Pith. sign in

REVIEW 4 major objections 6 minor 34 references

Feeding a frozen VLA policy a predicted execution-time observation-and-state pair restores delay-free performance in asynchronous robot control.

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 · deepseek-v4-flash

2026-07-31 23:16 UTC pith:W7SUYXSR

load-bearing objection FutureRTC is a genuinely useful plug-and-play adapter for async VLA execution—the action-conditional latent warping idea is new and the controlled analysis is solid—but the central mechanism needs direct validation and the empirical reporting lacks variances. the 4 major comments →

arxiv 2607.24008 v1 pith:W7SUYXSR submitted 2026-07-27 cs.RO

FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking

Hai Jiang , Yixian Zou , Binbin Liang , Boqian Liu , Fanman Meng , Shuaicheng Liu This is my paper
classification cs.RO
keywords asynchronous executionvision-language-action modelsaction chunkingobservation predictionstate correctionlatent warppolicy consistencyrobot 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.

FutureRTC targets the staleness problem in async robot control: when a vision-language-action (VLA) policy computes the next chunk of actions, the robot has already moved on, so by the time the chunk is executed, both the camera view and the arm state are outdated. The paper's central claim is that the untouched policy can be made to act as if the delay never existed by feeding it a predicted 'execution-time context' — a forecast of the visual observation and the proprioceptive state at the exact moment the chunk will be run. The paper argues that existing fixes (smoothing chunk boundaries, fine-tuning the policy, or predicting only the future arm state) miss the dominant problem: the stale visual observation. If FutureRTC is right, asynchronous execution can achieve delay-invariant success rates close to the synchronous upper bound with only a small adapter, making real-time deployment of large VLAs practical.

Core claim

FutureRTC claims that prediction-execution misalignment in asynchronous VLA execution is resolved by recovering the execution-time context rather than by patching the generated actions. The adapter computes an anticipated pair (ˆz_{t+K}, ˆs_{t+K}) — a latent visual feature forecast (a compressed representation of the image) and a corrected proprioceptive state (the robot's own joint/arm configuration) at the moment the chunk will be executed — and the frozen policy is conditioned on this pair and executed directly from its first action. In experiments, this makes success rates degrade only mildly as inference delay grows (94.2% → 88.5% at d=20 on LIBERO with π0.5, and 75.8% → 69.4% with Smol

What carries the argument

Two trainable modules plus one loss. The State Correction Module (SCM) rolls the last committed actions forward and applies a small MLP residual to correct the gap between commanded and actual motion. The Observation Prediction Module (OPM) works in the VLA vision encoder's latent space: it builds a motion prior from the committed actions, moves the stale feature along the implied 2D flow field via gated bilinear warping (transport gate α), and adds a gated synthesis residual (gate β) that borrows from the initial observation to fill in content that warping cannot explain, such as disocclusions or deformation on contact. A policy consistency loss trains the whole adapter so that the action c

Load-bearing premise

The load-bearing premise is that the vision encoder's latent features preserve enough 2D spatial structure that the execution-time view can be produced by warping the stale feature along an action-derived flow and patching with a residual; if the latent space is not spatially warpable, or scene change is driven by something other than the robot's own motion, the predicted observation degrades.

What would settle it

Run FutureRTC on a VLA whose vision encoder uses a global-attention transformer without clear 2D token neighborhoods, and compare delay robustness against state-only correction; if the margin collapses, the spatial-warpability premise is the culprit. Alternatively, introduce an independently moving object into the LIBERO or real-world scene during the delay window and measure success: the paper's own limitation predicts a sharp drop toward the naive-async baseline.

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

If this is right

  • Executing a chunk from its first action is correct once the conditioning context is aligned to execution time; the naive habit of discarding the first d actions is only a stopgap for stale inputs.
  • Predicting the future visual observation, not just the future arm state, is the decisive component: state-only correction leaves most of the performance gap on π0.5, where the state is textual tokens, and only partially helps on SmolVLA.
  • Asynchronous execution itself carries no inherent penalty: with the true execution-time pair, success stays flat across all delays at the delay-free level.
  • The adapter transfers across two different VLA backbones with only ~5–6M added parameters and ~3 ms added latency.

Where Pith is reading between the lines

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

  • Editor's inference: The OPM's warpability premise implies FutureRTC should lose its edge on vision backbones whose tokens do not keep 2D spatial order (e.g., heavily global-attention ViTs); that is a direct, testable consequence not explored in the paper.
  • Editor's inference: The same recipe — build a motion prior from the agent's own actions, transport and patch the latent, then enforce policy consistency — could transfer to other settings where the agent's motion dominates visual change, but the paper's limitation section already concedes it will struggle when independent external agents alter the scene.
  • Editor's inference: Because the adapter is trained by matching downstream actions rather than by latent fidelity, it may exploit the policy's tolerance for imperfect predictions; measuring prediction loss directly would show whether the observed success comes from accurate forecasting or from the policy ignoring small latent errors.
  • Editor's inference: The uniform-delay training trick (sampling d in [1,dmax]) means robustness is bought by covering the delay range, so one would expect graceful degradation outside the trained range — a boundary worth probing at d>dmax.

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

4 major / 6 minor

Summary. The paper proposes FutureRTC, a plug-and-play adapter for asynchronous execution of frozen flow-matching VLA policies. At each chunk boundary, instead of conditioning the next chunk on the stale pair (o_{t+K-d}, s_{t+K-d}), the adapter predicts the execution-time context (z_hat_{t+K}, s_hat_{t+K}). The state correction module (SCM) learns a residual to improve forward integration of committed actions (Eqs. 3-4). The observation prediction module (OPM) warps the stale vision-encoder latent along a 2D flow derived from the committed action sequence, then adds a gated synthesis residual (Eqs. 5-7), trained with an L2 latent loss (Eq. 8). A policy consistency loss (Eq. 9) aligns the action chunk from predicted context with that from ground-truth context using a single-step flow approximation. The frozen VLA then generates the next chunk from the predicted context and executes it from index 0. Experiments on Kinetix, LIBERO with pi0.5 and SmolVLA-450M, and three real-world bimanual tasks report higher success rates, fewer execution steps, and smoother trajectories across delays, at the cost of ~5M parameters and ~3ms added latency.

Significance. If the results hold, FutureRTC is a practically valuable contribution to asynchronous VLA deployment: it avoids retraining the policy, introduces minimal compute, and the controlled analysis in Table 3 is a strong sanity check showing that asynchronous execution with oracle execution-time context is delay-invariant. The cross-backbone experiments and per-suite/per-task appendices add credibility. The main risk is that the core OPM mechanism (spatial warpability of latent features) is not directly validated, and the empirical comparisons lack uncertainty quantification. These issues are fixable and do not invalidate the idea, but they must be addressed before the quantitative claims can be accepted.

major comments (4)
  1. [Observation Prediction Module, Eqs. (6)-(8)] The central claim that the VLA is conditioned on execution-time observations inherits the OPM's assumption that the vision-encoder latent space is spatially warpable along an action-derived 2D flow. This premise is load-bearing but never directly tested. Lobs in Eq. (8) is reported only as a training loss; no held-out latent prediction error, no qualitative comparison of predicted vs. ground-truth latents, and no ablation isolating the transport gate alpha, synthesis gate beta, or the z0 reference in Eq. (7). Table 2 reports downstream success only, which can be insensitive to latent corruption. Since the Limitations section itself concedes that the OPM 'may struggle in highly dynamic scenes dominated by independent external agents'--precisely where warpability fails--the current evidence does not establish that the gains are due to the proposed mechanism. Please add direct fidelity metr
  2. [Tables 1-2 and Fig. 6] No error bars, confidence intervals, or significance tests appear anywhere. Real-world evaluation uses 20 trials per task; several advantages over baselines are small (e.g., Fig. 6, Fold Towel at d approx 5: 80% vs 75% for VLASH, i.e., one success out of 20). Simulated averages in Tables 1 and 2 are reported without variance or trial counts. Since the paper's headline claim is consistent, cross-delay improvements over strong baselines, the empirical claim is not yet fully supported. Please report per-seed/trial results, confidence intervals, and significance tests for at least the key LIBERO and real-world comparisons.
  3. [Policy Consistency Loss, Eq. (9)] The consistency loss is computed with a single Euler step of the frozen flow-matching policy. For flow matching, one Euler step is generally a very coarse approximation of the full generative trajectory; if it is far from the final action distribution, Lpolicy may encourage the adapter to produce contexts that match only the first integration step. The paper does not justify this approximation or ablate it. Please include a comparison with multi-step target generation (e.g., n=1 vs n=4/8) or an analysis of the approximation error.
  4. [Table 1 and Real-World Evaluation] For training-time baselines (T-RTC, REMAC), the paper states they are reproduced following their publications, but no reproduction details (hyperparameters, training steps, seeds) are given. For the real-world set, VLASH is fine-tuned on pi0.5; however, no comparable fine-tuning details are shown beyond Table 5. If reproduced baselines are undertrained, the headline improvements would be inflated. Please provide full training configurations and, ideally, release code/checkpoints.
minor comments (6)
  1. [Notation] Eq. (9) writes pi_theta(z_{t+K}, s_{t+K}) although Eq. (1) defines the policy on observations o_t; clarify that the latent is injected after the frozen vision encoder E(.).
  2. [Table 2 vs Table 1] Table 2's Baseline uses A[0:K-1] while Table 1's Naive Async. uses A[d:d+K-1]; the appendix Table 3 explains the different splicing conventions, but the main text should state this distinction explicitly when interpreting Table 2.
  3. [Implementation Details] The hyperparameter d_max is said to be the largest delay considered on each benchmark but is never given numerically per benchmark; please report it. Also report the exact value of lambda used in L_total in the experiments section.
  4. [Figure 6] Figure 6 is dense: the three metric panels and two delay conditions are not clearly separated, and the legend format is hard to read. Please annotate each panel with the corresponding delay and metric.
  5. [Overhead reporting] The overhead numbers are inconsistently quoted: the abstract emphasizes ~5M parameters, while Table 4 reports +6.45M for pi0.5. Use the per-backbone numbers when summarizing.
  6. [Reproducibility] No code/data release link is provided beyond a Project Website placeholder; a reproducibility package would strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: predictions are regressed against externally recorded execution-time context and evaluated on external benchmarks.

full rationale

FutureRTC's adapter is trained with direct supervised losses against ground-truth execution-time data: L_state in Eq. 4 regresses the residual between the rolled-forward state and the recorded state at t+K, and L_obs in Eq. 8 regresses the predicted latent ẑ_{t+K} to the recorded latent z_{t+K}. These targets are external to the model's own predictions, not derived from the adapter's outputs. The policy consistency loss (Eq. 9) uses the frozen policy on both sides, but it is a regularizer that encourages the predicted context to be faithfully consumed; it does not define the predicted context or the success metric. All central comparisons (LIBERO, Kinetix, real-world tasks) are evaluated by downstream task success and execution efficiency, which are independent of the adapter's training objective. The paper does not invoke a uniqueness theorem or a load-bearing self-citation chain; the cited prior works provide baselines and backbones, not the paper's own mechanism. The OPM's warpability assumption is a correctness risk, and the Limitations section candidly notes that the module 'may struggle in highly dynamic scenes dominated by independent external agents,' but this is an empirical robustness caveat, not an equation-level circularity. No step in the derivation reduces to its own inputs by construction, so no circularity is found.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities: α and β are learned per-token gates, not independent forces or objects. The main load-bearing assumptions are spatial warpability of latent features, robot-motion dominance of visual change, and the fidelity of the frozen policy when consuming predicted latents. The single-step flow approximation is an ad hoc simplification that is not justified with sensitivity analysis.

free parameters (2)
  • λ (policy consistency loss weight) = 10
    Set empirically in Implementation Details; no sensitivity study is provided, and λ directly balances the action-consistency term against the state/observation regression losses.
  • d_max (training delay sampling bound) = per-benchmark (unspecified in main text)
    Delay d is sampled uniformly in [1, d_max] for robustness; d_max is a hand-chosen hyperparameter that defines coverage of evaluated delays.
axioms (5)
  • domain assumption The frozen VLA policy can faithfully consume an injected latent vector z_hat in place of its own vision encoder output.
    Used at inference (the predicted latent is directly injected, bypassing the encoder) and trained via L_policy; if the policy's internal distribution is sensitive to latent statistics, the injection could silently degrade actions.
  • domain assumption Visual change during the delay window is dominated by robot motion and is representable as 2D feature transport plus a small gated residual.
    Eq. (6)-(7) warp the stale latent along a predicted flow and synthesize a residual; this fails for independent external motion, which the authors acknowledge in Limitations.
  • domain assumption The VLA vision encoder features retain a 2D spatial layout supporting bilinear grid sampling.
    Eq. (6) performs spatial warping of latent tokens; ViT features are only approximately spatial, and the paper provides no direct validation that predicted latents are accurate reconstructions of ground-truth latents.
  • domain assumption Forward integration of commanded actions plus a learned MLP residual yields the execution-time proprioceptive state.
    Eq. (3)-(4) assume residual bias is learnable and that commanded actions are the right kinematic input; contact, slip, or controller tracking error would violate this.
  • ad hoc to paper A single Euler step of the frozen flow-matching policy is a sufficient approximation for the action-consistency target.
    Section 'Policy Consistency Loss' replaces multi-step flow integration with a one-step approximation; the shared-initial-noise condition for comparing the two action chunks is not stated.

pith-pipeline@v1.3.0-alltime-deepseek · 22838 in / 14923 out tokens · 126195 ms · 2026-07-31T23:16:18.299537+00:00 · methodology

0 comments
read the original abstract

Real-time deployment of Vision-Language-Action (VLA) policies necessitates asynchronous execution, wherein subsequent action chunks are computed concurrently with the execution of the current chunk, leading to prediction-execution misalignment and manifesting as inter-chunk discontinuities. Existing methods either superficially smooth chunk boundaries, require costly policy optimization, or exclusively forward-predict proprioceptive states yet neglect critical visual observations. In this paper, we propose \textbf{FutureRTC}, a plug-and-play adaptation framework that predicts execution-time observations and states for asynchronous VLA control without modifying the underlying policy. Specifically, FutureRTC features a state correction module to compensate for the discrepancy between rolled-forward and actual execution-time proprioceptive states and an observation prediction module that forecasts execution-time visual representations by leveraging robot motion as an explicit physical prior through motion-aware feature transport and reconstruction. Furthermore, we introduce a policy consistency loss to align the action chunks generated from predicted contexts with those produced under the expected execution-time inputs of the VLA policy. Extensive experiments across simulated and real-world environments demonstrate that FutureRTC achieves superior robustness to inference delays, resulting in smoother trajectories, faster execution, and consistently higher task success rates.

Figures

Figures reproduced from arXiv: 2607.24008 by Binbin Liang, Boqian Liu, Fanman Meng, Hai Jiang, Shuaicheng Liu, Yixian Zou.

Figure 1
Figure 1. Figure 1: Comparison of the inference-time method RTC ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between our method and existing execution paradigms, illustrated with prediction horizon [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the detailed architecture of the pro [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the transport gate α (top) and syn￾thesis gate β (bottom) during execution. The transport gate activates on the moving arm and gripper, propagating ex￾isting visual features according to the committed motion, whereas the synthesis gate highlights motion boundaries and newly revealed regions that require feature synthesis. ture latent from F is highly ill-posed and physically uncon￾strained.… view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison in the Kinetix simula [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-world comparison on three bimanual manip [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Average action velocity and acceleration on the [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Real-world experimental platform. (a) The AgileX Cobot Magic robot. (b) The workspace of the three bimanual tasks, containing the plates, towel, and cups with a cup rack. 2025), and REMAC (Wang et al. 2026a) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance comparison in the Kinetix simulator ( [PITH_FULL_IMAGE:figures/full_fig_p014_10.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

34 extracted references · 7 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2512.05964 , year=

    Training-Time Action Conditioning for Efficient Real-Time Chunking , author=. arXiv preprint arXiv:2512.05964 , year=

    arXiv

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Adaptive Action Chunking at Inference-Time for Vision-Language-Action Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [3]

    International Conference on Learning Representations , pages=

    Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling , author=. International Conference on Learning Representations , pages=

  4. [4]

    International Conference on Learning Representations , year=

    Real-Time Robot Execution with Masked Action Chunking , author=. International Conference on Learning Representations , year=

  5. [5]

    Advances in Neural Information Processing Systems , pages=

    Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking , author=. Advances in Neural Information Processing Systems , pages=

  6. [6]

    Advances in Neural Information Processing Systems , pages=

    Real-Time Execution of Action Chunking Flow Policies , author=. Advances in Neural Information Processing Systems , pages=

  7. [7]

    Proceedings of Robotics: Science and Systems , year=

    Learning Native Continuation for Action Chunking Flow Policies , author=. Proceedings of Robotics: Science and Systems , year=

  8. [8]

    Proceedings of Robotics: Science and Systems , year=

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware , author=. Proceedings of Robotics: Science and Systems , year=

  9. [9]

    IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

    PD-VLA: Accelerating Vision-Language-Action Model Integrated with Action Chunking via Parallel Decoding , author=. IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

  10. [10]

    arXiv preprint arXiv:2512.01031 , year=

    VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference , author=. arXiv preprint arXiv:2512.01031 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2510.26742 , year=

    Running VLAs at Real-Time Speed , author=. arXiv preprint arXiv:2510.26742 , year=

    arXiv

  12. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages=

    Realtime-VLA V2: Learning to Run VLAs Fast, Smooth, and Accurate , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages=

  13. [13]

    Proceedings of Robotics: Science and Systems , year=

    3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations , author=. Proceedings of Robotics: Science and Systems , year=

  14. [14]

    The International Journal of Robotics Research , volume=

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author=. The International Journal of Robotics Research , volume=

  15. [15]

    Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and others , booktitle=. _

  16. [16]

    arXiv preprint arXiv:2501.09747 , year=

    FAST: Efficient Action Tokenization for Vision-Language-Action Models , author=. arXiv preprint arXiv:2501.09747 , year=

    Pith/arXiv arXiv

  17. [17]

    Black, Kevin and Brown, Noah and Darpinian, James and Dhabalia, Karan and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Galliker, Manuel Y and others , booktitle=. _

  18. [18]

    arXiv preprint arXiv:2506.01844 , year=

    Smolvla: A vision-language-action model for affordable and efficient robotics , author=. arXiv preprint arXiv:2506.01844 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2503.14734 , year=

    Gr00t n1: An open foundation model for generalist humanoid robots , author=. arXiv preprint arXiv:2503.14734 , year=

    Pith/arXiv arXiv

  20. [20]

    Annual Conference on Robot Learning , year=

    OpenVLA: An Open-Source Vision-Language-Action Model , author=. Annual Conference on Robot Learning , year=

  21. [21]

    Advanced Robotics , volume=

    Real-world robot applications of foundation models: A review , author=. Advanced Robotics , volume=

  22. [22]

    International Conference on Learning Representations , year=

    Flow Matching for Generative Modeling , author=. International Conference on Learning Representations , year=

  23. [23]

    Annual Conference on Robot Learning , pages=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. Annual Conference on Robot Learning , pages=

  24. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Acot-vla: Action chain-of-thought for vision-language-action models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  25. [25]

    International Conference on Learning Representations , pages=

    Rdt-1b: a diffusion foundation model for bimanual manipulation , author=. International Conference on Learning Representations , pages=

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Libero: Benchmarking knowledge transfer for lifelong robot learning , author=. Advances in Neural Information Processing Systems , volume=

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Mlp-mixer: An all-mlp architecture for vision , author=. Advances in Neural Information Processing Systems , volume=

  28. [28]

    arXiv preprint arXiv:2606.19774 , year=

    Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection , author=. arXiv preprint arXiv:2606.19774 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2604.25050 , year=

    DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors , author=. arXiv preprint arXiv:2604.25050 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2509.23224 , year=

    Leave No Observation Behind: Real-time Correction for VLA Action Chunks , author=. arXiv preprint arXiv:2509.23224 , year=

    arXiv

  31. [31]

    International Conference on Learning Representations , year=

    LeRobot: An Open-Source Library for End-to-End Robot Learning , author=. International Conference on Learning Representations , year=

  32. [32]

    International Conference on Learning Representations , year=

    Adam: A Method for Stochastic Optimization , author=. International Conference on Learning Representations , year=

  33. [33]

    arXiv preprint arXiv:2601.18692 , year=

    A Pragmatic VLA Foundation Model , author=. arXiv preprint arXiv:2601.18692 , year=

    Pith/arXiv arXiv

  34. [34]

    Annual Conference on Robot Learning , pages=

    DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control , author=. Annual Conference on Robot Learning , pages=