Pith. sign in

REVIEW 3 major objections 5 minor 38 references

A 0.5B vision-language-action model that treats robot actions as a 2D time-by-dimension token grid, with structured masking and multi-layer bridging, beats much larger models on long-horizon manipulation.

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

2026-07-14 15:18 UTC pith:3DAQF37Y

load-bearing objection Solid small-model VLA recipe with a clear 2D action-masking idea and strong LIBERO/CALVIN numbers; novelty is compositional, and the long-horizon claim needs variance and matched controls before you lean on it hard. the 3 major comments →

arxiv 2607.09818 v1 pith:3DAQF37Y submitted 2026-07-10 cs.RO cs.AIcs.CV

TS-Mask VLA: 2D Temporal-Spatial Masking for Vision-Language-Action Model with Effective Bridging

Shengzhuo Yang , Ronghao Yu , Chuanjie Lv , Linpeng Peng , Hang Yu , Jie Ren , Jiajun Lv , Yong Liu This is my paper
classification cs.RO cs.AIcs.CV
keywords vision-language-actiondiscrete diffusionrobot manipulationtemporal-spatial maskingBridge Attentionaction tokenizationlong-horizon control
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 systems turn robot control into next-token prediction or continuous trajectory denoising. This paper argues that those choices miss the natural structure of actions—dependencies across time steps and across joint dimensions—and blur the boundary between understanding scenes and producing control. TS-Mask VLA discretizes continuous actions into a fixed grid of tokens over time and action dimensions, applies a two-stage temporal-then-spatial mask, and recovers the masked tokens with a dedicated discrete diffusion expert. That expert is conditioned, layer by layer, by a small vision-language backbone through Bridge Attention, which mixes action self-tokens, action-query tokens, and task tokens. The result is a 0.5-billion-parameter policy that reaches 95.7% average success on LIBERO and the highest average sequence length on CALVIN, while also working on real-robot tasks. A sympathetic reader cares because the work claims that explicit spatiotemporal structure and clean conditioning, not parameter count alone, are what make long-horizon robot behavior reliable.

Core claim

Discretizing robot actions into a time-by-dimension token grid, training with structured temporal-spatial 2D masking, and generating those tokens via a discrete diffusion action expert that receives multi-layer vision-language features through Bridge Attention yields more structurally consistent action sequences than autoregressive decoding or continuous diffusion, enabling a 0.5B model to outperform substantially larger VLAs on LIBERO and CALVIN long-horizon benchmarks.

What carries the argument

Temporal-spatial 2D masking of discrete action tokens, paired with Bridge Attention multi-layer conditioning of a discrete diffusion action expert: the 2D mask forces recovery of cross-time and cross-dimension structure, while Bridge Attention injects hierarchical vision-language signals without collapsing representation learning into the policy.

Load-bearing premise

The paper assumes that splitting each continuous action coordinate into 256 fixed uniform bins keeps enough precision and structure that 2D masked discrete diffusion on that grid can still produce accurate robot control.

What would settle it

Keep the same backbone and Bridge Attention, replace the 256-bin uniform tokenizer with continuous diffusion or finer learned quantization, and re-measure LIBERO Long and CALVIN 4- and 5-task completion rates; a clear drop relative to the reported 91.6% and 66.9% would show the gains depend on the discrete 2D grid rather than bridging alone.

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

If this is right

  • Sub-billion-parameter VLAs can match or beat multi-billion-parameter systems on standard manipulation benchmarks when action structure is modeled explicitly.
  • Structured 2D temporal-spatial masking improves long-horizon completion more than ordinary 1D token masking.
  • Layer-aligned multi-layer conditioning from intermediate VLM states can replace reliance on final-layer features alone for control.
  • Discrete diffusion with remasking inference is a practical alternative to autoregressive action decoding for robot policies.
  • Real-world deployment of tiny VLAs becomes more plausible if the same structure-and-bridging gains transfer outside simulation.

Where Pith is reading between the lines

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

  • If 2D action structure is the main inductive bias, analogous masking over multi-finger or multi-agent action maps should transfer without full retraining of the vision-language backbone.
  • Uniform 256-bin quantization may become a ceiling on high-precision tasks; adaptive or learned binning is a natural next lever once masking and bridging are fixed.
  • The unroll loss that closes the train-test gap points to a general need for multi-step consistency objectives whenever discrete diffusion is used for sequential control.
  • The parameter efficiency shown here suggests future VLA progress may shift from scaling the language backbone toward better action-space inductive biases.

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 / 5 minor

Summary. The paper proposes TS-Mask VLA, a discrete vision–language–action framework for robot manipulation. A lightweight Qwen2.5-0.5B VLM encodes third-view and wrist images plus language (with an Action Query token); multi-layer hidden states are injected into a dedicated Discrete Diffusion Action Expert via Bridge Attention that fuses self, AQ, and task streams with a learnable task gate. Continuous actions are uniformly quantized into V=256 bins and reshaped into a T×D temporal–spatial grid; a two-stage 2D masking strategy (full-frame temporal masks then per-frame spatial masks under a cosine schedule) trains masked token recovery, with an optional step-unroll loss. Inference starts from a fully masked grid and iteratively remasks low-confidence tokens. On LIBERO the 0.5B model reports 95.7% average success (Spatial/Object/Goal/Long), and on CALVIN ABC→D it reports the best average sequence length 4.19; real-world UR5e trials and ablations of 2D vs 1D masking and unroll strength are also presented.

Significance. If the reported gains hold under proper statistical controls, the work is a useful systems contribution for resource-constrained VLA: it shows that a tiny backbone plus a discrete diffusion action expert and structured 2D masking can match or exceed much larger autoregressive and continuous-diffusion VLAs on standard long-horizon benchmarks, with favorable parameter efficiency and real-robot transfer. The Bridge Attention multi-layer conditioning and the explicit temporal–spatial inductive bias are concrete, implementable design choices that the community can reuse. Strengths include clear ablations that move in the expected direction (especially +6.6% on LIBERO-Long for 2D vs 1D) and evaluation on both LIBERO and CALVIN plus three real-world tasks. The paper does not claim theoretical novelty beyond the architectural combination; its value is empirical and engineering.

major comments (3)
  1. [Tables I–III; Sec. IV-C/D; Fig. 5] Tables I–II and the real-world results (Fig. 5) report point success rates without error bars, seed counts, or trial N (only real-world states N=20). The central claim that Bridge Attention + 2D masking yields superior long-horizon performance (LIBERO-Long 91.6%, CALVIN avg. len. 4.19) therefore cannot be distinguished from training-seed or hyperparameter variance. At minimum the authors should report multi-seed means±std (or bootstrap CIs) for the main suites and for the 2D-vs-1D ablation in Table III, and state the number of evaluation episodes per suite.
  2. [Sec. III-C.2; Table III; Sec. IV-D] Table III shows a large Long-suite gain for 2D vs 1D masking, but does not isolate whether that gain survives under a matched continuous-diffusion or finer-bin / VQ baseline with the same Bridge Attention backbone and training budget. Sec. III-C.2 asserts that uniform V=256 scalar quantization preserves fine-grained temporal and inter-dimensional structure better than VQ-VAE, yet no quantization-error or control-precision analysis is given. Without such a control, the attribution of long-horizon superiority specifically to the 2D discrete inductive bias remains under-supported.
  3. [Table I; Sec. IV-B] Several strong baselines in Table I are marked * (reproduced under the “same setting”) or † (non-VLM). The manuscript does not specify which hyperparameters, data mixtures, action chunk lengths, or evaluation protocols were matched. For a claim of outperforming π0 / GR00T N1 / OpenVLA-OFT with 14–19× fewer parameters, the reproduction protocol and any deviations must be stated explicitly so that the comparison is interpretable.
minor comments (5)
  1. [Sec. III-C.1, Eq. (3)] Eq. (3) uses tanh(g) on the task branch; the range and initialization of g, and whether it is per-layer or shared, are not stated.
  2. [Sec. III-E; Sec. IV-A] Inference remask schedule γ(i/I) and iteration count I appear in Sec. III-E but no default values or sensitivity are given; likewise action chunk length is fixed at 8 without justification.
  3. [Figs. 1–2] Fig. 1 and Fig. 2 captions are dense; a short legend clarifying [MASK] vs predicted tokens and the two-stage mask order would improve readability.
  4. [Sec. II] Related Work cites discrete diffusion VLAs (Liang et al., Wen et al.) only briefly; a clearer positioning of how Bridge Attention + 2D masking differs from those concurrent discrete-diffusion action heads would help.
  5. [Abstract; throughout] Minor typos: “andand” / duplicated “and” in the abstract contributions list; inconsistent spacing around “temporal–spatial”.

Circularity Check

0 steps flagged

No circularity: empirical VLA systems paper whose success rates are external benchmark metrics, not quantities defined by the training objective or self-cited uniqueness claims.

full rationale

TS-Mask VLA proposes two architectural components (Bridge Attention multi-layer conditioning of a discrete diffusion action expert, and temporal–spatial 2D masking of uniformly quantized action tokens) and evaluates them by success rate / average sequence length on the public LIBERO and CALVIN suites plus real-robot trials. The reported numbers (95.7 % LIBERO avg, CALVIN avg length 4.19) are measured against held-out task suites; they are not algebraic rearrangements of fitted free parameters. Ablations (1D vs 2D mask, unroll strength λ) compare design choices on the same external metrics rather than recovering quantities that were already used to set those choices. Citations to VLA-Adapter, discrete diffusion, DINOv2/SigLIP, and Qwen2.5 are ordinary prior-art references by distinct author groups; none supply a load-bearing uniqueness theorem that forces the present architecture. Uniform 256-bin quantization is an explicit modeling choice whose fidelity is an empirical assumption, not a circular definition of the success metric. Consequently the derivation chain contains no self-definitional step, no fitted-input-called-prediction, and no self-citation that collapses the central claim.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim is empirical performance of a designed architecture, not a theorem. Load-bearing modeling choices include uniform action binning, the 2D mask schedule, multi-layer Bridge Attention fusion, and several training/inference hyperparameters. Background assumptions are standard robot-learning practice (chunked actions, LoRA fine-tuning of a small VLM, discrete diffusion as a generative prior). Invented entities are architectural modules rather than physical objects; independent evidence is only the paper’s own ablations and benchmarks.

free parameters (6)
  • action vocabulary size V (uniform bins)
    Fixed at 256 bins over [-1,1]; bin count is a design choice that sets quantization granularity and is not derived from data or theory.
  • unroll loss weight λ
    Ablated at 0 / 0.5 / 1.0; best reported setting λ=0.5 on LIBERO-Spatial, so the training objective balance is tuned to validation performance.
  • masking ratio schedule r=cos(πt/2)
    Cosine schedule and two-stage temporal-then-spatial masking ratios are chosen design knobs controlling corruption severity.
  • action chunk length T
    Set to 8 by default; defines the temporal extent of the 2D action grid.
  • inference remask schedule γ(i/I) and iteration count I
    Coarse-to-fine ReMask keeps high-confidence tokens; exact I and γ shape are free inference hyperparameters affecting final actions.
  • task-branch gate g in Bridge Attention
    Learnable scalar modulating task keys; capacity and early-training behavior depend on this free gate.
axioms (5)
  • domain assumption Robot actions can be treated as a T×D discrete token grid after independent per-dimension uniform quantization without needing VQ-VAE compression.
    Stated in Sec. III-C.2 as justification for rejecting VQ-VAE; underpins the entire 2D masking design.
  • domain assumption Layer-aligned multi-layer VLM hidden states plus Action Query tokens provide sufficient conditioning for a separate discrete diffusion action expert.
    Core architectural premise in Sec. III-B/C, adapted from VLA-Adapter-style bridging.
  • domain assumption Masked discrete diffusion with parallel remasking is a valid generative model for action tokens under vision-language conditioning.
    Imports discrete diffusion practice from language/image literature (cited) into control; not re-proved here.
  • domain assumption Standard LIBERO and CALVIN success metrics and the authors’ real-world trial protocol are adequate proxies for manipulation competence.
    Evaluation sections treat these benchmarks as the primary evidence for the central claim.
  • standard math Transformer attention, residual FFNs, RoPE, and cross-entropy over categorical tokens behave as in standard deep learning practice.
    Used throughout Sec. III without novel proof obligations.
invented entities (2)
  • Discrete Diffusion Action Expert with Bridge Attention no independent evidence
    purpose: Separate action generator that fuses self, AQ, and task streams under multi-layer VLM conditioning to produce discrete action tokens.
    Named module combining discrete diffusion with a three-stream attention bridge; Bridge Attention is adapted from VLA-Adapter rather than independently evidenced outside this line of work.
  • Temporal–Spatial 2D Token Masking (plus ReMask inference) no independent evidence
    purpose: Structured corruption of the T×D action grid to force learning of cross-time and inter-dimension dependencies, with confidence-based remasking at inference.
    Primary methodological contribution; support is internal ablations (1D vs 2D) and benchmark gains, not external independent measurements.

pith-pipeline@v1.1.0-grok45 · 17362 in / 3626 out tokens · 33109 ms · 2026-07-14T15:18:28.105739+00:00 · methodology

0 comments
read the original abstract

Vision-language-action (VLA) models aim to understand natural-language instructions and visual observations, and to generate and execute corresponding actions as embodied agents. Recently, autoregressive token-based action generation has driven the development of many representative VLA models. However, this paradigm often reduces action generation to next-token prediction, thereby lacking explicit modeling of the spatiotemporal structure of action sequences and the disentanglement between vision-language representations and actions, which can limit performance in long-horizon and complex scenarios. In this paper, we propose TS-Mask VLA, a vision-language-action framework for robot manipulation. TS-Mask VLA is built upon two key designs: (1) a Discrete Diffusion Action Expert equipped with a Bridge Attention conditioning bridge, which enables multi-layer conditioning from the VLM and facilitates more accurate and stable action generation; and (2) a temporal-spatial 2D masking strategy for discrete action tokens that strengthens the model's understanding of cross-time dependencies and inter-dimensional coupling, leading to more structurally consistent action sequences. We conduct extensive experiments on simulation benchmarks and real-world tasks. On LIBERO, TS-Mask VLA achieves a 95.7 percent average success rate with only 0.5B parameters, outperforming significantly larger models. On CALVIN, it attains the best average sequence length of 4.19 and strong long-horizon performance. Comprehensive analyses and ablations further validate the effectiveness of our design.

Figures

Figures reproduced from arXiv: 2607.09818 by Chuanjie Lv, Hang Yu, Jiajun Lv, Jie Ren, Linpeng Peng, Ronghao Yu, Shengzhuo Yang, Yong Liu.

Figure 1
Figure 1. Figure 1: Overview of TS-Mask VLA. The input observation image and task description are encoded by a visual backbone and a text tokenizer, and together with Action Query fed into an N-layer VLM to produce hidden states, where we adopt Qwen2.5-0.5B as the backbone. The VLM hidden states are provided as key/value to a Discrete Diffusion Action Expert composed of N stacked blocks of Bridge Attention and FFN. On the act… view at source ↗
Figure 2
Figure 2. Figure 2: Temporal–Spatial 2D Token Masking. Given a 2D action-token map of size T × D, we adopt a two-stage masking strategy. We first perform Temporal masking (left), where several time frames are randomly selected and all action tokens within the selected frames are masked. Then, Spatial masking (right) is applied to the remaining unmasked frames, where a subset of action tokens within each frame is randomly mask… view at source ↗
Figure 3
Figure 3. Figure 3: Inference procedure of TS-Mask VLA. Given an input image, task description, and Action Query, the VLM encodes multimodal context and conditions the Discrete Diffusion Action Expert. Starting from a fully masked action-token sequence at t0, the model predicts token distributions and produces partially denoised tokens at t1. A ReMask strategy is then applied to re-mask low￾confidence tokens while preserving … view at source ↗
Figure 4
Figure 4. Figure 4: All Evaluation environments, We conduct compre￾hensive evaluations of TS-Mask VLA in both simulation and real￾world settings. In simulation, evaluations are performed on the LIBERO benchmark and CALVIN benchmark. In the real-world experiment, we set up three experimental setups. figure shows our physical experiment. TABLE III Ablation on Masking Strategy (Success Rate %),Bold indicates the best performance… view at source ↗
Figure 5
Figure 5. Figure 5: Real-World Settings and results. with effective Bridging. Our approach effectively in￾jects multimodal knowledge into an discrete diffusion action expert and employs a spatio-temporal 2D masking strategy to strengthen temporal dependencies and cross￾dimensional coupling in action sequences. The resulting discrete diffusion action expert enables accurate and con￾sistent action generation. Despite using a ti… 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

38 extracted references · 24 linked inside Pith

  1. [1]

    Discrete diffusion vla: Bring- ing discrete diffusion to action decoding in vision-language- action policies,

    Z. Liang, Y. Li, T. Yang, C. Wu, S. Mao, T. Nian, L. Pei, S. Zhou, X. Yang, J. Panget al., “Discrete diffusion vla: Bring- ing discrete diffusion to action decoding in vision-language- action policies,”arXiv preprint arXiv:2508.20072, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Llada- vla: Vision language diffusion action models,

    Y. Wen, H. Li, K. Gu, Y. Zhao, T. Wang, and X. Sun, “Llada- vla: Vision language diffusion action models,”arXiv preprint arXiv:2509.06932, 2025

    Pith/arXiv arXiv 2025

  3. [3]

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

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Bal- akrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. San- ketiet al., “Openvla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    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,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  5. [5]

    Smolvla: A vision-language-action model for affordable and efficient robotics,

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafiotiet al., “Smolvla: A vision-language-action model for affordable and efficient robotics,”arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Multimodal diffusion transformer: Learning versatile behav- ior from multimodal goals,

    M. Reuss, Ö. E. Yağmurlu, F. Wenzel, and R. Lioutikov, “Multimodal diffusion transformer: Learning versatile behav- ior from multimodal goals,”arXiv preprint arXiv:2407.05996, 2024

    Pith/arXiv arXiv 2024

  7. [7]

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

    Y. Wang, P. Ding, L. Li, C. Cui, Z. Ge, X. Tong, W. Song, H. Zhao, W. Zhao, P. Houet al., “Vla-adapter: An effective paradigm for tiny-scale vision-language-action model,”arXiv preprint arXiv:2509.09372, 2025

    arXiv 2025

  8. [8]

    Intermask: 3d human interaction generation via collaborative masked modeling,

    M. G. Javed, C. Guo, L. Cheng, and X. Li, “Intermask: 3d human interaction generation via collaborative masked modeling,”arXiv preprint arXiv:2410.10010, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Libero: Benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learning,”Advances in Neural Information Processing Sys- tems, vol. 36, pp. 44776–44791, 2023

    2023

  10. [10]

    Calvin: A benchmark for language-conditioned policy learn- ingforlong-horizonrobotmanipulationtasks,

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard, “Calvin: A benchmark for language-conditioned policy learn- ingforlong-horizonrobotmanipulationtasks,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7327–7334, 2022

    2022

  11. [11]

    π0: A vision-language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “π0: A vision-language-action flow model for general robot control,” arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  12. [12]

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

    A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu et al., “Rt-1: Robotics transformer for real-world control at scale,”arXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Rt-2: Vision- language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision- language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    2023

  14. [14]

    Structured denoising diffusion models in discrete state-spaces,

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. Van Den Berg, “Structured denoising diffusion models in discrete state-spaces,”Advances in neural information processing sys- tems, vol. 34, pp. 17981–17993, 2021

    2021

  15. [15]

    Argmax flows and multinomial diffusion: Learning categori- cal distributions,

    E. Hoogeboom, D. Nielsen, P. Jaini, P. Forré, and M. Welling, “Argmax flows and multinomial diffusion: Learning categori- cal distributions,”Advances in neural information processing systems, vol. 34, pp. 12454–12465, 2021

    2021

  16. [16]

    Vector quantized diffusion model for text-to- image synthesis,

    S. Gu, D. Chen, J. Bao, F. Wen, B. Zhang, D. Chen, L. Yuan, and B. Guo, “Vector quantized diffusion model for text-to- image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10696– 10706

    2022

  17. [17]

    Diffusion-lm improves controllable text gener- ation,

    X. Li, J. Thickstun, I. Gulrajani, P. S. Liang, and T. B. Hashimoto, “Diffusion-lm improves controllable text gener- ation,”Advances in neural information processing systems, vol. 35, pp. 4328–4343, 2022

    2022

  18. [18]

    Maskgit: Masked generative image transformer,

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman, “Maskgit: Masked generative image transformer,” inProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11315–11325

    2022

  19. [19]

    Large language diffusion models,

    S.Nie,F.Zhu,Z.You,X.Zhang,J.Ou,J.Hu,J.Zhou,Y.Lin, J.-R. Wen, and C. Li, “Large language diffusion models,” arXiv preprint arXiv:2502.09992, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Dinov2: Learning robust visual features without super- vision,

    M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., “Dinov2: Learning robust visual features without super- vision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Sig- moid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sig- moid loss for language image pre-training,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 11975–11986

    2023

  22. [22]

    Qwen2.5-coder technical report,

    B. Hui, J. Yang, Z. Cui, J. Yang, D. Liu, L. Zhang, T. Liu, J. Zhang, B. Yu, K. Luet al., “Qwen2.5-coder technical report,”arXiv preprint arXiv:2409.12186, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Flowvla: Thinking in motion with a visual chain of thought,

    Z. Zhong, H. Yan, J. Li, X. Liu, X. Gong, W. Song, J. Chen, and H. Li, “Flowvla: Thinking in motion with a visual chain of thought,”arXiv e-prints, pp. arXiv–2508, 2025

    2025

  24. [24]

    Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,

    Q. Zhao, Y. Lu, M. J. Kim, Z. Fu, Z. Zhang, Y. Wu, Z. Li, Q. Ma, S. Han, C. Finnet al., “Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1702–1713

    2025

  25. [25]

    Thinkact: Vision-language-action reason- ing via reinforced visual latent planning,

    C.-P. Huang, Y.-H. Wu, M.-H. Chen, Y.-C. F. Wang, and F.-E. Yang, “Thinkact: Vision-language-action reason- ing via reinforced visual latent planning,”arXiv preprint arXiv:2507.16815, 2025

    Pith/arXiv arXiv 2025

  26. [26]

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

    Q. Bu, Y. Yang, J. Cai, S. Gao, G. Ren, M. Yao, P. Luo, and H. Li, “Univla: Learning to act anywhere with task-centric latent actions,”arXiv preprint arXiv:2505.06111, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Fine-tuning vision- language-actionmodels:Optimizingspeedandsuccess,

    M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision- language-actionmodels:Optimizingspeedandsuccess,”arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Towards synergistic, generalized, and effi- cient dual-system for robotic manipulation,

    Q. Bu, H. Li, L. Chen, J. Cai, J. Zeng, H. Cui, M. Yao, and Y. Qiao, “Towards synergistic, generalized, and effi- cient dual-system for robotic manipulation,”arXiv preprint arXiv:2410.08001, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Openhelix: A short survey, empirical analysis, and open-source dual-system vla model for robotic manipulation,

    C. Cui, P. Ding, W. Song, S. Bai, X. Tong, Z. Ge, R. Suo, W. Zhou, Y. Liu, B. Jiaet al., “Openhelix: A short survey, empirical analysis, and open-source dual-system vla model for robotic manipulation,”arXiv preprint arXiv:2505.03912, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Reconvla: Re- constructive vision-language-action model as effective robot perceiver,

    W. Song, Z. Zhou, H. Zhao, J. Chen, P. Ding, H. Yan, Y. Huang, F. Tang, D. Wang, and H. Li, “Reconvla: Re- constructive vision-language-action model as effective robot perceiver,”arXiv preprint arXiv:2508.10333, 2025

    Pith/arXiv arXiv 2025

  31. [31]

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

    K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine, “Fast: Efficient action tokenization for vision-language-action models,”arXiv preprint arXiv:2501.09747, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Gr00t n1: An open foundation model for generalist humanoid robots,

    J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huanget al., “Gr00t n1: An open foundation model for generalist humanoid robots,” arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Deer-vla: Dynamic inference of multimodal large language models for efficient robot execution,

    Y. Yue, Y. Wang, B. Kang, Y. Han, S. Wang, S. Song, J. Feng, and G. Huang, “Deer-vla: Dynamic inference of multimodal large language models for efficient robot execution,”Advances in Neural Information Processing Systems, vol. 37, pp. 56619– 56643, 2024

    2024

  34. [34]

    Vision-language foun- dation models as effective robot imitators,

    X. Li, M. Liu, H. Zhang, C. Yu, J. Xu, H. Wu, C. Cheang, Y. Jing, W. Zhang, H. Liuet al., “Vision-language foun- dation models as effective robot imitators,”arXiv preprint arXiv:2311.01378, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Zero-shot robotic manipulation with pretrained image-editing diffusion models,

    K. Black, M. Nakamoto, P. Atreya, H. Walke, C. Finn, A. Ku- mar, and S. Levine, “Zero-shot robotic manipulation with pretrained image-editing diffusion models,”arXiv preprint arXiv:2310.10639, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Predictive inverse dynamics models are scalable learners for robotic manipulation,

    Y. Tian, S. Yang, J. Zeng, P. Wang, D. Lin, H. Dong, and J. Pang, “Predictive inverse dynamics models are scalable learners for robotic manipulation,”arXiv preprint arXiv:2412.15109, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Vla-os: Structuring and dissecting planning representations and paradigms in vision- language-action models,

    C. Gao, Z. Liu, Z. Chi, J. Huang, X. Fei, Y. Hou, Y. Zhang, Y. Lin, Z. Fang, Z. Jianget al., “Vla-os: Structuring and dissecting planning representations and paradigms in vision- language-action models,”arXiv preprint arXiv:2506.17561, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Efficient diffusion transformer policies with mixture of expert denois- ers for multitask learning,

    M. Reuss, J. Pari, P. Agrawal, and R. Lioutikov, “Efficient diffusion transformer policies with mixture of expert denois- ers for multitask learning,”arXiv preprint arXiv:2412.12953, 2024

    Pith/arXiv arXiv 2024