Pith. sign in

REVIEW 3 major objections 7 minor 48 references

A world-action model that predicts both crisp video and semantic object-state features more than doubles real-world robot success when lighting, background, and object appearance shift.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 15:44 UTC pith:SLZ2NTRG

load-bearing objection A well-executed architecture paper with real empirical gains; the robustness story is plausible, but the semantic-stability mechanism isn't directly validated under the real-world shifts that matter most. the 3 major comments →

arxiv 2607.28993 v1 pith:SLZ2NTRG submitted 2026-07-31 cs.RO cs.CV

ST-WAM: Semantic-Temporal World Action Model for Robust Manipulation under Visual Distribution Shifts

Mingxin Wang , Bin Hu , Bin Qian , Kaitao Jiang , Haoning Wu , Feng Yan , Bowen Jing , Ruiyang Hao
show 7 more authors
Enyi Wang Kangning Niu Yandan Yang Mu Xu Yan Wang Houde Liu Tianlun Li
This is my paper
classification cs.RO cs.CV
keywords world action modelvisual distribution shiftDINOv3semantic future predictionhistory retrievalrobust manipulationflow matchingLIBERO-Plus
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.

This paper argues that video-generative world-action models fail under visual distribution shifts because their pixel-level future predictions entangle task-relevant changes with irrelevant scene details, causing what they term Training-Distribution Hallucination. The proposed remedy, ST-WAM, keeps fine-grained VAE dynamics but adds DINOv3 semantic features as a shared representation for both future prediction and history retrieval. The central claim is that this semantic-temporal complement improves zero-shot LIBERO-Plus success by 21.3 percentage points and raises real-world shift robustness from 25.8% to 61.5%, all without embodied pretraining or test-time future generation. If correct, this points to semantic future supervision as a broadly useful complement to pixel-generative objectives.

Core claim

The paper's central claim is that pixel-generative future supervision alone is insufficient for robust manipulation under visual distribution shifts because it allows the model to hallucinate training-domain content when given visually shifted observations. Instead, the authors show that a frozen self-supervised encoder, DINOv3, yields feature representations that remain stable under visual shifts while preserving task-state discriminability, and they build ST-WAM to exploit this: Dual-Space Future Experts jointly predict future VAE latents and DINO features, while Current-Anchored Intent Retrieval pulls task-relevant evidence from recent DINO history. The paper's empirical result is a subst

What carries the argument

The central mechanism is the use of DINOv3 features as a shared semantic representation across time. DSFE (Dual-Space Future Experts) runs three DiT branches—visual future, semantic future, and action—as a Mixture-of-Transformers, with an asymmetric attention mask that lets the two future streams refine each other while preventing future-target leakage into action generation. CAIR (Current-Anchored Intent Retrieval) uses Qwen3-VL to form a current visual-language anchor, then retrieves evidence from a short DINO history via cross-attention, producing label-free intent tokens that condition only the action branch. Together they supply prospective semantic state transitions and retrospective t

Load-bearing premise

The central assumption is that DINOv3 features, frozen from internet-pretrained weights, remain stable under exactly the visual shifts encountered at deployment while still distinguishing the task-relevant state; this stability is demonstrated only in controlled simulation triplets, not in the real-world shift conditions.

What would settle it

A direct test would be to evaluate a LIBERO-trained ST-WAM on real-world episodes where the background, lighting, and object colors are all changed simultaneously beyond the paper's current shift set, and measure whether the DINO features' same-state cosine similarity stays above the Wan-VAE baseline; a large drop would indicate the semantic anchor also fails under deeper distributional changes.

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

If this is right

  • If DINOv3 features stay stable under real-world shifts, the same dual-space recipe could be applied to other policy architectures without changing the base model's pretraining.
  • Because ST-WAM needs no embodied pretraining or task-specific annotations, it lowers the barrier for robust policy learning in new manipulation environments.
  • The 1.24x inference overhead over Fast-WAM is small enough that robustness gains come at acceptable latency for real-time control.
  • The structured attention mask and flow-matching objective decouple action from future streams, so the future branches can be dropped at inference while preserving the robustness benefit.
  • The consistent gains across all seven LIBERO-Plus perturbation categories, especially large gains under camera and sensor-noise shifts, suggest the approach addresses a general representation problem rather than a single failure mode.

Where Pith is reading between the lines

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

  • The paper's diagnosis of Training-Distribution Hallucination implies that any pixel-generative future objective, not just VAE-based ones, may suffer from the same entangling; the proposed semantic supervision could be a generally applicable corrective.
  • The success of DINO history retrieval suggests that long-horizon tasks with visual aliasing may particularly benefit from current-anchored memory, a direction the paper only begins to explore.
  • One could test whether the same robustness transfers to novel embodiments or physics changes by replacing DINOv3 with an alternative self-supervised encoder, or by probing how much of the gain comes from feature stability versus task-state discriminability.
  • The real-world compound shift result (48.0% vs 15.3%) hints that the method is not just fixing a single visual attribute but is addressing compositional robustness; checking how far this extends under unseen combinations would be a natural next experiment.

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 / 7 minor

Summary. The paper proposes ST-WAM, a World Action Model that augments VAE-latent future prediction with DINOv3 semantic future supervision (DSFE) and current-anchored retrieval from recent DINO history (CAIR). The authors identify a failure mode they call Training-Distribution Hallucination in video-generative WAMs under visual shifts, support it with a manual audit and a frame-triplet diagnosis comparing DINOv3 and Wan-VAE representations, and then evaluate ST-WAM on LIBERO, LIBERO-Plus, RoboTwin 2.0, and five real-world tasks. The headline results are 98.7% on LIBERO, 72.8% on LIBERO-Plus (21.3 percentage points above Fast-WAM), and 61.5% real-world success under visual shifts versus 25.8% for Fast-WAM. Ablations on LIBERO-Plus and in the real world indicate that both the semantic future objective and current-anchored DINO history retrieval contribute to the gains.

Significance. If the results hold, ST-WAM is a practically meaningful step toward robust action-centric world models: it does not require embodied pretraining or test-time video generation, and it outperforms several embodied-pretrained baselines under zero-shot visual shifts. The paper's strengths are the breadth of the evaluation (two simulation benchmarks plus a five-task real-world study), the component ablations that isolate the semantic future expert and the CAIR mechanism, and the careful comparison to a closely matched Fast-WAM baseline. The design rationale is plausible, and the ablation evidence supports the claim that the two proposed components are complementary. The main weaknesses are that the DINOv3 stability diagnosis is simulation-only, the real-world results lack statistical reporting, and some protocol details are underspecified.

major comments (3)
  1. [§1, Fig. 1(b), Table 4] The frame-triplet diagnosis that motivates the choice of DINOv3 is conducted only on LIBERO/LIBERO-Plus simulation. The real-world evaluation in Table 4 includes object-appearance and compound shifts, but no real-world DINOv3 stability or state-discriminability measurement is reported. Since the CAIR/DSFE mechanism is explicitly grounded in DINOv3's stability under visual shifts, the causal explanation for the real-world gains is not directly supported in the deployment regime. Please add a real-world triplet or per-shift feature-stability analysis (e.g., same-state DINO cosine similarity under the shifts used in Table 4), or alternatively soften the mechanism claim and present the real-world gains as empirical rather than mechanism-confirming.
  2. [Table 4] Real-world results are reported as aggregate point estimates from 30 trials per task and condition, with no confidence intervals, per-task breakdowns, or significance tests. The phrase 'predefined object initializations' leaves it unclear whether the 30 trials are deterministic repetitions or independent random draws. Since the central headline is the 25.8%→61.5% improvement under visual shifts, the paper should report task-wise results and confidence intervals and clarify the trial protocol. This is needed to establish that the gap is not partly an artifact of a particular fixed initialization set.
  3. [§1] The Training-Distribution Hallucination audit is central to the paper's motivation, but the protocol is underspecified. The claim that 70.6% of 180 predicted futures 'distinctly exhibit' the phenomenon relies on manual audit, yet the paper does not state the annotation criteria, whether the audit was performed on reconstructed videos or latent visualizations, the number of annotators, or inter-annotator agreement. Please provide this information, or supplement the manual audit with a quantitative measure (e.g., distance of the predicted future to training-domain content). Without this, the prevalence figure is difficult to assess.
minor comments (7)
  1. [Abstract / Methodology] The term 'pixel-generative future supervision' is used loosely; the future supervision is in VAE latent space, not raw pixels. Consider consistent phrasing such as 'video-generative' or 'latent-video' throughout.
  2. [Eq. (10)] The timestep-dependent weight w(τ_r) is not defined. Clarify whether it is the flow-matching weighting, a constant, or part of the shifted schedule; otherwise the objective is ambiguous.
  3. [Implementation Details] The 'shifted flow-matching schedule with shift 5.0' is not explained. Specify the timestep distribution (e.g., logit-normal with shift) and how it interacts with the sampled τ_r in Eq. (8).
  4. [Table 5] For the 'Semantic Expert w/o Future Obj.' variant, the parameter-matched claim is unclear: is the semantic prediction head retained but untrained, or removed? If it is removed, the comparison conflates architecture and loss. Please clarify, even though the large gap to the full model supports the qualitative conclusion.
  5. [Real-World Evaluation] Please state how many random seeds or independent initial conditions are used for each of the 30 trials per task and condition. If the 'predefined object initializations' are fixed across methods, describe how they were generated.
  6. [Methodology, §'Dual-Space Future Experts'] The paper says DINOv3 features are 'dense, frame-wise semantic features' but does not specify whether the ViT-S/16 output uses CLS tokens, patch tokens, or both, and which layer is taken. This detail is needed for reproducibility.
  7. [References] CF-VLA appears in the reference list and is discussed in related work, but the in-text citation is missing; please ensure all references are cited in the text.

Circularity Check

0 steps flagged

No significant circularity: the central claims are empirical benchmark results, and the DINOv3-based representation is an external pretrained encoder used as a measured input, not a fitted output.

full rationale

The paper's load-bearing claims are empirical benchmark numbers (LIBERO, LIBERO-Plus, RoboTwin 2.0, real-world trials) obtained by training ST-WAM and comparing with baselines. No equation or construction in the paper defines a parameter in terms of the quantity it is later claimed to predict. The DINOv3 stability diagnosis in Section 1 / Fig. 1(b) is a motivational measurement (cosine similarity on 290 frame triplets), not a target used to fit the model; DINOv3 is a frozen, externally pretrained encoder, and the model is trained end-to-end with a flow-matching objective over actions, VAE latents, and DINO features. The improvements on LIBERO-Plus and in the real world are supported by ablations (removing the semantic future expert or CAIR reduces shifted-condition performance), which are empirical comparisons rather than definitional equivalences. The only self-citations (e.g., CF-VLA, RoboTron-Mani by overlapping authors) appear in related-work context and do not carry the argument. The skeptic concern that DINOv3 stability was not validated on real-world shifts is a correctness/evidence limitation, not circularity: even if the mechanism were unsupported, the reported success rates would stand as point estimates. Therefore no circular step meets the bar of exhibiting a specific reduction from the paper's equations or self-citation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The ledger is dominated by standard supervised-training hyperparameters and the domain assumption that DINOv3 features are shift-stable and state-discriminative. The paper introduces one named phenomenon and one internal latent representation; neither is a physical entity with an external falsifiable handle. The empirical claims are benchmarked externally, so the circularity burden is low.

free parameters (5)
  • Loss weights (λv, λs, λa) = 1.0, 0.02, 1.0
    Chosen by hand; the small λs presumably balances DINOv3 feature scale against VAE and action losses. No sensitivity analysis is reported, so it is unknown how much of the LIBERO-Plus gain depends on these values.
  • Flow-matching timestep shift = 5.0
    A training hyperparameter that changes the noise schedule; reported without a selection procedure or sensitivity study.
  • CAIR query count N_I = 8
    Number of learnable intent queries; an architectural choice not ablated.
  • History length M = 4
    Number of preceding observations used for retrieval; fixed without reported ablation.
  • Retrieval depth L = 2
    Number of CAIR cross-attention blocks; chosen by hand, not ablated.
axioms (4)
  • domain assumption DINOv3 ViT-S/16 features are stable under visual distribution shifts and preserve task-state discriminability
    Central premise of the method. Supported only by the 290-triplet diagnosis in §1/Fig. 1(b); not shown for real-world compound shifts or beyond LIBERO-style perturbations.
  • domain assumption Cosine similarity in DINO feature space is a valid measure of task-state equivalence
    The frame-triplet diagnosis uses cosine similarity as a proxy for stability and discriminability; this metric choice is not itself validated against task success.
  • standard math Flow matching with linear interpolation and independent Gaussian noises per branch is a valid training objective for coupled future and action branches
    Standard flow-matching formulation used in prior WAMs; the paper extends it to three branches with synchronized future timesteps. This is a reasonable background assumption, not a new derivation.
  • domain assumption Pretrained Wan2.2 Video DiT and Wan2.2 VAE priors transfer to the modified three-branch MoT architecture
    The method initializes visual, semantic, and action experts from Wan2.2 weights; the paper assumes these priors survive the architectural surgery and still provide useful dynamics.
invented entities (2)
  • Training-Distribution Hallucination (named failure mode) independent evidence
    purpose: Describes the tendency of video-generative WAM futures to drift to training-domain content under visual shifts; motivates the semantic representation design.
    Supported by a manual audit (70.6% of 180 predictions) and by reported success-rate drops. The audit is subjective and not externally benchmarked, so the evidence is partial.
  • Short-horizon intent tokens I_t (CAIR output) no independent evidence
    purpose: Latent, label-free summary of recent task progress used to condition the action expert.
    Internal learned variables with no falsifiable handle outside the model; the paper itself calls them latent and label-free. They are ablated indirectly through CAIR ablations, but are not independently observable.

pith-pipeline@v1.3.0-daily-deepseek · 13122 in / 15576 out tokens · 160703 ms · 2026-08-03T15:44:41.719238+00:00 · methodology

0 comments
read the original abstract

World Action Models (WAMs) have emerged as a promising paradigm by jointly modeling robot actions and future visual dynamics. However, their reliance on pixel-generative future supervision can entangle action-relevant state transitions with task-irrelevant visual content, limiting robustness under visual distribution shifts. We identify Training-Distribution Hallucination, a recurring phenomenon in which futures conditioned on visually shifted observations hallucinate training-domain content rather than remain faithful to the current scene. A controlled frame-triplet diagnosis further shows that DINOv3 features remain more stable across visual shifts while better preserving task-state distinctions than Wan-VAE latents. Rather than correcting the predicted futures, we propose Semantic-Temporal WAM (ST-WAM) to improve action robustness by using DINOv3 as a shared semantic representation for future prediction and history retrieval while retaining fine-grained VAE dynamics. Its Dual-Space Future Experts (DSFE) jointly predict future VAE latents and DINO features, while Current-Anchored Intent Retrieval (CAIR) retrieves task-relevant evidence from recent DINO history under the current visual-language context. ST-WAM is trained end-to-end without additional embodied pretraining or task-specific annotations, and requires no explicit future generation at inference. It achieves 98.7% on LIBERO and 92.8% on RoboTwin 2.0; more importantly, compared with Fast-WAM, it improves zero-shot LIBERO-Plus performance by 21.3 percentage points and more than doubles real-world success under visual shifts from 25.8% to 61.5%. These results demonstrate that semantic-temporal modeling effectively complements pixel-generative dynamics for robust manipulation.

Figures

Figures reproduced from arXiv: 2607.28993 by Bin Hu, Bin Qian, Bowen Jing, Enyi Wang, Feng Yan, Haoning Wu, Houde Liu, Kaitao Jiang, Kangning Niu, Mingxin Wang, Mu Xu, Ruiyang Hao, Tianlun Li, Yandan Yang, Yan Wang.

Figure 1
Figure 1. Figure 1: (a) Three representative Training-Distribution [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ST-WAM. (a): Dual-Space Future Experts (DSFE) jointly model future dynamics in the VAE visual [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Structured cross-branch attention masks during [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world evaluation of ST-WAM on five tasks [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention heatmaps on two representative tasks. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 34 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2602.15922 , year=

    World action models are zero-shot policies , author=. arXiv preprint arXiv:2602.15922 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2601.21998 , year=

    Causal World Modeling for Robot Control , author=. arXiv preprint arXiv:2601.21998 , year=

    Pith/arXiv arXiv

  3. [3]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Reconvla: Reconstructive vision-language-action model as effective robot perceiver , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [4]

    arXiv preprint arXiv:2606.15768 , year=

    Lawam: Latent world action models for efficient dynamics-aware robot policies , author=. arXiv preprint arXiv:2606.15768 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2603.22078 , year=

    Do world action models generalize better than vlas? a robustness study , author=. arXiv preprint arXiv:2603.22078 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2606.13515 , year=

    Maskwam: Unifying mask prompting and prediction for world-action models , author=. arXiv preprint arXiv:2606.13515 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2602.12215 , year=

    Lda-1b: Scaling latent dynamics action model via universal embodied data ingestion , author=. arXiv preprint arXiv:2602.12215 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2606.03188 , year=

    GeoSem-WAM: Geometry-and Semantic-Aware World Action Models , author=. arXiv preprint arXiv:2606.03188 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2605.14712 , year=

    IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation , author=. arXiv preprint arXiv:2605.14712 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2606.21088 , year=

    MV-WAM: Manifold-Aware World Action Model with Value Augmentation , author=. arXiv preprint arXiv:2606.21088 , year=

    Pith/arXiv arXiv

  11. [11]

    5: Unified mobility-and-manipulation world action model , author=

    Abot-m0. 5: Unified mobility-and-manipulation world action model , author=. arXiv preprint arXiv:2607.00678 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2603.16666 , year=

    Fast-wam: Do world action models need test-time future imagination? , author=. arXiv preprint arXiv:2603.16666 , year=

    Pith/arXiv arXiv

  13. [13]

    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 , journal =

  14. [14]

    arXiv preprint arXiv:2504.16054 , year =

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2604.15483 , year =

    Pith/arXiv arXiv

  16. [16]

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

    Motus: A unified latent action world model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [17]

    arXiv preprint arXiv:2603.17240 , year=

    GigaWorld-Policy: An Efficient Action-Centered World--Action Model , author=. arXiv preprint arXiv:2603.17240 , year=

    arXiv

  18. [18]

    arXiv preprint arXiv:2601.16163 , year=

    Cosmos policy: Fine-tuning video models for visuomotor control and planning , author=. arXiv preprint arXiv:2601.16163 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2406.09246 , year=

    Openvla: An open-source vision-language-action model , author=. arXiv preprint arXiv:2406.09246 , year=

    Pith/arXiv arXiv

  20. [20]

    2026 , note =

    Jonas Pai and Liam Achenbach and Victoriano Montesinos and Benedek Forrai and Oier Mees and Elvis Nava , booktitle =. 2026 , note =. 2512.15692 , url =

    Pith/arXiv arXiv 2026

  21. [21]

    arXiv preprint arXiv:2603.10448 , year=

    Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control , author=. arXiv preprint arXiv:2603.10448 , year=

    arXiv

  22. [22]

    arXiv preprint arXiv:2604.11135 , year=

    Aim: Intent-aware unified world action modeling with spatial value maps , author=. arXiv preprint arXiv:2604.11135 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2604.19683 , year=

    Mask World Model: Predicting What Matters for Robust Robot Policy Learning , author=. arXiv preprint arXiv:2604.19683 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2602.22010 , year=

    World guidance: World modeling in condition space for action generation , author=. arXiv preprint arXiv:2602.22010 , year=

    arXiv

  25. [25]

    7: A latent world-action model from egocentric videos , author=

    Being-h0. 7: A latent world-action model from egocentric videos , author=. arXiv preprint arXiv:2605.00078 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2508.10104 , year=

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

    Pith/arXiv arXiv

  27. [27]

    Advances in Neural Information Processing Systems , volume=

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

  28. [28]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2510.13626 , year=

    Libero-plus: In-depth robustness analysis of vision-language-action models , author=. arXiv preprint arXiv:2510.13626 , year=

    Pith/arXiv arXiv

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge , author=. Advances in Neural Information Processing Systems , volume=

  31. [31]

    arXiv preprint arXiv:2602.10098 , year=

    Vla-jepa: Enhancing vision-language-action model with latent world model , author=. arXiv preprint arXiv:2602.10098 , year=

    arXiv

  32. [32]

    arXiv preprint arXiv:2604.16391 , year=

    Disentangled robot learning via separate forward and inverse dynamics pretraining , author=. arXiv preprint arXiv:2604.16391 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2411.04996 , year=

    Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models , author=. arXiv preprint arXiv:2411.04996 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2506.18088 , year=

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation , author=. arXiv preprint arXiv:2506.18088 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  36. [36]

    2506.21539 , archivePrefix =

    Jun Cen and Chaohui Yu and Hangjie Yuan and Yuming Jiang and Siteng Huang and Jiayan Guo and Xin Li and Yibing Song and Hao Luo and Fan Wang and Deli Zhao and Hao Chen , year =. 2506.21539 , archivePrefix =

    Pith/arXiv arXiv

  37. [37]

    2504.19854 , archivePrefix =

    Chia-Yu Hung and Qi Sun and Pengfei Hong and Amir Zadeh and Chuan Li and U-Xuan Tan and Navonil Majumder and Soujanya Poria , year =. 2504.19854 , archivePrefix =

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2505.06111 , year=

    Univla: Learning to act anywhere with task-centric latent actions , author=. arXiv preprint arXiv:2505.06111 , year=

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2501.09747 , year=

    Fast: Efficient action tokenization for vision-language-action models , author=. arXiv preprint arXiv:2501.09747 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2505.17016 , year=

    Interactive post-training for vision-language-action models , author=. arXiv preprint arXiv:2505.17016 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2502.19645 , year=

    Fine-tuning vision-language-action models: Optimizing speed and success , author=. arXiv preprint arXiv:2502.19645 , year=

    Pith/arXiv arXiv

  42. [42]

    arXiv preprint arXiv:2510.10274 , year=

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model , author=. arXiv preprint arXiv:2510.10274 , year=

    Pith/arXiv arXiv

  43. [43]

    The eleventh international conference on learning representations , year=

    Flow matching for generative modeling , author=. The eleventh international conference on learning representations , year=

  44. [44]

    Robot World Models , year=

    EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data , author=. Robot World Models , year=

  45. [45]

    arXiv preprint arXiv:2607.05468 , year=

    Learning 4D Geometric Priors for Inference-Efficient World Action Models , author=. arXiv preprint arXiv:2607.05468 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2605.06481 , year=

    Oa-wam: Object-addressable world action model for robust robot manipulation , author=. arXiv preprint arXiv:2605.06481 , year=

    Pith/arXiv arXiv

  47. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Yan, Feng and Liu, Fanfan and Huang, Yiyang and Guan, Zechao and Zheng, Liming and Zhong, Yufeng and Feng, Chengjian and Ma, Lin , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  48. [48]

    2026 , eprint=

    CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies , author=. 2026 , eprint=

    2026