Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A post-training semantic-foresight objective lifts out-of-distribution success of video-pretrained world-action models without giving up their generative pretraining.

desk verdict Useful plug-in for making VAE-space WAMs robust to visual shifts, with convincing aggregate gains, but the semantic-foresight mechanism is under-verified and missing error bars and code keep it from being an accept. read the letter →

arxiv 2608.05903 v2 pith:I4R2AXJJ submitted 2026-08-06 cs.CV cs.RO

Haodong Yan , Junfeng Li , Junjie He , Zhide Zhong , MingMing Yu , Wenxuan Song , Jiaguan Zhu , Yangyang Zheng
show 7 more authors
Yuqiao Du Jiadi You Yingjie Cai Xu Yan Guanyi Zhao Bingbing Liu Haoang Li
This is my paper · ORCID
classification cs.CVcs.RO
keywords world-actionmodelsvideogenerationsemanticforesightDINOv3CLSout-of-distributionrobustnessrepresentationalignmentrobotmanipulationquerytokens
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

The paper aims to remove a trade-off faced by world-action models: VAE-space models inherit strong video-generation pretraining but are fragile under visual shifts, while semantic-space models are robust but forgo that pretraining. Robust-WAM keeps the VAE-based generative path untouched and instead trains a handful of learnable query tokens in the action stream to output the semantic content of future frames, supervised by frozen DINOv3 CLS embeddings of ground-truth future frames. Across two out-of-distribution benchmarks and a real-robot lighting-shift setup, the method improves success rates of several WAM baselines without sacrificing in-distribution performance. If correct, this shows that appearance robustness and large-scale generative pretraining are not mutually exclusive for robot policies.

What carries the argument

The central object is the semantic foresight alignment objective: a cosine loss between the output hidden state of each learnable query token and the frozen DINOv3 CLS embedding of the corresponding future ground-truth frame. Query tokens live in the action stream alongside action tokens, and each takes the positional encoding of the matched action step through Eq. 4-5, so the model knows which future step the query describes. The frozen DINOv3 CLS target is chosen because it summarizes scene content such as objects and layout while being comparatively insensitive to texture and illumination. The alignment loss is added to the existing flow-matching video and action losses, leaving the VAE tokenizer, video denoiser, and video-action attention unchanged; at inference the teacher and linear alignment head are dropped and only the query tokens remain in the action sequence.

What would settle it

Train the same WAM with the query tokens present but with the alignment target replaced by random or constant vectors; if out-of-distribution success stays as high as with DINOv3 CLS targets, the semantic foresight objective is not the cause of the gains. Alternatively, probe the query-token outputs at inference under illumination shifts and check whether their nearest neighbours in DINOv3 CLS space track the ground-truth future scenes; if the queries' outputs correlate with training-set appearance rather than future semantics, the inference-time transfer claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the appearance bias of VAE-based world-action models can be corrected in the action representation instead of replacing the latent space. It prepends learnable query tokens to the action tokens of the action diffusion transformer, gives each query the positional encoding of the action step that leads to the future frame it describes, and aligns the query outputs with the frozen DINOv3 CLS embedding of that future frame using a cosine loss. At inference the teacher and alignment head are dropped, leaving only the queries in the action stream, which are supposed to supply appearance-invariant foresight through the existing attention path. The reported result is consistent OOD success-rate gains on FastWAM, GE-Act, and LingBot-VA over LIBERO-Plus and RoboTwin clean-to-random, plus a sharp reduction of the in-distribution-to-OOD gap on a real Franka platform under unseen colored lighting.

Load-bearing premise

The method assumes that the small set of query tokens left in the model at test time actually carries the future-scene meaning learned during training, rather than merely memorizing familiar appearances, and that the action stream relies on that meaning when the visuals change.

Editorial extensions

If this is right

  • Existing video-generation-based WAMs can gain visual out-of-distribution robustness through a post-training objective, without re-pretraining or replacing the latent space.
  • The same recipe transfers across action-expert WAMs (FastWAM, GE-Act) and a unified autoregressive WAM (LingBot-VA), so it is a general add-on rather than a bespoke fix.
  • Robustness gains are not confined to photometric shifts: the paper reports improvements under camera, layout, background, language, and sensor-noise perturbations as well.
  • Because the teacher and alignment head are dropped at inference, the method adds no semantic-feature computation at test time beyond the fixed query tokens.
  • The paper's comparison with a semantic-latent WAM suggests semantic representations help most when they enhance, rather than replace, pretrained video-generation dynamics priors.

Reading between the lines

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

  • Editorial inference: if the query tokens genuinely carry DINOv3-like future semantics at inference, their outputs should be predictable targets for probing; one could test whether the retained query states cluster by future-scene content rather than by training-scene appearance.
  • Editorial inference: the method points to a cheaper route to semantic robustness—instead of retraining a world model in semantic space, align a small set of auxiliary tokens inside an existing VAE-space model—which may transfer to other generative-policy families.
  • Editorial inference: a natural extension not explored in the paper is combining the CLS target with a geometric target through staged or weighted optimization; the paper's combined DINOv3-CLS-plus-depth ablation underperformed the single CLS target, suggesting simple concatenation is not the right way to fuse semantic and geometric foresight.
  • Editorial inference: the success under camera and layout perturbations hints that the alignment objective may also capture viewpoint-invariant content, but the paper does not isolate which semantic dimensions the queries encode, so this remains untested.
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 Robust-WAM, a post-training method for video-generation-based World-Action Models (WAMs). It preserves the existing VAE-based generative path and adds a semantic-foresight alignment objective in the action stream: learnable query tokens are prepended to the action tokens, and their output hidden states are aligned during training with frozen DINOv3 CLS embeddings of future ground-truth frames. At inference the teacher and alignment head are dropped, leaving only the query tokens in the action sequence. The method is evaluated on LIBERO-Plus, RoboTwin clean-to-random, and a real-robot illumination-shift setup, applied to three WAM backbones (FastWAM, GE-Act, LingBot-VA). The reported results show consistent improvements in out-of-distribution success rates (e.g., +9.2 on FastWAM LIBERO-Plus, +4.6 on RoboTwin, +22.7 OOD-average points on the real robot) with no degradation on clean benchmarks. Ablations indicate that the DINOv3 CLS target and the temporal positional encoding of queries outperform alternatives.

Significance. If the mechanism is as claimed, Robust-WAM addresses a real trade-off in WAMs: it allows models to keep large-scale VAE-space video pretraining while gaining appearance robustness, without retraining a semantic world model from scratch. The method is simple, architecture-agnostic, and uses a frozen external teacher (DINOv3) and external benchmarks, which keeps circularity low and yields a falsifiable prediction: the retained query tokens should continue to produce DINOv3-CLS-like future semantics under OOD appearance shifts. The breadth of evaluation across three backbones, two simulation benchmarks, and a real robot is a strength, and the ablations support that the choice of target and temporal indexing matter. However, the central causal claim that query tokens provide appearance-invariant semantic foresight at inference is not directly verified; the empirical support also lacks error bars or repeated-seed statistics. These are load-bearing gaps that affect the interpretation of every headline number.

major comments (3)
  1. [Section 3.2, Eq. (6) and inference paragraph] The mechanism claim that retained query tokens provide DINOv3-CLS-like semantic foresight at inference is unverified. At test time each query is a fixed learned embedding plus a positional encoding; it has no observation-dependent semantic input, so any semantics must arise through attention to the video and action tokens. The alignment loss L_align supervises only the query outputs after N action blocks and never constrains what the query inputs encode. Nothing in the training objective prevents the queries from acting as extra parameters that memorize training-distribution appearance, and the paper provides no probe of query outputs at inference, no attention analysis, and no control that removes or randomizes the semantic target. A random-target control (e.g., aligning queries to fixed random vectors or to an untrained encoder) and a probing experiment measuring cosine similarity between query outputs and DINOv3 CLS of future frames under OOD appearance shifts would directly test the mechanism. Without such evidence, the paper's central claim that the OOD gains are due to 'semantic foresight' rather than added capacity is not established.
  2. [Tables 1, 3, and 5] All success rates are reported as single runs with no error bars, confidence intervals, or repeated-seed statistics. Some headline gains are small (e.g., +0.3 and +0.8 on clean LIBERO in Table 1; +2.9 on GE-Act LIBERO-Plus), and the real-robot evaluation uses only 25 rollouts per condition with no variance reported. These omissions make it impossible to assess whether the improvements are statistically meaningful, especially for GE-Act on LIBERO-Plus and for the clean-suite comparisons. The paper should report means and standard deviations (or at least confidence intervals) over at least three seeds for the main comparisons, and ideally for the ablations in Table 4 as well.
  3. [Section 4.4, Table 4] The ablations do not include a control that keeps the K query tokens but removes the semantic alignment objective, or replaces the DINOv3 targets with random targets. The comparison between 'Baseline' (49.7) and 'Full temporal query' (58.9) confounds the effect of the alignment loss with the presence of additional learnable parameters. For example, K = 16 extra tokens in the action stream could improve OOD robustness by providing extra capacity or a different attention bias, independent of DINOv3 semantics. A variant with the same queries and no L_align, or with random fixed targets, would isolate the contribution of semantic foresight and is necessary to support the causal reading of Eq. (6) that the paper asserts.
minor comments (5)
  1. [Section 3.2] The target dimension d_z is used in the alignment objective before it is defined; consider stating d_z = 768 when introducing the DINOv3 CLS target in Section 3.2 rather than only in Section 4.1.
  2. [Table 2] The text says the robot-initial-state axis is the 'only axis that does not improve' for FastWAM, but the table shows a regression from 44.4 to 43.4; this should be described as a small degradation, not merely 'no improvement.'
  3. [Table 5] The 'gap ↓' column is visually confusing because the arrow suggests a lower value is better but the header does not define the gap; consider reporting the OOD success rate and the InD-minus-OOD difference directly, or stating the gap definition explicitly in the caption.
  4. [Section 4.1 and Section 3.3] The query count K is specified in the text (K=16 for LIBERO and real robot, K=24 for RoboTwin) but the derivation from T_f=8 and camera counts could be made more explicit with a small table or a formula near Eq. (4), since it is load-bearing for the reader's understanding of the method.
  5. [Section 5] The conclusion describes the alignment as having 'low computational overhead,' but no training-time or inference-time cost comparison (e.g., added parameters, FLOPs, or wall-clock time) is provided; a brief quantitative statement would support this claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Robust-WAM's alignment target is an external frozen encoder and its claimed gains are measured on external benchmarks; the unverified inference-time transfer mechanism is an empirical gap, not a circular reduction.

full rationale

The derivation chain is: Eqs. (1)-(3) keep the original WAM flow-matching losses, Eq. (5) prepends query tokens with the action-stream positional encoding, Eq. (6) supervises query outputs with cosine alignment to frozen DINOv3 CLS embeddings of ground-truth future frames, and Eq. (7) adds this as an auxiliary loss. The reported endpoint, success rate on LIBERO-Plus, RoboTwin clean-to-random, and real-robot illumination shifts, is measured by evaluation rollouts and is not computed from L_align or from the DINOv3 targets. The teacher is frozen and external to the paper's method, so Eq. (6) is not a self-defined target. The Section 3.2 claim that retained query tokens provide appearance-invariant foresight at inference is an inductive mechanism assumption: it could fail if the queries memorize training-distribution appearance, and the paper does not probe query outputs. However, an unverified mechanism is an evidence gap, not circularity, because the gains are still benchmark-measured. The Table 4 ablations choose DINOv3 CLS and lambda_align=0.1 using LIBERO-Plus itself; this is test-set selection and weakens statistical strength, but it is standard model selection rather than a fitted quantity renamed as a prediction. Self-citations such as S-VAM, FRAPPE, ReconVLA, FlowVLA, and DualCoT appear only in the introduction and related-work enumeration; none is invoked as justification for Robust-WAM's alignment objective or for the OOD improvement claim, and no uniqueness theorem or ansatz is imported from prior author work. The paper also states that teacher features are never provided as inputs to the queries during training and only supervise query outputs through Lalign, which further confirms that the alignment target is not fed into the model as a prediction input. Accordingly, no circular step can be exhibited with the required reduction, and the paper is not circular.

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

The central mechanism rests on a frozen external semantic encoder (DINOv3) and on the assumption that query tokens transfer semantic foresight at inference. The only hand-set hyperparameter is lambda_align=0.1; query tokens are learned parameters whose inference-time behavior is unverified.

free parameters (1)
  • Alignment loss weight lambda_align = 0.1
    Set to 0.1 across all experiments and not ablated; it may have been tuned on the evaluation benchmarks. The central OOD gains depend on this weighting.
assumptions (4)
  • domain assumption DINOv3 CLS embeddings are largely invariant to appearance (texture, illumination) while preserving scene content.
    Invoked in Section 3.2 to justify the alignment target; no invariance measurements under the tested perturbations are provided.
  • ad hoc to paper Query tokens continue to encode appearance-invariant future semantics at inference after the teacher is dropped.
    Section 3.2: 'At inference, we therefore drop the DINOv3 teacher and the head g while retaining the K query tokens.' The transfer of semantic foresight from training targets to inference-time query outputs is not directly verified.
  • domain assumption The pretrained VAE-space VGM dynamics prior is worth retaining and is compatible with adding semantic alignment to the action stream.
    Foundation of the method's motivation (Section 1); supported by experimental comparison to LDA-1B but not independently proven.
  • domain assumption LIBERO-Plus and RoboTwin clean-to-random perturbations are representative proxies for visual OOD shifts in real deployment.
    Used for evaluation; real-robot test only covers illumination shifts.
invented entities (1)
  • Learnable query tokens q (K tokens, one per future frame and camera view)
    purpose: Carry appearance-invariant semantic foresight into the action DiT stream; aligned to DINOv3 CLS of future GT frames during training and retained at inference.
    The paper provides no direct evidence (e.g., probing or attention analysis) that these tokens encode the intended semantic content at inference; their efficacy is inferred only from end-task success rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models." pith.science (2026). https://pith.science/paper/I4R2AXJJ

@misc{pith2026260805903,
  author       = {Pith},
  title        = {Pith review of: Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4R2AXJJ}},
  note         = {Machine review of arXiv:2608.05903}
}
read the original abstract

Mainstream World-Action Models (WAMs) adapt pretrained video generation models (VGMs) for robot control, transferring their learned dynamics prior for action prediction. These VGMs are typically trained in a variational autoencoder (VAE) latent space. However, the VAE latent space is optimized for pixel reconstruction, which rewards fine appearance detail and leaves the action prediction fragile under visual shifts. Recent works build WAMs in semantic latent space, which are more robust to appearance shifts. However, these models cannot leverage the large-scale VGM pretraining that exists only in VAE space. To overcome this dilemma, we propose Robust-WAM, a general post-training method for video-generation-based WAMs that preserves the VAE-based generative path and adds a lightweight semantic foresight alignment objective on the action stream. This retains the large-scale VGM pretraining while grounding actions in appearance-invariant dynamics that stay reliable under illumination shifts and other visual out-of-distribution conditions. Specifically, we employ learnable query tokens to bring future-scene semantics into the action stream by aligning their output hidden states with the semantic foresight of future ground-truth frames. To establish the temporal correspondence between each query and the future step it describes, we give it the positional encoding of the matching action tokens. Experiments on out-of-distribution generalization simulation benchmarks and a real-robot setup show that our Robust-WAM consistently improves the success rates of multiple WAM baselines without sacrificing in-distribution performance.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 5 canonical work pages

  1. [1]

    V-JEPA 2 : Self-supervised video models enable understanding, prediction and planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, Xiao...

    arXiv doi:10.48550/arxiv.2506.09985 2025

  2. [2]

    How learning by reconstruction produces uninformative features for perception

    Randall Balestriero and Yann LeCun. How learning by reconstruction produces uninformative features for perception. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 2566--2585. PMLR, 2024. https://proceedings.mlr.press/v235/balestriero24b.html

    2024

  3. [3]

    Motus: A unified latent action world model

    Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, Hongyan Zhao, Hanyu Liu, Zhizhong Su, Lei Ma, Hang Su, and Jun Zhu. Motus: A unified latent action world model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 35101--35113, 202...

    2026

  4. [4]

    Gr00t n1: An open foundation model for generalist humanoid robots

    Johan Bjorck, Fernando Casta \ n eda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025

    arXiv 2025

  5. [5]

    _0 : A vision-language-action flow model for general robot control

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. _0 : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024

    arXiv

  6. [6]

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

    Qingwen Bu, Yanting Yang, Jisong Cai, Shenyuan Gao, Guanghui Ren, Maoqing Yao, Ping Luo, and Hongyang Li. Univla: Learning to act anywhere with task-centric latent actions. arXiv preprint arXiv:2505.06111, 2025

    arXiv 2025

  7. [7]

    Worldvla: Towards autoregressive action world model

    Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, et al. Worldvla: Towards autoregressive action world model. arXiv preprint arXiv:2506.21539, 2025

    arXiv 2025

  8. [8]

    GR-2 : A generative video-language-action model with web-scale knowledge for robot manipulation

    Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, Hanbo Zhang, and Minzhao Zhu. GR-2 : A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. doi:10.48550/arXiv.2410.06158. https://arxiv.org/abs/2410.06158

    arXiv doi:10.48550/arxiv.2410.06158

Show all 49 references
  1. [9]

    Lawam: Latent world action models for efficient dynamics-aware robot policies

    Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, et al. Lawam: Latent world action models for efficient dynamics-aware robot policies. arXiv preprint arXiv:2606.15768, 2026 a

    2026

  2. [10]

    RoboTwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, Weiliang Deng, Yubin Guo, Tian Nian, Xuanbing Xie, Qiangyu Chen, Kailun Su, Tianling Xu, Guodong Liu, Mengkang Hu, Huan-ang Gao, Kaixuan Wang, Zhixuan L...

    2026

  3. [11]

    Tenenbaum, Dale Schuurmans, and Pieter Abbeel

    Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. https://arxiv.org/abs/2302.00111

    2023 arXiv

  4. [12]

    LIBERO-Plus : A progressive robustness benchmark for visual-language-action models

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. LIBERO-Plus : A progressive robustness benchmark for visual-language-action models. In Proceedings of the IEEE/CVF Conf...

    2026

  5. [13]

    Prediction with action: Visual policy learning via joint denoising process

    Yanjiang Guo, Yucheng Hu, Jianke Zhang, Yen-Jen Wang, Xiaoyu Chen, Chaochao Lu, and Jianyu Chen. Prediction with action: Visual policy learning via joint denoising process. In Advances in Neural Information Processing Systems, volume 37, pages 112386--112410. Curran Associates...

    doi

  6. [14]

    World models

    David Ha and J \"u rgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2 0 (3): 0 440, 2018

    2018 arXiv

  7. [15]

    Nora: A small open-sourced generalist vision language action model for embodied tasks, 2025

    Chia-Yu Hung, Qi Sun, Pengfei Hong, Amir Zadeh, Chuan Li, U-Xuan Tan, Navonil Majumder, and Soujanya Poria. Nora: A small open-sourced generalist vision language action model for embodied tasks, 2025. https://arxiv.org/abs/2504.19854

    2025 arXiv

  8. [16]

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

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024

    2024 arXiv

  9. [17]

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

    Moo Jin Kim, Chelsea Finn, Percy Liang, et al. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025

    2025 arXiv

  10. [18]

    Cosmos policy: Fine-tuning video models for visuomotor control and planning

    Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, and Jinwei Gu. Cosmos policy: Fine-tuning video models for visuomotor control and planning. arXiv preprint arXiv:2601.16163, 2026. doi:10.48550/arXiv...

    2026 doi arXiv

  11. [19]

    A path towards autonomous machine intelligence version 0.9

    Yann LeCun et al. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62 0 (1): 0 1--62, 2022

    2022

  12. [20]

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

    Fuhao Li, Wenxuan Song, Han Zhao, Jingbo Wang, Pengxiang Ding, Donglin Wang, Long Zeng, and Haoang Li. Spatial forcing: Implicit spatial representation alignment for vision-language-action model. arXiv preprint arXiv:2510.12276, 2025

  13. [21]

    Causal world modeling for robot control

    Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, Yujun Shen, and Yinghao Xu. Causal world modeling for robot control. arXiv preprint arXiv:2601.21998, 2026. doi:10.48550/arXiv.2601.21998. https://arxiv.org/abs/2601.21998

    2026 doi arXiv

  14. [22]

    Genie envisioner: A unified world foundation platform for robotic manipulation

    Yue Liao, Pengfei Zhou, Siyuan Huang, Donglin Yang, Shengcong Chen, Yuxin Jiang, Yue Hu, Si Liu, Jianlan Luo, Liliang Chen, Shuicheng Yan, Maoqing Yao, and Guanghui Ren. Genie envisioner: A unified world foundation platform for robotic manipulation. In The Fourteenth Internati...

    2026

  15. [23]

    Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. In The Fourteenth International Conference on Learning Repres...

    2026

  16. [24]

    Evo-0: Vision-language-action model with implicit spatial understanding

    Tao Lin, Gen Li, Yilei Zhong, Yanwen Zou, Yuxin Du, Jiting Liu, Encheng Gu, and Bo Zhao. Evo-0: Vision-language-action model with implicit spatial understanding. arXiv preprint arXiv:2507.00416, 2025

    2025

  17. [25]

    LIBERO : Benchmarking knowledge transfer for lifelong robot learning

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO : Benchmarking knowledge transfer for lifelong robot learning. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2023. https://arxiv.org/abs/2...

    2023 arXiv

  18. [26]

    LDA-1B : Scaling latent dynamics action model via universal embodied data ingestion

    Jiangran Lyu, Kai Liu, Xuheng Zhang, Haoran Liao, Yusen Feng, Wenxuan Zhu, Tingrui Shen, Jiayi Chen, Jiazhao Zhang, Yifei Dong, Wenbo Cui, Senmao Qi, Shuo Wang, Yixin Zheng, Mi Yan, Xuesong Shi, Haoran Li, Dongbin Zhao, Ming-Yu Liu, Zhizheng Zhang, Li Yi, Yizhou Wang, and He W...

    2026

  19. [27]

    Reconstruction or semantics? what makes a latent space useful for robotic world models

    Nilaksh , Saurav Jha, Artem Zholus, and Sarath Chandar. Reconstruction or semantics? what makes a latent space useful for robotic world models. arXiv preprint arXiv:2605.06388, 2026. doi:10.48550/arXiv.2605.06388. https://arxiv.org/abs/2605.06388

    2026 doi arXiv

  20. [28]

    Cosmos world foundation model platform for physical AI

    NVIDIA et al. Cosmos world foundation model platform for physical AI . arXiv preprint arXiv:2501.03575, 2025. doi:10.48550/arXiv.2501.03575. https://arxiv.org/abs/2501.03575

    2025 doi arXiv

  21. [29]

    mimic-video : Video-action models for generalizable robot control beyond VLA s

    Jonas Pai, Liam Achenbach, Oliver Sanchez, Stefanos Charalambous, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava. mimic-video : Video-action models for generalizable robot control beyond VLA s. In Proceedings of Robotics: Science and Systems (RSS), 2026. http...

    2026

  22. [30]

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

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025

    2025 arXiv

  23. [31]

    World action models: A survey

    Qiuhong Shen, Shihua Zhang, Yue Liao, Qi Li, Zhenxiong Tan, Shizun Wang, Shuicheng Yan, and Xinchao Wang. World action models: A survey. arXiv preprint arXiv:2606.20781, 2026. doi:10.48550/arXiv.2606.20781. https://arxiv.org/abs/2606.20781

    2026 doi arXiv

  24. [32]

    Oriane Sim \'e oni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha \"e l Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth \'e e Darcet, Th \'e o Moutakanni, Le...

    2025 doi arXiv

  25. [33]

    ReconVLA : Reconstructive vision-language-action model as effective robot perceiver

    Wenxuan Song, Ziyang Zhou, Han Zhao, Jiayi Chen, Pengxiang Ding, Haodong Yan, Yuxin Huang, Feilong Tang, Donglin Wang, and Haoang Li. ReconVLA : Reconstructive vision-language-action model as effective robot perceiver. arXiv preprint arXiv:2508.10333, 2025. doi:10.48550/arXiv....

    2025 doi arXiv

  26. [34]

    Wan: Open and advanced large-scale video generative models

    Team Wan et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. doi:10.48550/arXiv.2503.20314. https://arxiv.org/abs/2503.20314

    2025 doi arXiv

  27. [35]

    Repwam: World action modeling with representation visual-action tokenizers

    Junke Wang, Qihang Zhang, Shuai Yang, Yiming Luo, Yujun Shen, Zuxuan Wu, Yu-Gang Jiang, and Yinghao Xu. Repwam: World action modeling with representation visual-action tokenizers. arXiv preprint arXiv:2606.13674, 2026

    2026 arXiv

  28. [36]

    Unleashing large-scale video generative pre-training for visual robot manipulation

    Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations (ICLR), 2024. https://a...

    2024 arXiv

  29. [37]

    FutureVLA : Joint visuomotor prediction for vision-language-action model

    Xiaoxu Xu, Hao Li, Jinhui Ye, Yilun Chen, Jia Zeng, Xinyi Chen, Linning Xu, Dahua Lin, Weixin Li, and Jiangmiao Pang. FutureVLA : Joint visuomotor prediction for vision-language-action model. arXiv preprint arXiv:2603.10712, 2026. doi:10.48550/arXiv.2603.10712. https://arxiv.o...

    2026 doi

  30. [38]

    Open-world hand-object interaction video generation based on structure and contact-aware representation

    Haodong Yan, Hang Yu, Zhide Zhong, Weilin Yuan, Xin Gong, Zehang Luo, Chengxi Heyu, Junfeng Li, Wenxuan Song, Shunbo Zhou, and Haoang Li. Open-world hand-object interaction video generation based on structure and contact-aware representation. arXiv preprint arXiv:2512.01677, 2...

    2025 doi

  31. [39]

    S-VAM : Shortcut video-action model by self-distilling geometric and semantic foresight

    Haodong Yan, Zhide Zhong, Jiaguan Zhu, Junjie He, Weilin Yuan, Wenxuan Song, Xin Gong, Yingjie Cai, Guanyi Zhao, Xu Yan, Bingbing Liu, Ying-Cong Chen, and Haoang Li. S-VAM : Shortcut video-action model by self-distilling geometric and semantic foresight. arXiv preprint arXiv:2...

    2026 doi

  32. [40]

    Starvla- : Reducing complexity in vision-language-action systems

    Jinhui Ye, Ning Gao, Senqiao Yang, Jinliang Zheng, Zixuan Wang, Yuxin Chen, Pengguang Chen, Yilun Chen, Shu Liu, and Jiaya Jia. Starvla- : Reducing complexity in vision-language-action systems. arXiv preprint arXiv:2604.11757, 2026 a

    2026 arXiv

  33. [41]

    World action models are zero-shot policies

    Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, Ayaan Malik, Kyungmin Lee, William Liang, Nadun Ranawaka, Jiasheng Gu, Yinzhen Xu, Guanzhi Wang, Fengyuan Hu, Avnish Narayan, Johan Bjor...

    2026 doi arXiv

  34. [42]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024

    arXiv

  35. [43]

    Fast-WAM : Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026

    Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-WAM : Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026. doi:10.48550/arXiv.2603.16666. https://arxiv.org/abs/2603.16666

    2026 doi arXiv

  36. [44]

    Do world action models generalize better than VLA s? a robustness study

    Zhanguang Zhang, Zhiyuan Li, Behnam Rahmati, Rui Heng Yang, Yintao Ma, Amir Rasouli, Sajjad Pakdamansavoji, Yangzheng Wu, Lingfeng Zhang, Tongtong Cao, Feng Wen, Xinyu Wang, Xingyue Quan, and Yingxue Zhang. Do world action models generalize better than VLA s? a robustness stud...

    2026 doi arXiv

  37. [45]

    FRAPPE : Infusing world modeling into generalist policies via multiple future representation alignment

    Han Zhao, Jingbo Wang, Wenxuan Song, Shuai Chen, Yang Liu, Yan Wang, Haoang Li, and Donglin Wang. FRAPPE : Infusing world modeling into generalist policies via multiple future representation alignment. arXiv preprint arXiv:2602.17259, 2026. doi:10.48550/arXiv.2602.17259. https...

    2026 doi

  38. [46]

    FLARE : Robot learning with implicit world modeling

    Ruijie Zheng, Jing Wang, Scott Reed, Johan Bjorck, Yu Fang, Fengyuan Hu, Joel Jang, Kaushil Kundalia, Zongyu Lin, Loic Magne, Avnish Narayan, You Liang Tan, Guanzhi Wang, Qi Wang, Jiannan Xiang, Yinzhen Xu, Seonghyeon Ye, Jan Kautz, Furong Huang, Yuke Zhu, and Linxi Fan. FLARE...

    2025

  39. [47]

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

    Zhide Zhong, Haodong Yan, Junfeng Li, Xiangchen Liu, Xin Gong, Tianran Zhang, Wenxuan Song, Jiayi Chen, Xinhu Zheng, Hesheng Wang, and Haoang Li. FlowVLA : Visual chain of thought-based motion reasoning for vision-language-action models. arXiv preprint arXiv:2508.18269, 2025. ...

    2025 doi

  40. [48]

    DualCoT-VLA : Visual-linguistic chain of thought via parallel reasoning for vision-language-action models

    Zhide Zhong, Junfeng Li, Junjie He, Haodong Yan, Xin Gong, Guanyi Zhao, Yingjie Cai, Jiantao Gao, Xu Yan, Bingbing Liu, Yingcong Chen, Liuqing Yang, and Haoang Li. DualCoT-VLA : Visual-linguistic chain of thought via parallel reasoning for vision-language-action models. arXiv ...

    2026 doi

  41. [49]

    DINO-WM : World models on pre-trained visual features enable zero-shot planning

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. DINO-WM : World models on pre-trained visual features enable zero-shot planning. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 79115...

    2025

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.