Pith. sign in

REVIEW 3 major objections 5 minor 40 references

CAC-VLA: Context-Gated Action Conditioning for Vision-Language-Action Models

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read A vision-language model can guide continuous robot control by predicting latent future-action codes and gating how strongly they condition the action expert.

desk verdict Solid VLA engineering: VLM-native OAT latents plus a context gate, strong LIBERO-Plus and real-robot deltas, but train-time teacher latents and tiny saturated-suite gains keep the claim conditional. read the letter →

arxiv 2607.04816 v1 pith:674YNXJI submitted 2026-07-06 cs.RO

classification cs.RO
keywords VLAActionConditioningRoboticManipulationLatentActionsContextGateVision-Language-ActionLIBERO
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Vision-language-action systems usually hand multimodal features to a continuous action expert even though those features were never built for motor control. This paper argues the vision-language backbone itself can learn a lightweight latent-action interface: dedicated query tokens predict compact coarse-to-fine codes of upcoming action segments, and a context gate then decides residual strength when those codes condition the expert. On standard multi-task tabletop suites the method reaches 98.3 percent average success and 89.5 percent under distribution shifts, with supporting real-world pick-and-place gains. A sympathetic reader cares because the design supplies action-structured guidance without a separate trajectory generator and without treating intermediate signals as fixed commands.

What carries the argument

Context-gated latent-action conditioning: the VLM predicts raw ordered-action-tokenizer codes of future segments; expert layers retrieve them by cross-attention on current action hidden states; a channel-wise gate computed from pooled action state and update sets residual injection strength so guidance is calibrated rather than fixed.

What would settle it

Retrain the identical architecture with latent alignment removed or with random latent targets instead of ordered-action encodings; if LIBERO Object, Goal, and Long success rates remain statistically indistinguishable from the full model, the claimed benefit of context-gated latent-action conditioning is falsified.

Watch

Extended reading notes

Core claim

The authors claim that training the vision-language model to predict ordered latent actions encoded from future robot action segments, then adaptively injecting those predictions into the continuous action expert through a context gate, forms an effective VLM-native interface for expert control, evidenced by 98.3 percent average success on LIBERO and 89.5 percent on LIBERO-Plus under supervised fine-tuning.

Load-bearing premise

The scheme assumes that fixed-horizon raw codes of future actions from a frozen tokenizer are both predictable from vision and language and informative enough that residual guidance measurably helps the continuous expert.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CAC-VLA, a VLA architecture that equips a pretrained VLM with learnable query tokens that predict coarse-to-fine latent actions (raw latents from a frozen ordered action tokenizer applied to future action segments) and injects those latents into a flow-matching action expert via cross-attention residual updates modulated by a learned context gate. Training uses ground-truth OAT latents both as alignment targets (Smooth-L1) and as the expert conditioning source; at inference the expert is conditioned only on VLM predictions. On LIBERO the method reports 98.3% average success (second to ACoT-VLA), on LIBERO-Plus supervised fine-tuning 89.5% (best among reported methods), with ablations on latent horizon and gating and a small matched real-robot pick-and-place / stacking comparison against π0.5.

Significance. If the gains are attributable to the deployed VLM-native interface rather than to privileged training-time conditioning, the work supplies a lightweight, architecture-compatible way to give continuous action experts action-structured residual guidance without a separate trajectory reasoner. The LIBERO-Plus and real-robot results are the more informative empirical contributions; the design is simple enough to be adopted by other VLA stacks. Strengths include clear train/inference separation of the OAT source, horizon and gate ablations, gate-behavior analysis in the appendix, and a fair real-robot protocol with matched data and hyperparameters.

major comments (3)
  1. §3.3 Eq. (7) and §3.4: during training the expert is conditioned on ground-truth OAT latents zoat, while at inference it receives only VLM predictions ˆz. The alignment loss (Eq. 6) is therefore not the sole path by which future-action structure reaches the expert; the expert itself is trained under perfect latent conditioning. Table 4’s modest ablation deltas (98.3 → 98.0 / 97.9) and the near-saturated LIBERO numbers do not isolate how much of the reported performance is due to the VLM-native interface that is actually deployed. A controlled experiment that trains the expert under predicted (or noised / dropout-only) latents, or reports train-vs-inference latent error correlation with success, is needed to support the central claim.
  2. Table 1 / Table 4: absolute gains on LIBERO are small (0.3–0.4 points over the no-latent / no-gate variants; second place overall). The strongest numerical support is therefore LIBERO-Plus SFT (Table 2, 89.5%) and the real-robot comparison (Fig. 2). The abstract and §4.2 should not present the 98.3% LIBERO figure as primary evidence of effectiveness without acknowledging saturation and the limited ablation margin; otherwise the central claim is overstated relative to the load-bearing tables.
  3. §3.2 Eqs. (3)–(6) and App. A: the informativeness of frozen OAT raw latents (Dz=4, fixed Hl) as a predictable intermediate is assumed rather than validated. There is no report of latent prediction error on held-out trajectories, no comparison to alternative latent constructions (e.g., learned autoencoder, discrete codes, or longer/shorter hierarchical horizons), and no analysis of when ˆz is inaccurate relative to task phase. Without this, it remains unclear whether the interface is robust or merely adequate on the evaluated suites.
minor comments (5)
  1. Table 2 caption: clarify which π0.5 / π∗0.5 entries are author-reproduced versus taken from prior work; the asterisk note is easy to miss.
  2. Fig. 1(c) and Eq. (11): state explicitly that the gate is channel-wise and shared across action tokens; the figure alone is slightly ambiguous.
  3. App. C / Fig. 2: report confidence intervals or binomial standard errors for the 25-trial real-robot rates; 16/25 vs 4/25 is large but still small-N.
  4. Related Work §2.2–2.3: a short explicit comparison table (guidance source, train-time teacher forcing, adaptive vs fixed fusion) would help situate CAC-VLA against ACoT-VLA, LAPA, and UniVLA.
  5. Typos / polish: abstract “89.5% LIBERO-Plus” missing “on”; consistent use of CAC-VLA vs CAC-VLA spacing; ensure all arXiv citations that are concurrent are marked as such.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard supervised latent-alignment + teacher-forced conditioning with explicit train/inference split; empirical success rates are not forced by construction.

full rationale

CAC-VLA is an empirical VLA architecture paper. Latent actions are defined by a frozen external OAT tokenizer applied to future ground-truth action segments (Eq. 3); the VLM is trained to regress those latents via Smooth-L1 (Eqs. 5–6) and the expert is conditioned on them. During training the expert sees the OAT target zoat; at inference it sees only the VLM prediction ˆz (Eq. 7). This is ordinary teacher-forcing / supervised imitation, not a self-definitional loop: the quantity being predicted (ˆz) is not defined in terms of the reported success rates, nor are the LIBERO / LIBERO-Plus numbers obtained by fitting a free parameter that is then re-labeled a “prediction.” Ablations (Tables 3–4) and the explicit train/inference distinction further show that the central claim is an empirical claim about residual guidance, not a tautology. Self-citations (π0.5 baseline, OAT, ACoT-VLA) are ordinary architectural references, not load-bearing uniqueness theorems that force the result. No fitted-input-as-prediction, no ansatz smuggled via self-citation, and no renaming of a known identity. The derivation chain is therefore self-contained against external benchmarks; score 0 is the correct outcome.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on standard VLA training assumptions plus a handful of design choices (latent horizon, gate form, OAT as the latent source) that are either taken from prior work or selected by ablation. No new physical entities are postulated; the free parameters are ordinary ML hyperparameters whose values are reported.

free parameters (5)
  • latent-action horizon Hl = 20 (LIBERO) / 10 (LIBERO-Plus)
    Chosen by ablation (Tab. 3); set to 20 on LIBERO and 10 on LIBERO-Plus; directly controls the temporal scope of the claimed guidance.
  • latent alignment weight λalign = 0.1
    Balances Lalign against the flow-matching action loss; fixed at 0.1 without further sensitivity study.
  • number of latent query tokens Nq and latent dim Dz = Nq=8, Dz=4
    Architectural free choices that determine capacity of the latent interface.
  • latent-action conditioning dropout = 0.1
    Regularization rate applied to the conditioning path; set to 0.1.
  • peak learning rate and schedule = 1.25e-5 (sim) / 5e-5 (stacking)
    1.25e-5 cosine with 10k warmup; task-specific override for real-robot stacking (5e-5).
assumptions (4)
  • domain assumption A frozen ordered action tokenizer (OAT) produces raw latents that are a useful coarse-to-fine encoding of future continuous action segments.
    Invoked in §3.2 Eq. 3 as the sole source of latent-action supervision and training-time conditioning; taken from the cited OAT paper without re-derivation.
  • domain assumption Flow-matching action experts of the π0.5 family remain the appropriate continuous generators once latent conditioning is added.
    The entire expert stack and L_act loss are inherited from π0.5 (§3.4, App. A).
  • ad hoc to paper Cross-attention residual injection modulated by a channel-wise sigmoid gate is a sufficient mechanism for adaptive conditioning strength.
    Defined in §3.3 Eqs. 8–12; alternative fusion methods are not exhaustively compared beyond the ‘w/o gate’ ablation.
  • domain assumption LIBERO and LIBERO-Plus success rates under the official protocols are valid proxies for generalist manipulation competence.
    Standard community assumption used for all reported claims in §4.
invented entities (2)
  • Context-Gated Action Conditioning module (context gate + latent-action cross-attention residual)
    purpose: Adaptively scale the residual contribution of predicted latent actions into each expert layer.
    Newly introduced architectural block (§3.3, Fig. 1c); independent evidence is only the ablation and gate visualizations inside this paper.
  • VLM-native latent-action interface (learnable query tokens + OAT-aligned prediction head)
    purpose: Let the VLM itself emit action-structured conditioning without a separate action reasoner.
    Defined in §3.2; purpose is architectural convenience and claimed performance; no external falsifiable prediction beyond the reported success rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAC-VLA: Context-Gated Action Conditioning for Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/674YNXJI

@misc{pith2026260704816,
  author       = {Pith},
  title        = {Pith review of: CAC-VLA: Context-Gated Action Conditioning for Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/674YNXJI}},
  note         = {Machine review of arXiv:2607.04816}
}
read the original abstract

Vision-Language-Action (VLA) models have become a promising paradigm for generalist robot manipulation, where visual-language representations are used to condition continuous action generation. However, these representations are not explicitly optimized for action conditioning, leaving the action expert to bridge the gap between multimodal understanding and precise motor control. Recent action-reasoning methods introduce additional modules to generate explicit action plans or action-space reasoning signals, demonstrating the benefit of action-level guidance but often requiring separate action-generation frameworks. We propose CAC-VLA, a Context-Gated Action Conditioning framework that learns a lightweight latent-action interface directly within the VLM. Instead of generating executable trajectories, CAC-VLA trains the VLM to predict coarse-to-fine latent actions, which are structured representations encoded from future action segments, and adaptively leverages them to condition the action expert via a context gate. This enables VLM-native action conditioning while calibrating the influence of latent-action guidance on expert action generation. Experiments on LIBERO and LIBERO-Plus demonstrate the effectiveness of CAC-VLA, achieving 98.3% average success rate on LIBERO and 89.5% LIBERO-Plus, suggesting that context-gated latent-action conditioning is an effective interface for continuous expert control.

Figures

Figures reproduced from arXiv: 2607.04816 by the authors.

Figure 1
Figure 1. Overview of CAC-VLA, the expert-aware attention mask, and the context-gate module. resentations, the policy first predicts latent actions from the multimodal context and then conditions the action expert on them: zt = fθ (ot, l), at:t+He−1 = πθ (ot, l, zt). (2) Here, fθ denotes the VLM-side latent-action predictor, and zt provides action-structured condition￾ing for continuous expert control. 3.2 VLM-native Latent A… view at source ↗
Figure 2
Figure 2. Quantitative real-world evaluation on pick-and-place and block-stacking tasks. We report [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Visualization of raw context-gate values across flow steps. We report the average gate [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Effective latent-conditioning strength on LIBERO-Long and LIBERO-Goal. We report [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Phase-level visualization of effective latent-conditioning strength for a representative [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Representative successful and failed real-world executions of [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 29 linked inside Pith

  1. [1]

    Beyer, A

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Al- abdulmohsin, M. Tschannen, E. Bugliarello, T. Unterthiner, D. Keysers, S. Koppula, F. Liu, A. Grycner, A. Gritsenko, N. Houlsby, M. Kumar, K. Rong, J. Eisenschlos, R. Kabra, M. Bauer, M. Boˇsnjak, X. Chen, M. Minderer, P. V oigtlaender, I. Bica, I. Balazevic, J. Puigcer...

  2. [2]

    C. Ni, C. Chen, X. Wang, Z. Zhu, W. Zheng, B. Wang, T. Chen, G. Zhao, H. Li, Z. Dong, Q. Zhang, Y . Ye, Y . Wang, G. Huang, and W. Mei. Swiftvla: Unlocking spatiotemporal dy- namics for lightweight vla models at minimal overhead, 2025. URLhttps://arxiv. org/abs/2512.00903

  3. [3]

    Zhang, H

    W. Zhang, H. Liu, Z. Qi, Y . Wang, X. Yu, J. Zhang, R. Dong, J. He, F. Lu, H. Wang, Z. Zhang, L. Yi, W. Zeng, and X. Jin. Dreamvla: A vision-language-action model dreamed with compre- hensive world knowledge, 2025. URLhttps://arxiv.org/abs/2507.04447

  4. [4]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, M. Y . Galliker, D. Ghosh, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. V...

  5. [5]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Haus- man, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.π 0: A vision-language-action flow model for general robot control, 2026. URLhttps://arxiv. o...

  6. [6]

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

  7. [7]

    Marafioti, O

    A. Marafioti, O. Zohar, M. Farr ´e, M. Noyan, E. Bakouch, P. Cuenca, C. Zakka, L. B. Allal, A. Lozhkov, N. Tazi, et al. Smolvlm: Redefining small and efficient multimodal models.arXiv preprint arXiv:2504.05299, 2025

  8. [8]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakr- ishnan, K. Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022

Show all 40 references
  1. [9]

    Q. Zhao, Y . Lu, M. J. Kim, Z. Fu, Z. Zhang, Y . Wu, Z. Li, Q. Ma, S. Han, C. Finn, A. Handa, M.-Y . Liu, D. Xiang, G. Wetzstein, and T.-Y . Lin. Cot-vla: Visual chain-of-thought reason- ing for vision-language-action models, 2025. URLhttps://arxiv.org/abs/2503. 22020

  2. [10]

    J. Cen, C. Yu, H. Yuan, Y . Jiang, S. Huang, J. Guo, X. Li, Y . Song, H. Luo, F. Wang, et al. Worldvla: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539, 2025

  3. [11]

    Zhong, Y

    L. Zhong, Y . Liu, Y . Wei, Z. Xiong, M. Yao, S. Liu, and G. Ren. Acot-vla: Action chain-of- thought for vision-language-action models.arXiv preprint arXiv:2601.11404, 2026

  4. [12]

    C. Liu, X. Han, J. Gao, Y . Zhao, H. Chen, and Y . Du. Oat: Ordered action tokenization, 2026. URLhttps://arxiv.org/abs/2602.04215

  5. [13]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. Libero: Benchmarking knowl- edge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023. 9

  6. [14]

    S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, et al. Libero-plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626, 2025

  7. [15]

    Driess, F

    D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y . Chebotar, P. Sermanet, D. Duckworth, S. Levine, V . Van- houcke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence. Palm-e: An e...

  8. [16]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. G. Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. L...

  9. [17]

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

  10. [18]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, C. Xu, J. Luo, T. Kreiman, Y . Tan, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine. Octo: An open-source generalist robot policy. InProceedings of Robotics: Science and S...

  11. [19]

    W. Yu, J. Peng, H. Yang, J. Zhang, Y . Duan, J. Ji, and Y . Zhang. Ldp: A local diffusion planner for efficient robot navigation and collision avoidance. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5466–5472. IEEE, 2024

  12. [20]

    Y . Duan, H. Li, Y . Wu, W. Yu, X. Zhang, Y . Shen, J. Ji, and Y . Zhang. Stdarm: Transferring visuomotor policies from static data training to dynamic robot manipulation.arXiv preprint arXiv:2504.18792, 2025

  13. [21]

    Y . Gao, Y . Shen, S. Zhang, W. Yu, Y . Duan, J. Wu, J. Deng, Y . Zhang, et al. Drift-based policy optimization: Native one-step policy learning for online robot control.arXiv preprint arXiv:2604.03540, 2026

  14. [22]

    S. Ye, J. Jang, B. Jeon, S. Joo, J. Yang, B. Peng, A. Mandlekar, R. Tan, Y .-W. Chao, B. Y . Lin, et al. Latent action pretraining from videos.arXiv preprint arXiv:2410.11758, 2024

  15. [23]

    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

  16. [24]

    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, 44(10-11):1684–1704, 2025

  17. [25]

    Q. Lv, W. Kong, H. Li, J. Zeng, Z. Qiu, D. Qu, H. Song, Q. Chen, X. Deng, and J. Pang. F1: A vision-language-action model bridging understanding and generation to actions.arXiv preprint arXiv:2509.06951, 2025. 10

  18. [26]

    Y . Liao, P. Zhou, S. Huang, D. Yang, S. Chen, Y . Jiang, Y . Hu, J. Cai, S. Liu, J. Luo, et al. Genie envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint arXiv:2508.05635, 2025

  19. [27]

    Zheng, Y

    R. Zheng, Y . Liang, S. Huang, J. Gao, H. Daum ´e III, A. Kolobov, F. Huang, and J. Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies.arXiv preprint arXiv:2412.10345, 2024

  20. [28]

    D. Qu, H. Song, Q. Chen, Y . Yao, X. Ye, Y . Ding, Z. Wang, J. Gu, B. Zhao, D. Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model.arXiv preprint arXiv:2501.15830, 2025

  21. [29]

    Huang, Y .-H

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

  22. [30]

    Pertsch, K

    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

  23. [31]

    Y . Yang, Z. Duan, T. Xie, F. Cao, P. Shen, P. Song, P. Jin, G. Sun, S. Xu, Y . You, et al. Fpc-vla: A vision-language-action framework with a supervisor for failure prediction and correction. arXiv preprint arXiv:2509.04018, 2025

  24. [32]

    Shukor, D

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

  25. [33]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

  26. [34]

    Q. Bu, J. Cai, L. Chen, X. Cui, Y . Ding, S. Feng, S. Gao, X. He, X. Hu, X. Huang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025

  27. [35]

    Liang, Y

    Z. Liang, Y . Li, T. Yang, C. Wu, S. Mao, L. Pei, X. Yang, J. Pang, Y . Mu, and P. Luo. Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies. arXiv preprint arXiv:2508.20072, 2025

  28. [36]

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, and G. Huang. Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic ma- nipulation.arXiv preprint arXiv:2508.19236, 2025

  29. [37]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

  30. [38]

    Y . Wang, P. Ding, L. Li, C. Cui, Z. Ge, X. Tong, W. Song, H. Zhao, W. Zhao, P. Hou, et al. Vla- adapter: An effective paradigm for tiny-scale vision-language-action model. InProceedings of the AAAI conference on artificial intelligence, volume 40, pages 18638–18646, 2026

  31. [39]

    C.-Y . Hung, Q. Sun, P. Hong, A. Zadeh, C. Li, U. Tan, N. Majumder, S. Poria, et al. Nora: A small open-sourced generalist vision language action model for embodied tasks.arXiv preprint arXiv:2504.19854, 2025

  32. [40]

    S. Tan, K. Dou, Y . Zhao, and P. Kr ¨ahenb¨uhl. Interactive post-training for vision-language- action models.arXiv preprint arXiv:2505.17016, 2025. 11 A Additional Training Details This appendix provides additional implementation and training details forCAC-VLA. Unless other- ...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.