LIBERO-10 evaluation with π0.5 under the paper’s protocol: if SEAM does not clearly cut boundary jerk and chunk discontinuity while holding success near the unguided baseline and denoising cost near 1×—or if success collapses when the prior points at a bad mode—the central claim fails.

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

If this is right

  • Action-chunked flow VLAs can smooth chunk boundaries without retraining, rejection sampling, or policy backprop.
  • Denoising-loop cost can stay near the unguided baseline while boundary jerk and discontinuity fall by roughly a quarter.
  • Weak closed-form overlap guidance can preserve task success better than aggressive temporal averaging of incompatible modes.
  • Under synchronous execution, the unexecuted tail is a free analytic consistency reference that does not require a learned continuation model.
  • Guided window length and strength act as explicit knobs trading smoothness against policy freedom on a fixed trained model.

Where Pith is reading between the lines

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

  • The same free-overlap idea may apply to other generative action models that chunk and re-sample noise, not only flow matching.
  • When the previous tail itself encodes a failing mode, adaptive or gated guidance strength may matter more than a fixed λ schedule.
  • Real-time stacks that cannot afford a policy backward pass per ODE step are a natural fit for closed-form steering of this kind.
  • The reported gap between boundary and interior jerk suggests chunk-boundary artifacts are a first-order deployment issue for chunked VLAs in general.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes SEAM, a training-free inference-time method for flow-matching VLA policies that use action chunking. It identifies multimodal bifurcation at chunk boundaries, where independently sampled adjacent chunks can select incompatible trajectory modes and produce high jerk/discontinuity. SEAM constructs an aligned prior from the previous chunk’s unexecuted tail and applies Velocity-guided Loss Steering (VLS): after each Euler step, a closed-form quadratic correction steers only a guided overlap window toward a time-dependent target r_i = (1−t_i)·a_al[1:M], scaled by λ(1−t), without backpropagating through the policy. On LIBERO-10 with π0.5, SEAM reports ~28% lower boundary jerk, ~27% lower chunk discontinuity, near-baseline success (95.7% vs 94.8%), and ~1.01× denoising-loop cost, positioned between RTC (stronger smoothing, higher cost) and ACT-TE (aggressive smoothing, lower success).

Significance. If the result holds, SEAM fills a practical gap in the cost–smoothness spectrum for chunked generative robot policies: an analytic, backpropagation-free overlap correction that improves boundary motion quality without retraining or rejection sampling. The contribution is useful for real-time VLA deployment, where RTC-style ΠGDM guidance is expensive and temporal ensembling can harm task reliability. Strengths include a clear synchronous-execution insight, a simple closed-form mechanism (Eqs. 5–8), matched cost-aware comparisons (Table 2), and ablations over λ, dimension subset, and guided window M (Tables 4–5). The work is incremental rather than foundational, but the operating regime—smoother chunk transitions at near-baseline latency while preserving task success—is well motivated for robotics practice.

major comments (2)
  1. The headline claim that SEAM “preserves baseline-level task success” (Abstract; Table 2: 95.7% vs 94.8%) is load-bearing for the intended operating regime versus RTC/ACT-TE, but is supported only by point estimates over 130 episodes/task with no SEMs, CIs, bootstrap intervals, or significance tests. Table 3 shows mixed per-task deltas (gains on T1–T3/T5, losses on T7/T10), so a small true regression could be hidden by aggregate noise. Please report uncertainty (e.g., binomial/bootstrap CIs or episode-level SEMs) and, if possible, multi-seed or multi-backbone checks so that “preserves success” is statistically interpretable rather than a single-run aggregate.
  2. The weakest load-bearing assumption is that the previous unexecuted tail is a safe analytic consistency reference (Aligned Prior Construction; VLS, Eqs. 3–8). The paper correctly notes that strong continuation can lock failed modes (RTC qualitative T1 analysis; Figure 4) and intentionally uses weak λ, but the main evaluation remains a single backbone (π0.5) and single benchmark (LIBERO-10). Because the method’s value proposition is “smooth without locking,” the claim would be substantially stronger with at least one additional policy family or a harder contact-rich setting where wrong-mode locking is more likely, or with an explicit failure analysis when the prior itself is incorrect.
minor comments (5)
  1. Abstract and main text round the same reductions slightly differently (28%/27% vs 27.7%/26.7% in the Main Results paragraph). Align the reported percentages with Table 2.
  2. Notation for the reverse Euler schedule is slightly awkward: Eq. (2) writes x_{t+Δt} with Δt = −1/N, while later text uses t_i = 1−i/N and x_i. A short clarifying sentence would help readers track discrete solver indices versus continuous flow time.
  3. Figure 3 is useful for the cost–smoothness trade-off; ensure axis labels and point annotations remain legible in print and that D-Cost is defined in the caption as well as the table.
  4. Related Work cites several concurrent/arXiv-only methods (Legato, FASTER, BID, etc.). Briefly clarify which are concurrent and which are direct baselines to avoid overstating novelty relative to unpublished work.
  5. Table 1 is a helpful algorithm box; consider explicitly stating that observation conditioning o enters v_θ in step 3.1 (as in the surrounding text) for self-contained reproducibility.

Circularity Check

0 steps flagged

No significant circularity: SEAM/VLS is an empirical inference method whose smoothness and success claims are measured outcomes, not algebraically forced by the prior or by self-citation.

full rationale

SEAM does not claim a first-principles derivation that reduces to its inputs. The aligned prior a_al = Extend(a_tail, H) and the time-dependent target r_i = (1−t_i)·a_al[1:M] (Eqs. 3–4) are intentional design choices motivated by the standard flow-matching interpolation (Eq. 1) and the synchronous chunk-overlap structure; they define how VLS steers, not a quantity later re-presented as an independent prediction. The closed-form correction (Eqs. 6–8) is a lightweight post-Euler nudge whose effect on executed trajectories is evaluated empirically (boundary jerk, CD, AVb, success on LIBERO-10 with π0.5), not forced by construction: steering the guided window toward the prior does not algebraically determine post-processed jerk or task success, which depend on the policy, environment, and full trajectory. Hyperparameters (λ, M, dimension set) are chosen by ablation (Tables 4–5)—ordinary method tuning, not fitted inputs renamed as predictions. Related-work citations (RTC, BID, ACT-TE, flow matching) are external baselines and background; there is no load-bearing self-citation uniqueness theorem or ansatz smuggled from the same authors. The paper is self-contained against external benchmarks (LIBERO-10, π0.5). Score 0 with empty steps is the correct outcome.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on standard flow-matching and action-chunking machinery plus a few hand-chosen guidance hyperparameters and the modeling choice that the previous tail is a good soft target. No new physical entities are postulated; the invented pieces are algorithmic constructs (aligned prior, VLS correction).

free parameters (3)
  • guidance strength λ
    Main results use λ=0.1 chosen via ablation on LIBERO-10; larger λ hurts success. The reported smoothness–success trade-off depends on this choice.
  • guided window length M
    Main setting M=20 (of L=40 overlap). Table 5 shows M is a smoothness knob selected on the same benchmark.
  • optional L_max / dimension subset
    Design allows guiding a subset of dimensions and a bounded prefix; main claim uses all physical dimensions and M=20.
axioms (4)
  • domain assumption Flow matching reverse ODE with Euler integration generates action chunks from Gaussian latents (Eqs. 1–2).
    SEAM is defined as a post-Euler correction inside this solver; validity is scoped to flow-matching VLAs like π0.5.
  • domain assumption Synchronous chunked execution executes K≪H actions then re-queries, leaving an unexecuted tail of length L=H−K as available state.
    Aligned prior construction depends on this execution model; asynchronous or non-overlapping schemes fall outside the method as stated.
  • ad hoc to paper The linear path target r_i=(1−t_i)·a_al[1:M] is a useful local consistency surrogate for the reverse state of the overlap.
    Paper explicitly calls this a surrogate, not an exact conditional posterior; VLS success hinges on this modeling choice.
  • domain assumption Independent Gaussian chunk latents can select incompatible modes (multimodal bifurcation) under similar observations.
    Motivation for cross-chunk steering; supported by cited contact-rich multimodality literature and the paper’s qualitative examples.
invented entities (2)
  • Aligned prior a_al no independent evidence
    purpose: Extend previous unexecuted tail to chunk length as the only cross-chunk consistency reference for guidance.
    Algorithmic construct defined by Extend(a_tail, H); no independent physical status beyond the previous prediction.
  • Velocity-guided Loss Steering (VLS) no independent evidence
    purpose: Closed-form post-Euler correction of the guided overlap window toward the time-dependent target without policy backprop.
    Core invented mechanism of SEAM; evidence is empirical on LIBERO-10 rather than external independent validation.

pith-pipeline@v1.1.0-grok45 · 16389 in / 3435 out tokens · 27749 ms · 2026-07-11T16:28:20.593157+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action (VLA) policies that execute fixed-length action chunks can exhibit multimodal bifurcation: a cross-chunk inconsistency in which adjacent chunks generated from independent Gaussian latents can converge to incompatible trajectory modes, producing abrupt discontinuities at chunk boundaries. Existing remedies either require backpropagation through the policy at each denoising step, rely on rejection sampling, or require retraining, each trading computational cost or task reliability for smoother transitions. We propose SEAM (Smooth Execution of Action-Chunked Motion), a training-free inference-time method for flow matching VLAs. SEAM exploits a simple synchronous-execution insight: after the robot consumes the executed prefix, the previous chunk's unexecuted tail is already available as an analytic consistency reference. Its core mechanism, Velocity-guided Loss Steering (VLS), derives a time-dependent target from this tail and applies a closed-form correction after each Euler step without backpropagating through the policy network. On LIBERO-10 with pi_0.5, SEAM reduces boundary jerk by 28%, reduces chunk transition discontinuity by 27%, preserves baseline-level task success, and keeps denoising-loop cost near the unguided baseline.

Figures

Figures reproduced from arXiv: 2607.04609 by Dijia Zhan, Jie Tang, Jinyi Li, Xuemiao Xu.

Figure 1
Figure 1. Figure 1: SEAM overview. (a) Independent chunk sampling can produce two boundary artifacts: an intention conflict ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SEAM. The previous decoded chunk [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cost–smoothness trade-off. D-Cost denotes relative [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative T1 failure analysis. Representative T1 rollouts explain the largest gap in Table 3: the baseline shows intent [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 20 linked inside Pith

  1. [1]

    S.; Kolesnikov, A.; Wang, X.; Salz, D.; Neumann, M.; Alabdulmohsin, I.; Tschannen, M.; Bugliarello, E.; et al

    Beyer, L.; Steiner, A.; Pinto, A. S.; Kolesnikov, A.; Wang, X.; Salz, D.; Neumann, M.; Alabdulmohsin, I.; Tschannen, M.; Bugliarello, E.; et al. 2024. PaliGemma: A Versatile 3B VLM for Transfer. arXiv preprint arXiv:2407.07726

    Pith/arXiv arXiv 2024

  2. [2]

    Black, K.; Brown, N.; Driess, D.; Esmail, A.; Equi, M.; Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; et al. 2024. _0 : A Vision-Language-Action Flow Model for General Robot Control. arXiv preprint arXiv:2410.24164

    Pith/arXiv arXiv 2024

  3. [3]

    Black, K.; Galliker, M.; and Levine, S. 2025. Real-Time Execution of Action Chunking Flow Policies. In Advances in Neural Information Processing Systems, volume 38, 33383--33407

    2025

  4. [4]

    Brohan, A.; Brown, N.; Carbajal, J.; Chebotar, Y.; Dabis, J.; Finn, C.; Gopalakrishnan, K.; Hausman, K.; Herzog, A.; Hsu, J.; et al. 2022. RT-1: Robotics Transformer for Real-World Control at Scale. arXiv preprint arXiv:2212.06817

    Pith/arXiv arXiv 2022

  5. [5]

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

    2025

  6. [6]

    T.; Klasky, M

    Chung, H.; Kim, J.; Mccann, M. T.; Klasky, M. L.; and Ye, J. C. 2022. Diffusion Posterior Sampling for General Noisy Inverse Problems. arXiv preprint arXiv:2209.14687

    Pith/arXiv arXiv 2022

  7. [7]

    Dhariwal, P.; and Nichol, A. 2021. Diffusion Models Beat GANs on Image Synthesis. Advances in Neural Information Processing Systems, 34: 8780--8794

    2021

  8. [8]

    S.; Love, J.; et al

    Gemma Team ; Mesnard, T.; Hardin, C.; Dadashi, R.; Bhupatiraju, S.; Pathak, S.; Sifre, L.; Rivi \`e re, M.; Kale, M. S.; Love, J.; et al. 2024. Gemma: Open Models Based on Gemini Research and Technology. arXiv preprint arXiv:2403.08295

    Pith/arXiv arXiv 2024

  9. [9]

    Ho, J.; and Salimans, T. 2022. Classifier-Free Diffusion Guidance. arXiv preprint arXiv:2207.12598

    Pith/arXiv arXiv 2022

  10. [10]

    Jia, J.; Li, G.; Chen, X.; An, T.; Hu, Y.; Li, J.; Guo, X.; and Yang, J. 2026. Action-to-Action Flow Matching. arXiv preprint arXiv:2602.07322

    Pith/arXiv arXiv 2026

  11. [11]

    Kawar, B.; Elad, M.; Ermon, S.; and Song, J. 2022. Denoising Diffusion Restoration Models. Advances in Neural Information Processing Systems, 35: 23593--23606

    2022

  12. [12]

    J.; Pertsch, K.; Karamcheti, S.; Xiao, T.; Balakrishna, A.; Nair, S.; Rafailov, R.; Foster, E.; Lam, G.; Sanketi, P.; et al

    Kim, M. J.; Pertsch, K.; Karamcheti, S.; Xiao, T.; Balakrishna, A.; Nair, S.; Rafailov, R.; Foster, E.; Lam, G.; Sanketi, P.; et al. 2024. OpenVLA: An Open-Source Vision-Language-Action Model. arXiv preprint arXiv:2406.09246

    Pith/arXiv arXiv 2024

  13. [13]

    T.; Ben-Hamu, H.; Nickel, M.; and Le, M

    Lipman, Y.; Chen, R. T.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2022. Flow Matching for Generative Modeling. arXiv preprint arXiv:2210.02747

    Pith/arXiv arXiv 2022

  14. [14]

    Liu, B.; Zhu, Y.; Gao, C.; Feng, Y.; Liu, Q.; Zhu, Y.; and Stone, P. 2023. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. Advances in Neural Information Processing Systems, 36: 44776--44791

    2023

  15. [15]

    I.; Xie, A.; Lee, Y.; Du, M.; and Finn, C

    Liu, Y.; Hamid, J. I.; Xie, A.; Lee, Y.; Du, M.; and Finn, C. 2025. Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. In International Conference on Learning Representations

    2025

  16. [16]

    Liu, Y.; Yu, H.; Zhao, J.; Li, B.; Zhang, D.; Li, M.; Wu, W.; Hu, Y.; Xie, J.; Guo, J.; et al. 2026. Learning Native Continuation for Action Chunking Flow Policies. arXiv preprint arXiv:2602.12978

    Pith/arXiv arXiv 2026

  17. [17]

    Lu, Y.; Liu, Z.; Fan, X.; Yang, Z.; Hou, J.; Li, J.; Ding, K.; and Zhao, H. 2026. FASTER: Rethinking Real-Time Flow VLAs. arXiv preprint arXiv:2603.19199

    Pith/arXiv arXiv 2026

  18. [18]

    Lugmayr, A.; Danelljan, M.; Romero, A.; Yu, F.; Timofte, R.; and Van Gool, L. 2022. RePaint: Inpainting Using Denoising Diffusion Probabilistic Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11461--11471

    2022

  19. [19]

    Meng, C.; He, Y.; Song, Y.; Song, J.; Wu, J.; Zhu, J.-Y.; and Ermon, S. 2021. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. arXiv preprint arXiv:2108.01073

    Pith/arXiv arXiv 2021

  20. [20]

    Octo Model Team ; Ghosh, D.; Walke, H.; Pertsch, K.; Black, K.; Mees, O.; Dasari, S.; Hejna, J.; Kreiman, T.; Xu, C.; et al. 2024. Octo: An Open-Source Generalist Robot Policy. arXiv preprint arXiv:2405.12213

    Pith/arXiv arXiv 2024

  21. [21]

    Physical Intelligence ; Black, K.; Brown, N.; Darpinian, J.; Dhabalia, K.; Driess, D.; Esmail, A.; Equi, M.; Finn, C.; Fusai, N.; et al. 2025. _ 0.5 : A Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054

    Pith/arXiv arXiv 2025

  22. [22]

    M.; Cui, Z.; Altanzaya, A

    Shafiullah, N. M.; Cui, Z.; Altanzaya, A. A.; and Pinto, L. 2022. Behavior Transformers: Cloning k Modes with One Stone. Advances in Neural Information Processing Systems, 35: 22955--22968

    2022

  23. [23]

    P.; Kumar, A.; Ermon, S.; and Poole, B

    Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2020. Score-Based Generative Modeling through Stochastic Differential Equations. arXiv preprint arXiv:2011.13456

    Pith/arXiv arXiv 2020

  24. [24]

    Tong, A.; Fatras, K.; Malkin, N.; Huguet, G.; Zhang, Y.; Rector-Brooks, J.; Wolf, G.; and Bengio, Y. 2023. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. arXiv preprint arXiv:2302.00482

    Pith/arXiv arXiv 2023

  25. [25]

    Wang, R. 2026. Noise-Space Attribution and Control of Chunk-Boundary Artifact. arXiv preprint arXiv:2603.11642

    Pith/arXiv arXiv 2026

  26. [26]

    Wen, J.; Zhu, M.; Liu, J.; Liu, Z.; Yang, Y.; Zhang, L.; Zhang, S.; Zhu, Y.; and Xu, Y. 2025 a . dVLA: Diffusion Vision-Language-Action Model with Multimodal Chain-of-Thought. arXiv preprint arXiv:2509.25681

    arXiv 2025

  27. [27]

    Wen, Y.; Li, H.; Gu, K.; Zhao, Y.; Wang, T.; and Sun, X. 2025 b . LLaDA-VLA: Vision Language Diffusion Action Models. arXiv preprint arXiv:2509.06932

    Pith/arXiv arXiv 2025

  28. [28]

    Wu, W.; Lu, F.; Wang, Y.; Yang, S.; Liu, S.; Wang, F.; Zhu, Q.; Sun, H.; Wang, Y.; Ma, S.; et al. 2026. A Pragmatic VLA Foundation Model. arXiv preprint arXiv:2601.18692

    Pith/arXiv arXiv 2026

  29. [29]

    Yashima, D.; Seno, K.; Kurita, S.; Oda, Y.; and Sugiura, K. 2026. HiFlow: Tokenization-Free Scale-Wise Autoregressive Policy Learning via Flow Matching. arXiv preprint arXiv:2603.27281

    arXiv 2026

  30. [30]

    Zhai, X.; Mustafa, B.; Kolesnikov, A.; and Beyer, L. 2023. Sigmoid Loss for Language Image Pre-Training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 11975--11986

    2023

  31. [31]

    Z.; Kumar, V.; Levine, S.; and Finn, C

    Zhao, T. Z.; Kumar, V.; Levine, S.; and Finn, C. 2023. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. arXiv preprint arXiv:2304.13705

    Pith/arXiv arXiv 2023

  32. [32]

    Zitkovich, B.; Yu, T.; Xu, S.; Xu, P.; Xiao, T.; Xia, F.; Wu, J.; Wohlhart, P.; Welker, S.; Wahid, A.; et al. 2023. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In Conference on Robot Learning, 2165--2183. PMLR

    2023