Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior

T0 review · 3 major / 7 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Dense 3D scene flow, used as a latent prior inside a dual-expert VLA, grounds robot action prediction in geometry and lifts success rates on long-horizon and out-of-distribution manipulation.

desk verdict Solid dual-expert VLA that puts dense 3D scene flow inside a pretrained backbone via one-step partial denoising and gated attention; tables and ablations back the gains under matched budgets. read the letter →

arxiv 2603.25399 v2 pith:KYL4HAUB submitted 2026-03-26 cs.CV cs.RO

Xinkai Wang , Chenyi Wang , Yifu Xu , Mingzhe Ye , Fucheng Zhang , Jialin Tian , Xinyu Zhan , Lifeng Zhu
show 2 more authors
Cewu Lu Lixin Yang
This is my paper · ORCID
classification cs.CVcs.RO
keywords vision-language-action3Dsceneflowlatentmotionpriormatchinggatedcross-attentionroboticmanipulationout-of-distributionrobustness
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

Current vision-language-action models map 2D semantic features straight to robot commands, so they must discover 3D contact physics only from action labels. That implicit route fails when viewpoints, layouts or object placements change. LaMP inserts an intermediate Motion Expert that predicts dense 3D scene flow (how every point on a grid will move in depth and image coordinates) and feeds one partially denoised hidden state of that flow into an Action Expert through a gated cross-attention layer. Because the gate starts near zero and opens only when the geometric cue helps, the pretrained language-vision backbone is not overwritten. The resulting policy records the highest average success rates on the standard LIBERO suites, on the harder LIBERO-Plus zero-shot perturbations, and on the SimplerEnv-WidowX transfer benchmark under identical data budgets, and it also improves real-robot pick-and-place, folding and multi-step serving tasks.

What carries the argument

Gated one-step motion guidance: the Motion Expert runs a single denoising step of a 10-step ODE (to τ = 0.1) on a 20×20×32 grid of (Δu, Δv, Δd) displacements; the resulting hidden state is fused into the last-layer vision-language features by a learnable scalar-gated cross-attention that begins at zero and rises only when beneficial.

What would settle it

Replace the one-step Motion-Expert features with pure Gaussian noise or with fully denoised flow and re-measure LIBERO-Long and LIBERO-Plus success; if the gated model no longer outperforms the no-motion ablation by the reported margins, the partial-denoising claim is false.

Watch

Extended reading notes

Core claim

Dense 3D scene flow functions as an effective latent motion prior for language-conditioned robot policies: a flow-matching Motion Expert that produces a single partially denoised 3D flow state, injected by gated cross-attention into a frozen vision-language backbone, yields higher success and greater out-of-distribution robustness than policies that must learn 3D dynamics only from action supervision.

Load-bearing premise

A single partial denoising step already produces hidden states that contain enough true future 3D geometry to condition the action policy; if those states are mostly noise or appearance artefacts, the claimed gains disappear.

Editorial extensions

If this is right

  • Explicit 3D geometric foresight can be added to existing VLAs without full multi-step world-model rollouts at inference time.
  • Camera-frame scene-flow priors transfer across embodiments and visual domains more readily than joint-space or pixel-space representations.
  • Long-horizon error accumulation shrinks when the policy is conditioned on predicted scene dynamics rather than on 2D semantics alone.
  • Gated injection of geometric features protects pretrained vision-language representations from collapse while still supplying contact-aware cues.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same one-step partial-denoising trick may generalise to other intermediate geometric tokens (contact maps, affordance fields) that are cheaper than full video prediction.
  • If TraceForge-style supervision can be replaced by self-supervised flow from unlabelled video, the motion prior could scale beyond curated robot datasets.
  • Adaptive spatial or temporal resolution of the flow grid would be a natural next control knob for tasks that need either finer contact or longer horizons.
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 / 7 minor

Summary. LaMP proposes a dual-expert Vision-Language-Action architecture that inserts dense 3D scene flow as an explicit latent motion prior between a pretrained VLM and an action policy. A flow-matching Motion Expert is trained on TraceForge-generated 3D scene flow; at policy training and inference only a single partial denoising step (τ=0.1 of a 10-step ODE) is run, and the resulting hidden states condition an Action Expert via a zero-initialized gated cross-attention layer. The paper reports state-of-the-art average success rates under matched training budgets on LIBERO (98.3%), LIBERO-Plus zero-shot OOD (79.3%, +9.7 pp over the strongest prior baseline), and SimplerEnv-WidowX (79.2%), with supporting ablations (No-Motion, 2D-Flow, Add, Concat+MLP) and real-robot multi-task experiments on a Flexiv Rizon 4.

Significance. If the empirical ranking holds under matched budgets, the work is a clear contribution to VLA design: it is among the first to integrate dense 3D scene flow as an internal latent prior inside a pretrained VLM-based policy rather than as a standalone world model, and the gated one-step conditioning is a practical way to inject geometric foresight without multi-step reconstruction latency. The consistent gains on long-horizon LIBERO-Long, camera/robot OOD axes of LIBERO-Plus, and Stack Block in SimplerEnv, together with the No-Motion and 2D-Flow ablations, make a credible case that depth-aware motion priors improve contact-aware control and cross-domain robustness. The two-stage training recipe, embodiment-agnostic (u,v,d) representation, and real-robot transfer with limited target data are useful engineering results for the community.

major comments (3)
  1. [Section III-D / Inference] Section III-D and the inference procedure make one-step partial denoising (τ=0.1 of N=10) load-bearing for both efficiency and the claim that hidden states already carry task-relevant 3D dynamics. The No-Motion ablation (Table I: 98.3%→92.4% LIBERO avg; Fig. 4: 79.2%→56.3% SimplerEnv) shows that some motion signal helps, but does not isolate whether one step is sufficient versus 0, 2, 5, or full reconstruction, nor whether the hidden states remain faithful to future geometry. A short sensitivity study over τ (or number of solver steps) on at least one benchmark is needed to ground this central design choice.
  2. [Table I, Table II] Tables I and II report point success rates and claim the highest averages under the same training budgets, yet no episode-level variance, confidence intervals, or statistical tests are provided. With 50 episodes per LIBERO task and large but not uniform margins (e.g., LIBERO-Goal 97.4% vs π0.5 98.0%; SimplerEnv Put Eggplant 95.8% vs FlowVLA 100%), ranking claims require uncertainty quantification. Please add error bars or bootstrap CIs and clarify that all baselines were trained or evaluated under identical data and compute budgets where re-runs were performed.
  3. [Section V / Fig. 7] Real-world evaluation (Section V) uses average task-progress scores with N=10 in-domain and N=5 OOD trials per condition, and compares against a re-implemented 3D FDP without an official codebase. Progress scores are reasonable for multi-stage tasks, but the small N and re-implementation make the OOD margin (62.5% vs 52.5% / 26.8%) harder to interpret. Please report per-stage success or binary success where possible, state the re-implementation fidelity for 3D FDP, and note limitations of sample size more explicitly.
minor comments (7)
  1. [Abstract / Fig. 1] Fig. 1 caption and abstract both claim superior real-world and simulation performance; ensure the abstract's 'highest reported average success rates under the same training budgets' is scoped only to the baselines and budgets actually compared in Tables I–II.
  2. [Section III / Appendix Algorithms 1–2] Notation: M_t is used for both the full motion tensor and the latent prior; z_m vs zm and m_0 vs M_0 appear inconsistently between Section III and the Appendix algorithms. Unify.
  3. [Section III-C / Limitations] Section III-C states K_h × K_w = 400 (20×20) while the Limitations section writes '20×20 grid' and Table III 'Grid resolution 20×20'; after 2×2 patchification tokens are 10×10—state the pre- and post-patch grid clearly once.
  4. [Section II] Related Work claims LaMP is the first to integrate dense 3D scene flow as a latent prior within a pretrained VLA; this is plausible given the cited standalone 3D flow pipelines, but a brief comparison table of intermediate representations (2D flow, traces, depth, latent actions, 3D flow) would help readers place the contribution.
  5. [Section III-A / Fig. 3] Fig. 3 is referenced as the overview but the manuscript text also refers to Fig. 1 for the same high-level idea; ensure figure numbering and callouts are consistent in the camera-ready version.
  6. [Section V-B / Tables I–II] Typos and wording: 'Unsenn Background' (Section V-B); 'w/omotion' spacing in tables; 'cam_ref' introduced without formal definition in the main text.
  7. [Appendix Table IV] Hyperparameters (Appendix Tables III–IV) list action chunk horizons H=9/29/15 across settings; a one-sentence justification in the main text for why SimplerEnv uses a much longer horizon would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical VLA method with external-benchmark evaluation and ablations; success rates are not forced by construction or self-citation.

full rationale

LaMP is a dual-expert architecture (Motion Expert via conditional flow matching on dense 3D scene flow + Action Expert via gated cross-attention) trained in two stages and evaluated on public external benchmarks (LIBERO, LIBERO-Plus zero-shot OOD, SimplerEnv-WidowX) plus real-robot trials under matched data budgets. The central performance claims (highest average success rates, 9.7-point OOD gain) are measured quantities, not algebraic consequences of the inputs. Ablations (w/o motion, 2D-flow prior, Add vs. Concat+MLP vs. gated fusion) produce large drops, confirming the reported gains are not definitional. TraceForge and starVLA are ordinary external tooling citations with no author overlap and no uniqueness theorem invoked to forbid alternatives. One-step partial denoising (τ=0.1) is an explicit design choice inspired by prior non-overlapping work (VPP, mimic-video), not a self-referential fit renamed as prediction. No self-definitional equations, fitted-input-as-prediction, load-bearing self-citation chains, or renamed known results appear in the derivation or evaluation chain.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The central performance claim rests on architectural choices (one-step τ=0.1, gated scalar, 20×20×32 flow grid), the quality of TraceForge pseudo-labels, and the assumption that last-layer VLM features plus partial motion states are jointly sufficient for action flow-matching. No free parameters are fitted to the final success-rate tables; the free parameters are design hyper-parameters chosen before evaluation.

free parameters (4)
  • partial-denoising fraction τ=0.1 (1 of N=10 steps)
    Hand-chosen operating point that trades completeness of motion reconstruction for latency; performance claims depend on this specific partial state being informative.
  • spatial grid K=20×20 and temporal horizon T=32
    Fixed resolution of the dense 3D flow representation; not derived from first principles.
  • gate scalar g initialized at 0
    Initialization and single shared scalar control how much motion signal enters the VLM features; chosen for training stability.
  • action chunk horizons H (9/29/15) and learning rates
    Benchmark-specific hyper-parameters listed in Table IV that affect absolute success rates.
assumptions (3)
  • domain assumption Conditional flow matching on TraceForge-generated 3D scene flow yields a useful latent motion prior for downstream action prediction.
    Invoked throughout Section III-B–D; the quality of the pseudo-labels is taken as given.
  • ad hoc to paper Last-layer VLM hidden states z_t plus one-step motion hidden states are jointly sufficient to condition an action flow-matching expert.
    Core factorization (Eq. 1) and gated fusion (Eq. 4); not independently verified outside the reported ablations.
  • domain assumption Camera-frame (Δu,Δv,Δd) trajectories are embodiment-agnostic enough for cross-robot transfer after light fine-tuning.
    Stated in Section III-C and used to justify real-world and SimplerEnv results.
invented entities (1)
  • LaMP dual-expert architecture (Motion Expert + gated Action Expert with one-step partial denoising)
    purpose: To inject dense 3D geometric foresight into a pretrained VLA without full multi-step reconstruction or representational collapse.
    The specific combination of one-step hidden-state conditioning and zero-init gated cross-attention is introduced by the paper; independent evidence is limited to the ablations and external benchmarks reported herein.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior." pith.science (2026). https://pith.science/paper/KYL4HAUB

@misc{pith2026260325399,
  author       = {Pith},
  title        = {Pith review of: LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYL4HAUB}},
  note         = {Machine review of arXiv:2603.25399}
}
read the original abstract

We introduce \textbf{LaMP}, a dual-expert Vision-Language-Action framework that embeds dense 3D scene flow as a latent motion prior for robotic manipulation.Existing VLA models regress actions directly from 2D semantic visual features, forcing them to learn complex 3D physical interactions implicitly.This implicit learning strategy degrades under unfamiliar spatial dynamics.LaMP addresses this limitation by aligning a flow-matching \emph{Motion Expert} with a policy-predicting \emph{Action Expert} through gated cross-attention.Specifically, the Motion Expert generates a one-step partially denoised 3D scene flow, and its hidden states condition the Action Expert without full multi-step reconstruction.We evaluate LaMP on the LIBERO, LIBERO-Plus, and SimplerEnv-WidowX simulation benchmarks as well as real-world experiments.LaMP consistently outperforms evaluated VLA baselines across LIBERO, LIBERO-Plus, and SimplerEnv-WidowX benchmarks, achieving the highest reported average success rates under the same training budgets. On LIBERO-Plus OOD perturbations, LaMP shows improved robustness with an average 9.7\% gain over the strongest prior baseline.Our project page is available at https://summerwxk.github.io/lamp-project-page/.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Track4Action: Distilling World-Centric 3D Tracker into Vision-Language-Action Policies

    cs.RO 2026-08 conditional novelty 6.0 of 10

    Track4Action distills a frozen 3D tracker's pooled feature over demonstration clips into track queries that condition a VLA action head, reporting gains on LIBERO, LIBERO-Plus, RoboTwin 2.0, and physical bimanual task...

  2. ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A geometry-consistent memory plus a self-supervised consistency-reinforcement stage improves video spatial reasoning by roughly 7–15 points on three benchmarks, with an advertised average gain of 12.6 points.

Reference graph

Works this paper leans on

43 extracted references · 15 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Thinkact: Vision-language-action reasoning via reinforced visual la- tent planning,

    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 la- tent planning,”arXiv preprint arXiv:2507.16815, 2025

    arXiv 2025

  2. [2]

    3d-vla: A 3d vision-language-action generative world model,

    H. Zhen, X. Qiu, P. Chen, J. Yang, X. Yan, Y . Du, Y . Hong, and C. Gan, “3d-vla: A 3d vision-language-action generative world model,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 61 229–61 245

    2024

  3. [3]

    Ddp-wm: Disentangled dynamics prediction for efficient world models,

    S. Yin, K. Yin, W. Chen, Y . Liu, G. Li, and L. Lin, “Ddp-wm: Disentangled dynamics prediction for efficient world models,”arXiv preprint arXiv:2602.01780, 2026

    arXiv 2026

  4. [4]

    Octo: An open-source generalist robot policy,

    D. Ghosh, H. R. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y . L. Tan, L. Y . Chen, Q. Vuong, T. Xiao, P. R. Sanketi, D. Sadigh, C. Finn, and S. Levine, “Octo: An open-source generalist robot policy,” inProceedings of Robotics: Science and Systems, 2024

    2024

  5. [5]

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

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kol- lar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “OpenVLA: An open-source vision-language-action model,” inProceedings of The 8th Conference on Robot Learning, 2025, pp. 2679–2713

    2025

  6. [6]

    GR00T N1: An open foundation model for generalist humanoid robots,

    J. Bjorck, F. Casta neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. LLontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y . L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y . Xie, Y . Xu, Z. Xu, S. Ye, Z. Yu, A. Zhang,...

    arXiv 2025

  7. [7]

    π {0.5}: a vision-language-action model with open-world generalization,

    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...

    2025

  8. [8]

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

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

    arXiv 2025

Show all 43 references
  1. [9]

    Eo-1: Interleaved vision-text-action pretrain- ing for general robot control,

    D. Qu, H. Song, Q. Chen, Z. Chen, X. Gao, X. Ye, Q. Lv, M. Shi, G. Ren, C. Ruanet al., “Eo-1: Interleaved vision-text-action pretrain- ing for general robot control,”arXiv preprint arXiv:2508.21112, 2025

    2025

  2. [10]

    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

    2025 arXiv

  3. [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. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. Shi, L. Smith, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky, “π0: A ...

    2024 arXiv

  4. [12]

    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, “Learning to act anywhere with task-centric latent actions,” inProceedings of Robotics: Science and Systems, 2025

    2025

  5. [13]

    Flowvla: Visual chain of thought-based motion reasoning for vision-language-action models,

    Z. Zhong, H. Yan, J. Li, X. Liu, X. Gong, T. Zhang, W. Song, J. Chen, X. Zheng, H. Wanget al., “Flowvla: Visual chain of thought-based motion reasoning for vision-language-action models,”arXiv preprint arXiv:2508.18269, 2025

    2025

  6. [14]

    LIBERO-Plus: In-depth robustness analysis of vision- language-action models,

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

    2025 arXiv

  7. [15]

    TraceVLA: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies,

    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

    2024 arXiv

  8. [16]

    F1: A vision-language-action model bridging understanding and generation to actions,

    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

    2025 arXiv

  9. [17]

    Flare: Robot learning with implicit world modeling,

    R. Zheng, J. Wang, S. Reed, J. Bjorck, Y . Fang, F. Hu, J. Jang, K. Kun- dalia, Z. Lin, L. Magne, A. Narayan, Y . L. Tan, G. Wang, Q. Wang, J. Xiang, Y . Xu, S. Ye, J. Kautz, F. Huang, Y . Zhu, and L. Fan, “Flare: Robot learning with implicit world modeling,” inProceedings of ...

    2025

  10. [18]

    3D diffusion policy: Generalizable visuomotor policy learning via simple 3D representations,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3D diffusion policy: Generalizable visuomotor policy learning via simple 3D representations,” inProceedings of Robotics: Science and Systems, 2024

    2024

  11. [19]

    RISE: 3D perception makes real-world robot imitation simple and effective,

    C. Wang, H. Fang, H.-S. Fang, and C. Lu, “RISE: 3D perception makes real-world robot imitation simple and effective,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 2870–2877

    2024

  12. [20]

    TraceGen: World modeling in 3d trace space enables learning from cross-embodiment videos,

    S. Lee, Y . Jung, I. Chun, Y .-C. Lee, Z. Cai, H. Huang, A. Talreja, T. D. Dao, Y . Liang, J.-B. Huanget al., “TraceGen: World modeling in 3d trace space enables learning from cross-embodiment videos,” arXiv preprint arXiv:2511.21690, 2025

    2025

  13. [21]

    3DFlowAction: Learning cross-embodiment manipulation from 3d flow world model,

    H. Zhi, P. Chen, S. Zhou, Y . Dong, Q. Wu, L. Han, and M. Tan, “3DFlowAction: Learning cross-embodiment manipulation from 3d flow world model,”arXiv preprint arXiv:2506.06199, 2025

    2025 arXiv

  14. [22]

    Video prediction policy: A generalist robot policy with predictive visual representations,

    Y . Hu, Y . Guo, P. Wang, X. Chen, Y .-J. Wang, J. Zhang, K. Sreenath, C. Lu, and J. Chen, “Video prediction policy: A generalist robot policy with predictive visual representations,” inInternational Conference on Machine Learning. PMLR, 2025, pp. 24 328–24 346

    2025

  15. [23]

    Mimic-Video: Video-action models for generalizable robot control beyond VLAs,

    J. Pai, L. Achenbach, V . Montesinos, B. Forrai, O. Mees, and E. Nava, “Mimic-Video: Video-action models for generalizable robot control beyond VLAs,”arXiv preprint arXiv:2512.15692, 2025

    2025 arXiv

  16. [24]

    LIBERO: Benchmarking knowledge transfer for lifelong robot learn- ing,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “LIBERO: Benchmarking knowledge transfer for lifelong robot learn- ing,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 776–44 791, 2023

    2023

  17. [25]

    Evaluating real-world robot manipulation policies in simulation,

    X. Li, K. Hsu, J. Gu, O. Mees, K. Pertsch, H. R. Walke, C. Fu, I. Lunawat, I. Sieh, S. Kirmaniet al., “Evaluating real-world robot manipulation policies in simulation,” inConference on Robot Learn- ing. PMLR, 2024, pp. 3705–3728

    2024

  18. [26]

    Learning universal policies via text-guided video generation,

    Y . Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schu- urmans, and P. Abbeel, “Learning universal policies via text-guided video generation,”Advances in neural information processing systems, vol. 36, pp. 9156–9172, 2023

    2023

  19. [27]

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

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

    2023 arXiv

  20. [28]

    Worldvla: Towards autoregressive action world model,

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

    2025 arXiv

  21. [29]

    Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge,

    W. Zhang, H. Liu, Z. Qi, Y . Wang, X. Yu, J. Zhang, R. Dong, J. He, H. Wang, Z. Zhanget al., “Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  22. [30]

    3d flow diffusion policy: Visuomotor policy learning via generating flow in 3d space,

    S. Noh, D. Nam, K. Kim, G. Lee, Y . Yu, R. Kang, and K. Lee, “3d flow diffusion policy: Visuomotor policy learning via generating flow in 3d space,”arXiv preprint arXiv:2509.18676, 2025

    2025

  23. [31]

    Robotic control via embodied chain-of-thought reasoning,

    M. Zawalski, W. Chen, K. Pertsch, O. Mees, C. Finn, and S. Levine, “Robotic control via embodied chain-of-thought reasoning,” inPro- ceedings of The 8th Conference on Robot Learning, 2024, pp. 3157– 3181

    2024

  24. [32]

    Qdepth- vla: quantized depth prediction as auxiliary supervision for vision- language-action models,

    Y . Li, Y . Chen, M. Zhou, H. Li, Z. Zhang, and D. Zhao, “Qdepth- vla: quantized depth prediction as auxiliary supervision for vision- language-action models,” inConference on Autonomous Agents and Multiagent Systems, 2026

    2026

  25. [33]

    MolmoAct: Action reasoning models that can reason in space,

    J. Lee, J. Duan, H. Fang, Y . Deng, S. Liu, B. Li, B. Fang, J. Zhang, Y . R. Wang, S. Leeet al., “MolmoAct: Action reasoning models that can reason in space,”arXiv preprint arXiv:2508.07917, 2025

    2025 arXiv

  26. [34]

    Villa-x: enhancing latent action modeling in vision-language-action models,

    X. Chen, H. Wei, P. Zhang, C. Zhang, K. Wang, Y . Guo, R. Yang, Y . Wang, X. Xiao, L. Zhaoet al., “Villa-x: enhancing latent action modeling in vision-language-action models,”arXiv preprint arXiv:2507.23682, 2025

    2025 arXiv

  27. [35]

    Flow as the cross-domain manipulation interface,

    M. Xu, Z. Xu, Y . Xu, C. Chi, G. Wetzstein, M. Veloso, and S. Song, “Flow as the cross-domain manipulation interface,” inProceedings of The 8th Conference on Robot Learning, 2024, pp. 2475–2499

    2024

  28. [36]

    Robotic vla benefits from joint learning with motion image diffusion,

    Y . Fang, K. Ranasinghe, L. Xue, H. Zhou, J. Tan, R. Xu, S. Hei- necke, C. Xiong, S. Savarese, D. Szafiret al., “Robotic vla benefits from joint learning with motion image diffusion,”arXiv preprint arXiv:2512.18007, 2025

    2025

  29. [37]

    Spatial traces: Enhancing vla models with spatial-temporal understanding,

    M. A. Patratskiy, A. K. Kovalev, and A. I. Panov, “Spatial traces: Enhancing vla models with spatial-temporal understanding,”Optical Memory and Neural Networks, vol. 34, no. Suppl 1, pp. S72–S82, 2025

    2025

  30. [38]

    Cogvideox: Text-to-video diffusion models with an expert transformer,

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Fenget al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  31. [39]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  32. [40]

    Bridgedata v2: A dataset for robot learning at scale,

    H. R. Walke, K. Black, T. Z. Zhao, Q. Vuong, C. Zheng, P. Hansen- Estruch, A. W. He, V . Myers, M. J. Kim, M. Duet al., “Bridgedata v2: A dataset for robot learning at scale,” inConference on Robot Learning, 2023, pp. 1723–1736

    2023

  33. [41]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Elliset al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

    2024 arXiv

  34. [42]

    Interndata-a1: Pioneering high- fidelity synthetic data for pre-training generalist policy,

    Y . Tian, Y . Yang, Y . Xie, Z. Cai, X. Shi, N. Gao, H. Liu, X. Jiang, Z. Qiu, F. Yuanet al., “Interndata-a1: Pioneering high- fidelity synthetic data for pre-training generalist policy,”arXiv preprint arXiv:2511.16651, 2025

    2025

  35. [43]

    Starvla: A lego-like codebase for vision- language-action model developing,

    starVLA Contributors, “Starvla: A lego-like codebase for vision- language-action model developing,” 1 2025. APPENDIX We define the notation shared across Algorithm 1 and Algorithm 2.o t is the visual observation at timestept.lis the language instruction.s t is the robot propri...

    2025

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.