Pith. sign in

REVIEW 3 major objections 6 minor 92 references

A compact robot policy can map vision and language straight to actions at 32 Hz without routing every step through a large language model.

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

2026-07-30 10:54 UTC pith:MTQYB5LS

load-bearing objection Solid efficiency result on a known recipe: drop the LLM from the control loop and you still sit in the LIBERO top tier at ~30 Hz / <1 GB — the “outperforming” framing is thinner than the latency win. the 3 major comments →

arxiv 2607.27205 v1 pith:MTQYB5LS submitted 2026-07-29 cs.CV cs.RO

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Hengyi Xie , Chenfei Yao , Xianjin Wu , Xuanyang Xi , Yiping Tang , Di Xu , Yingying Zhu , Dingkang Liang
show 2 more authors
Xiang Bai Han Ding
This is my paper
classification cs.CV cs.RO
keywords vision-language-actionrobotic manipulationreal-time inferencecross-attention fusionaction chunkingefficient VLALIBERO
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.

Most vision-language-action robot policies send every camera frame through a multi-billion-parameter language model before producing motor commands. That design works, but it is slow and memory-heavy, which blocks real-time control on ordinary hardware. TurboVLA argues that execution-level manipulation does not need that bottleneck: once an instruction already names the skill, the policy mainly needs language to select the right visual cues and turn them into continuous motions. The system encodes images and text separately, fuses them with a small bidirectional cross-attention block, and decodes action chunks in one pass. On the LIBERO benchmark it reaches 97.7% average success with 0.2B parameters, 31.2 ms latency, and under 1 GB of GPU memory on a consumer RTX 4090, matching or beating much larger policies, and it also holds up on bimanual and real-robot tasks.

Core claim

For concrete, execution-level language-conditioned manipulation, a direct V+L→A pathway—lightweight text encoding, compact vision features, bidirectional vision-language interaction, and non-autoregressive continuous action-chunk decoding—can match or exceed the success rates of substantially larger LLM-centric V→L→A policies while cutting parameters, latency, and inference memory by a large margin.

What carries the argument

The V+L→A pathway: independent vision and BERT-style text encoders, a stacked bidirectional cross-attention interaction module that produces instruction-conditioned visual features and vision-aware instruction features, plus an ACT-style parallel action-chunk decoder conditioned on robot state.

Load-bearing premise

Once the instruction already specifies the intended skill, low-level control only needs language semantics to gate visual features for continuous actions, and does not need a large generative language model at every control step.

What would settle it

Train and evaluate the same compact V+L→A architecture against strong LLM-centric VLAs on tasks whose instructions require multi-step semantic planning or open-ended decomposition rather than concrete skill labels; if success collapses while large LLM-centric policies hold up, the central claim fails for general language-conditioned control.

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

If this is right

  • Language-conditioned robot policies can run locally on consumer GPUs at tens of hertz without remote LLM servers.
  • Efficiency gains for execution-level VLA need not come only from faster action heads or pruning; redesigning the multimodal pathway itself can move the performance–latency frontier.
  • Benchmarking VLA systems should jointly report success, end-to-end latency, parameter count, and inference VRAM, not success alone.
  • High-level LLM planning and lightweight execution policies can be stacked rather than forcing one large model to do both.

Where Pith is reading between the lines

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

  • If the assumption holds, many deployed manipulators may prefer a small always-on executor plus occasional high-level planning calls instead of a permanent multi-billion-parameter backbone in the control loop.
  • The same bidirectional fusion pattern may transfer to other closed-loop embodied settings where language is a task gate rather than a generative reasoner, such as mobile navigation with short commands.
  • Stress tests that gradually increase instruction ambiguity and horizon would map the boundary between ‘execution’ and ‘planning’ more sharply than current suite averages.

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

3 major / 6 minor

Summary. The paper proposes TurboVLA, a vision-language-action policy that replaces the LLM-centric V→L→A pathway with a direct V+L→A mapping: DINOv3 visual features and BERT instruction features are fused through N=6 bidirectional cross-attention layers (initialized from Grounding-DINO feature-enhancement weights) and decoded by an ACT-style transformer into 12-step continuous action chunks. The resulting 0.2B-parameter policy runs at 31.2 ms latency and 0.9 GB inference VRAM on an RTX 4090 while achieving 97.7% average success on LIBERO (2,000 rollouts), 60.2% on RoboTwin 2.0 bimanual multi-task (clean setting), and 80–92.5% on four real-world AgileX Piper tasks, compared against π0.5 under a shared fine-tuning protocol. Ablations cover language conditioning, text-encoder choice, interaction directionality, depth N, and horizon H. The central claim is that execution-level manipulation does not require an LLM at the center of the control pathway, and that removing it improves the performance–efficiency frontier.

Significance. If the results hold, this is a useful and timely contribution to the efficiency debate in VLA design. The efficiency half of the claim is well supported: latency, VRAM, and parameter counts are re-measured by the authors on a single RTX 4090 using official implementations and checkpoints at batch size one (§5.2), so the ~3× latency gap over π0.5 and the sub-1-GB footprint are robust to measurement noise. The evaluation is unusually broad for an architecture paper — four LIBERO suites with 2,000 trials, a 50-task bimanual setting, a real-robot comparison against π0.5 under a shared protocol, and a genuinely informative ablation suite (Tab. 3's language-removal collapse on LIBERO-Goal, 97.4%→11.6%, is a clean demonstration that the policy uses instruction semantics rather than visual priors). Code is promised at a public repository. The work also makes a falsifiable architectural claim — that bidirectional cross-attention between compact encoders suffices for execution-level control — and the directionality ablation (Tab. 5) directly tests it. The scope is honestly bounded: the authors state in the Conclusion that the approach targets concrete execution-level instructions and does no

major comments (3)
  1. [Tab. 1 / §5.3 / Abstract] The headline phrase 'matching or outperforming substantially larger VLA policies' is only half supported. At 50 rollouts per task (§5.2), i.e., 500 trials per suite, the standard error on a ~95–99% success rate is roughly 0.6–1.0 points. TurboVLA's margins over the three nearest baselines — CogVLA 97.4%, VLA-Adapter 97.3%, VLA-JEPA 97.2% — are 0.3–0.5 points, i.e., within sampling noise even before protocol differences are counted. Moreover, §5.2 states that efficiency metrics were re-measured 'for all other runnable methods' but makes no analogous statement for the success-rate column, which appears to be assembled from the cited papers, each with its own data processing, checkpoint selection, and rollout harness; LIBERO success rates are known to shift by several points across evaluation setups. The 'matching' claim is secure (TurboVLA is clearly in the top tier, and the LIBERO-Long ga
  2. [Fig. 4 / §5.3 (real-world evaluation)] The claim that TurboVLA 'consistently outperform[s] π0.5' on the real platform rests on 40 trials per task. At success rates of 80–92.5%, the binomial standard error is ~5–7 points per task, and per-task gaps over π0.5 visible in Fig. 4 appear to be of comparable magnitude. The shared fine-tuning protocol (same platform, 4×65 demonstrations, same evaluation) is commendable and removes the usual confounds, but the trial count is too small to support 'consistently outperforming' as stated. Please report the exact π0.5 per-task numbers alongside TurboVLA's (the bar chart alone is insufficient), give confidence intervals, and either increase trial counts on the tasks where the margin is smallest or temper the wording to 'competitive with or better than π0.5 in this evaluation.'
  3. [§5.1 / Tab. 1 ('Emb. PT.' column)] TurboVLA is marked ✗ for embodied pretraining, which is accurate for robot data, but the comparison understates the role of non-robot pretraining in the efficiency story: the interaction module is initialized from Grounding-DINO feature-enhancement weights (§5.1), and DINOv3 and BERT are both large-scale pretrained encoders. Since several baselines (e.g., VLA-Adapter, Evo-1) are also ✗ yet carry different pretraining inheritances, the table invites an apples-to-apples reading that the text does not fully qualify. This does not undermine the result, but the paper's efficiency argument would be strengthened by (i) an ablation training the interaction module from scratch to quantify how much of the 97.7% depends on grounding-pretrained initialization, and (ii) a sentence in §5.1 clarifying the pretraining provenance of every component.
minor comments (6)
  1. [Title / §1] The title advertises '32 Hz' while the abstract and Tab. 1 report 31.2 ms (≈32.1 Hz) as the chunk-prediction latency. Since the policy emits 12-step action chunks, the effective control frequency depends on how many chunk steps are executed before re-planning; please clarify the execution schedule (full-chunk open-loop vs. receding-horizon) and state the resulting closed-loop control rate, which is the quantity that matters for the real-time claim.
  2. [§5.2 (latency protocol)] Latency is defined as 'multimodal input to producing an action chunk or an equivalent number of autoregressive action tokens.' Please specify the equivalence rule (e.g., 12 tokens for OpenVLA-class models?) and whether autoregressive baselines were measured with KV-caching enabled, since this materially affects the fairness of the 202.9 ms OpenVLA figure.
  3. [Tab. 2] RoboTwin 2.0 evaluation is restricted to clean demonstrations 'given our available compute budget.' Since RoboTwin 2.0's domain randomization is a main selling point of the benchmark, please note explicitly in the text (not only the table caption) that randomized-scene generalization is untested, and ideally report FlowPolicy's missing latency ('–') or explain its absence.
  4. [§4.2 / Fig. 3(b)] The FusionLayer is described only at the level of 'layer normalization, bidirectional cross-attention, and modality-specific FFNs.' Please give the exact ordering (pre- vs. post-norm), whether attention is multi-head with how many heads, and the FFN hidden ratio — the module is the paper's core contribution and should be reproducible from the text alone.
  5. [Tab. 4] The parameter counts for the text-encoder ablation (216.9M for SigLIP-Base vs 216.1M for BERT) suggest these are total policy parameters; please label the column accordingly and explain why the SigLIP variant, despite a larger text tower, does not change the total more.
  6. [General] The camera setup uses RGB-D sensors (Fig. 4), but §4.1 and the architecture diagram describe only RGB processing. Please state whether depth is used anywhere in the real-world pipeline.

Circularity Check

0 steps flagged

No circularity: empirical architecture paper with BC training and external rollouts, not a derivation that re-labels inputs as predictions.

full rationale

TurboVLA is an engineering/architecture paper. Its load-bearing claims are measured quantities—LIBERO/RoboTwin success rates from closed-loop rollouts, and latency/VRAM/params on an RTX 4090—not first-principles predictions. Training is standard behavior cloning with L1 on expert action chunks (§4.3); the V+L→A pathway is a design choice (compact encoders, bidirectional cross-attention, ACT-style chunk decoder), not a theorem derived from axioms. Ablations of interaction depth N and horizon H on LIBERO (§5.4) are ordinary hyperparameter selection, not fitted constants renamed as out-of-sample predictions. Citations (Grounding DINO for interaction init, ACT, DINOv3, BERT, π0.5 baselines) supply components or comparison points; none import a uniqueness theorem or self-proved premise that forces the central result. The paper is self-contained against external benchmarks. No self-definitional loop, no fitted-input-as-prediction, no load-bearing self-citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Load-bearing commitments are standard supervised imitation-learning assumptions plus the design choice that compact text encoding and bidirectional cross-attention suffice for execution-level instruction following. Free parameters are ordinary architecture/training hyperparameters selected on validation-style ablations, not physical constants. No new physical entities are postulated.

free parameters (4)
  • interaction depth N = 6
    Number of bidirectional fusion layers; ablated on LIBERO and fixed at 6 for main results.
  • action horizon H = 12 (LIBERO); 50 (RoboTwin)
    Length of predicted continuous action chunk; ablated and fixed at 12 for LIBERO main experiments.
  • shared hidden dimension d = 256
    Projection width for vision/language/state features before interaction and decoding.
  • learning rate and training schedule = 5e-5; 80k steps LIBERO / 55k RoboTwin
    Behavior-cloning optimization hyperparameters (5e-5, step counts, warmup, batch size) chosen for reported runs.
axioms (5)
  • domain assumption Expert demonstrations are an adequate supervision signal; minimizing L1 action-chunk error yields a competent closed-loop policy (behavior cloning).
    Stated training objective in §4.3; standard IL assumption underlying all reported success rates.
  • domain assumption Execution-level instructions specify objects, attributes, and relations sufficiently that a non-generative text encoder (e.g., BERT) captures the needed semantics.
    Core design premise in Intro and §4.1; if false for the target task distribution, removing the LLM would fail.
  • ad hoc to paper Bidirectional cross-attention between independently encoded vision and language features can build action-ready multimodal representations without an LLM latent space.
    Architectural thesis of §4.2; supported empirically by Table 5 but not derived from theory.
  • domain assumption LIBERO / RoboTwin clean / four real tasks are representative enough to support claims about efficient language-conditioned manipulation.
    Evaluation scope in §5; generalization beyond these regimes is assumed when discussing deployability.
  • standard math Standard transformer attention, residual fusion, and ACT-style parallel decoding are valid computational primitives.
    Used throughout §4 without modification to the underlying math.
invented entities (1)
  • TurboVLA V+L→A execution pathway independent evidence
    purpose: Name the proposed policy architecture that encodes vision and language separately, fuses them with lightweight bidirectional interaction, and decodes continuous action chunks without an LLM core.
    This is a system design, not a new physical entity. Independent evidence is the empirical benchmarks and promised code rather than an external predicted observable.

pith-pipeline@v1.2.0-grok45-kimik3 · 22708 in / 3807 out tokens · 79787 ms · 2026-07-30T10:54:04.146928+00:00 · methodology

0 comments
read the original abstract

Vision-language-action (VLA) models commonly adopt an LLM-centric $V \to L \to A$ pathway, where visual observations are projected into the representation space of a large language model before being decoded into robot actions. Although effective, this design incurs substantial computation and memory overhead at every policy invocation. In this work, we introduce TurboVLA, a new VLA paradigm that reformulates the conventional $V \to L \to A$ pathway as a direct $V + L \to A$ mapping. Instead of using a large language model as the central interface between perception and action, TurboVLA independently encodes visual observations and language instructions, directly exchanges information between them through lightweight bidirectional vision-language interaction, and predicts continuous action chunks with a compact decoder. This simple design constructs task-conditioned representations directly from visual and linguistic features, significantly reducing the computational and memory costs of VLA inference. On LIBERO, TurboVLA achieves 97.7% average success with only 0.2B parameters, 31.2 ms inference latency, and 0.9 GB inference VRAM on a consumer-grade RTX 4090, matching or outperforming substantially larger VLA policies. These results establish TurboVLA as a simple and effective alternative to the prevailing LLM-centric VLA paradigm, offering a new perspective on how vision, language, and action can be connected for efficient robotic manipulation. Code is available at https://github.com/H-EmbodVis/TurboVLA.

Figures

Figures reproduced from arXiv: 2607.27205 by Chenfei Yao, Dingkang Liang, Di Xu, Han Ding, Hengyi Xie, Xiang Bai, Xianjin Wu, Xuanyang Xi, Yingying Zhu, Yiping Tang.

Figure 1
Figure 1. Figure 1: TurboVLA enables compact local deployment and real-time language-conditioned manipulation with 0.2B parameters, 0.9 GB inference VRAM and 31.2 ms policy latency. ABSTRACT Vision-language-action (VLA) models commonly adopt an LLM-centric V → L → A pathway, where visual observations are projected into the repre￾sentation space of a large language model before being decoded into robot actions. Although effect… view at source ↗
Figure 2
Figure 2. Figure 2: From LLM-centric VLA to TurboVLA. (a) LLM-centric VLA predicts actions from large [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of TurboVLA. (a) TurboVLA simply yet efficiently maps visual observations [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world evaluation on the AgileX Piper platform. Left: our single-arm setup with a [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Vision-language interaction variants. (a) Directly concatenating visual and instruction [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of action horizon H on LIBERO. In this paper, we propose TurboVLA, a simple yet effi￾cient V + L → A paradigm that moves beyond the con￾ventional LLM-centered execution pathway for vision￾language-action learning. By combining lightweight in￾struction encoding, compact visual representations, bidi￾rectional vision-language interaction, and action-chunk decoding, TurboVLA preserves task-conditioned m… view at source ↗

Review history (2 revisions) →

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

92 extracted references · 13 linked inside Pith

  1. [1]

    Openvla: An open-source vision-language-action model , author=

  2. [2]

    Rt-1: Robotics transformer for real-world control at scale , author=

  3. [3]

    2023 , organization=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. 2023 , organization=

    2023

  4. [4]

    Fine-tuning vision-language-action models: Optimizing speed and success , author=

  5. [5]

    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=RSS, year=

  6. [6]

    Vla-adapter: An effective paradigm for tiny-scale vision-language-action model , author=

  7. [7]

    2024 , organization=

    Open x-embodiment: Robotic learning datasets and rt-x models , author=. 2024 , organization=

    2024

  8. [8]

    Octo: An open-source generalist robot policy , author=

  9. [9]

    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

  10. [10]

    Rdt-1b: a diffusion foundation model for bimanual manipulation , author=

  11. [11]

    arXiv preprint arXiv:2411.19650 , year=

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation , author=. arXiv preprint arXiv:2411.19650 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2605.30280 , year=

    Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments , author=. arXiv preprint arXiv:2605.30280 , year=

    Pith/arXiv arXiv

  13. [13]

    Learning fine-grained bimanual manipulation with low-cost hardware , author=

  14. [14]

    Diffusion policy: Visuomotor policy learning via action diffusion , author=

  15. [15]

    IEEE Robotics and Automation Letters , year=

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation , author=. IEEE Robotics and Automation Letters , year=

  16. [16]

    arXiv preprint arXiv:2504.19854 , year=

    Nora: A small open-sourced generalist vision language action model for embodied tasks , author=. arXiv preprint arXiv:2504.19854 , year=

    Pith/arXiv arXiv

  17. [17]

    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

  18. [18]

    arXiv preprint arXiv:2511.14659 , year=

    Nora-1.5: A vision-language-action model trained using world model-and action-based preference rewards , author=. arXiv preprint arXiv:2511.14659 , year=

    arXiv

  19. [19]

    2024 , organization=

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection , author=. 2024 , organization=

    2024

  20. [20]

    arXiv preprint arXiv:2405.10300 , year=

    Grounding dino 1.5: Advance the" edge" of open-set object detection , author=. arXiv preprint arXiv:2405.10300 , year=

    Pith/arXiv arXiv

  21. [21]

    Robopoint: A vision-language model for spatial affordance prediction for robotics , author=

  22. [22]

    Roboground: Robotic manipulation with grounded vision-language priors , author=

  23. [23]

    Roborefer: Towards spatial referring with reasoning in vision-language models for robotics , author=

  24. [24]

    Spatialvla: Exploring spatial representations for visual-language-action model , author=

  25. [25]

    Spatial forcing: Implicit spatial representation alignment for vision-language-action model , author=

  26. [26]

    From spatial to actions: Grounding vision-language-action model in spatial foundation priors , author=

  27. [27]

    arXiv preprint arXiv:2510.13778 , year=

    Internvla-m1: A spatially guided vision-language-action framework for generalist robot policy , author=. arXiv preprint arXiv:2510.13778 , year=

    Pith/arXiv arXiv

  28. [28]

    St4vla: Spatially guided training for vision-language-action models , author=

  29. [29]

    Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies , author=

  30. [30]

    Darabi, Nastaran and Trivedi, Amit Ranjan , booktitle=CVPRF, pages=

  31. [31]

    Restoring linguistic grounding in vla models via train-free attention recalibration , author=

  32. [32]

    arXiv preprint arXiv:2603.22760 , year=

    SG-VLA: Learning Spatially-Grounded Vision-Language-Action Models for Mobile Manipulation , author=. arXiv preprint arXiv:2603.22760 , year=

    arXiv

  33. [33]

    Univla: Learning to act anywhere with task-centric latent actions , author=

  34. [34]

    Unified vision-language-action model , author=

  35. [35]

    Language conditioned imitation learning over unstructured data , author=

  36. [36]

    Language-conditioned imitation learning for robot manipulation tasks , author=

  37. [37]

    2022 , organization=

    Cliport: What and where pathways for robotic manipulation , author=. 2022 , organization=

    2022

  38. [38]

    QVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization , author=

  39. [39]

    arXiv preprint arXiv:2506.07530 , year=

    Bitvla: 1-bit vision-language-action models for robotics manipulation , author=. arXiv preprint arXiv:2506.07530 , year=

    arXiv

  40. [40]

    Efficientvla: Training-free acceleration and compression for vision-language-action models , author=

  41. [41]

    Specprune-vla: Accelerating vision-language-action models via action-aware self-speculative pruning , author=

  42. [42]

    Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation , author=

  43. [43]

    IEEE Robotics and Automation Letters , year=

    Rlrc: Reinforcement learning-based recovery for compressed vision-language-action models , author=. IEEE Robotics and Automation Letters , year=

  44. [44]

    Shallow- : Knowledge Distillation for Flow-based VLAs , author=

  45. [45]

    IEEE Robotics and Automation Letters , volume=

    Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  46. [46]

    2022 , organization=

    Bc-z: Zero-shot task generalization with robotic imitation learning , author=. 2022 , organization=

    2022

  47. [47]

    VIMA: General Robot Manipulation with Multimodal Prompts , author=

  48. [48]

    2023 , organization=

    Perceiver-actor: A multi-task transformer for robotic manipulation , author=. 2023 , organization=

    2023

  49. [49]

    CogVLA: Cognition-aligned vision-language-action models via instruction-driven routing & sparsification , author=

  50. [50]

    Libero: Benchmarking knowledge transfer for lifelong robot learning , author=

  51. [51]

    Palm-e: An embodied multimodal language model , author=

  52. [52]

    Do as i can, not as i say: Grounding language in robotic affordances , author=

  53. [53]

    Inner monologue: Embodied reasoning through planning with language models , author=

  54. [54]

    Fast: Efficient action tokenization for vision-language-action models , author=

  55. [55]

    2021 , organization=

    Learning transferable visual models from natural language supervision , author=. 2021 , organization=

    2021

  56. [56]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  57. [57]

    Evo-1: Lightweight vision-language-action model with preserved semantic alignment , author=

  58. [58]

    Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies , author=

  59. [59]

    Mantis: A versatile vision-language-action model with disentangled visual foresight , author=

  60. [60]

    Mm-act: Learn from multimodal parallel generation to act , author=

  61. [61]

    Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models , author=

  62. [62]

    Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge , author=

  63. [63]

    IEEE Robotics and Automation Letters , volume=

    What matters in language conditioned robotic imitation learning over unstructured data , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  64. [64]

    Robomamba: Efficient vision-language-action model for robotic reasoning and manipulation , author=

  65. [65]

    Vla-cache: Efficient vision-language-action manipulation via adaptive token caching , author=

  66. [66]

    arXiv preprint arXiv:2509.12594 , year=

    The better you learn, the smarter you prune: Towards efficient vision-language-action models via differentiable token pruning , author=. arXiv preprint arXiv:2509.12594 , year=

    arXiv

  67. [67]

    Real-time execution of action chunking flow policies , author=

  68. [68]

    arXiv preprint arXiv:2603.19199 , year=

    Faster: Rethinking real-time flow vlas , author=. arXiv preprint arXiv:2603.19199 , year=

    Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2605.13778 , year=

    Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs , author=. arXiv preprint arXiv:2605.13778 , year=

    Pith/arXiv arXiv

  70. [70]

    arXiv preprint arXiv:2508.10104 , year=

    Dinov3 , author=. arXiv preprint arXiv:2508.10104 , year=

    Pith/arXiv arXiv

  71. [71]

    Vla-jepa: Enhancing vision-language-action model with latent world model , author=

  72. [72]

    Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding , author=

  73. [73]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation , author=

  74. [74]

    Sigmoid loss for language image pre-training , author=

  75. [75]

    Findings of the Association for Computational Linguistics , pages=

    Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models , author=. Findings of the Association for Computational Linguistics , pages=

  76. [76]

    PoseVLA: Universal Pose Pretraining for Generalizable Vision-Language-Action Policies , author=

  77. [77]

    Pointact: Vision-language-action models with multi-scale point-action interaction , author=

  78. [78]

    OAT: Ordered Action Tokenization , author=

  79. [79]

    arXiv preprint arXiv:2604.05014 , year=

    StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing , author=. arXiv preprint arXiv:2604.05014 , year=

    Pith/arXiv arXiv

  80. [80]

    arXiv preprint arXiv:2603.05117 , year=

    SeedPolicy: Horizon Scaling via Self-Evolving Diffusion Policy for Robot Manipulation , author=. arXiv preprint arXiv:2603.05117 , year=

    Pith/arXiv arXiv

Showing first 80 references.