Pith. sign in

REVIEW 3 major objections 4 minor 44 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a post-training objective aligning action-stream queries with semantic embeddings of future frames restores out-of-distribution robustness in video-generation world-action models without sacrificing in-distribution…

desk verdict A clean, plausible post-training fix for OOD robustness in VAE-latent world-action models, with a believable experiment suite but an under-tested mechanism. read the letter →

arxiv 2608.05903 v1 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-actionmodelssemanticforesightalignmentvideogenerationout-of-distributiongeneralizationrobotmanipulationDINOv3CLSembeddingsrepresentationVAElatentspace
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

World-action models that build on pretrained video generators inherit strong dynamics priors, but those priors live in a VAE latent space tuned for pixel reconstruction, so action predictions break under illumination and other appearance shifts. Semantic-latent world models are more robust but cannot reuse the large-scale video pretraining. Robust-WAM tries to get both: it keeps the VAE generative path untouched and adds a lightweight alignment loss on the action stream, steering learnable query tokens toward frozen semantic summaries of the ground-truth future frames. If correct, this gives a plug-and-play post-training recipe that lifts out-of-distribution success rates on simulation benchmarks and a real robot while preserving in-distribution performance.

What carries the argument

The load-bearing mechanism is semantic foresight alignment. Learnable query tokens are prepended to the noised action tokens, one per future frame and camera view, and each query reuses the positional encoding of the action step that reaches the future frame it describes. After the action DiT processes the augmented sequence, a linear head maps each query's output to be cosine-aligned with the frozen DINOv3 CLS embedding of that frame's ground-truth appearance. The queries and action tokens attend to each other through the WAM's existing attention, so action generation can read future semantics while the video-generation loss continues to train the VAE path; at inference only the query tokens remain, and the DINOv3 teacher and alignment head are dropped.

What would settle it

Train the same Robust-WAM variant but align the query outputs to a scrambled or non-semantic target, such as the CLS embedding of a time-shuffled or noise-corrupted future frame, while keeping the same query count, positional encodings, and loss weight; if out-of-distribution success stays as high as with the true future-frame CLS, the semantic content of the target is not carrying the robustness.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the appearance bias of VAE-latent world-action models can be corrected in the action representation rather than by replacing the latent space. Prepending one learnable query token per future frame and camera view to the action sequence, and aligning each query's output with the frozen DINOv3 CLS embedding of the corresponding ground-truth future frame, gives the action stream appearance-invariant foresight: the semantic target summarizes objects and layout while discarding texture and illumination. Because the video DiT, its VAE tokenization, and the video-generation objective remain unchanged, the large-scale generative pretraining is preserved, and the same recipe works on action-expert WAMs and on a unified video-action transformer.

Load-bearing premise

The load-bearing premise is that the DINOv3 CLS embedding of the ground-truth future frame is an appearance-invariant summary of exactly the task-relevant future state, so aligning the action stream to it, rather than the added parameters or auxiliary loss, is what produces the robustness gains.

Editorial extensions

If this is right

  • If Robust-WAM is correct, a VAE-latent world-action model can gain appearance robustness without retraining or replacing its video generator: the VAE tokenization, the video DiT, and the flow-matching video loss all stay as they are, and only the action stream receives the extra semantic supervision.
  • The recipe transfers across transformer architectures: it improves the action-expert WAMs FastWAM and GE-Act on LIBERO-Plus and the unified video-action WAM LingBot-VA on RoboTwin clean-to-random, so the post-training objective is not tied to one sequence layout.
  • The gains are not limited to pixel-level photometric shifts; on FastWAM the per-axis LIBERO-Plus improvements include sensor noise (+18.9 points), camera perturbations (+12.4), lighting (+11.3), and background changes (+10.3), with clean LIBERO success unchanged or slightly higher.
  • In the real-robot evaluation, Robust-WAM built on GE-Act reduces the average in-distribution-to-OOD gap from 22.7 to 2.7 points under unseen purple and cyan illumination, implying the method can close most of the appearance-shift failure in deployment.

Reading between the lines

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

  • Editorial inference: because the ablations test only semantic or geometric targets, they do not rule out that part of the gain is a regularization effect of the auxiliary loss; a scrambled-target control would isolate the semantic contribution.
  • Editorial inference: the query-plus-CLS mechanism does not logically depend on a video generator, so the same semantic foresight alignment could be applied to a vision-language-action policy or an inverse-dynamics model; the paper only demonstrates it inside WAMs.
  • Editorial inference: the per-axis gains on layout and camera perturbations suggest a testable extension, namely measuring whether the semantic target also stabilizes action prediction under unseen object poses and geometric rearrangements in real deployment, beyond illumination shifts.
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 / 4 minor

Summary. Robust-WAM is a post-training procedure for video-generation-based world-action models (WAMs). It prepends learnable query tokens to the action DiT sequence, assigns them the positional encoding of the action step that corresponds to each future frame, and trains their output hidden states to cosine-align with frozen DINOv3 CLS embeddings of ground-truth future frames, while leaving the VAE-space video generation path and its flow-matching losses unchanged. At inference the DINOv3 teacher and the alignment head are dropped, and only the query tokens remain in the action stream. The paper reports improved out-of-distribution success rates on LIBERO-Plus for FastWAM and GE-Act, on RoboTwin clean-to-random for LingBot-VA, and on a real Franka illumination-shift setup for GE-Act, with small or no loss on in-distribution performance. Ablations compare different alignment targets (DINOv3 CLS, DINOv3 Patch, DepthAnything3, and their combination) and query-design choices (with/without temporal positional encoding, future versus current-frame targets).

Significance. The empirical pattern is consistent and the method is simple, clearly specified, and plausibly general: it addresses a recognized trade-off between VAE-latent VGM pretraining and appearance robustness, and the reported gains appear across two benchmarks, three WAM architectures (two action-expert and one unified), and a real-robot setup. The paper contains useful ablations in Table 4 and draws a clear distinction from previous future-representation alignment methods such as FLARE and FRAPPE. However, the central mechanistic claim—that appearance-invariant semantic foresight from DINOv3 CLS is the causal ingredient behind the gains—is not isolated by the experiments, and none of the success-rate claims carry uncertainty estimates. If the missing controls are added and the statistical robustness is established, this would be a solid contribution to robot learning with generative world models.

major comments (3)
  1. [Sec. 4.4, Table 4] The ablation design does not establish that the DINOv3 CLS target, rather than the added query capacity or the auxiliary loss, drives the gains. Every variant in Table 4 adds the same query mechanism and an alignment loss with some target, so the +9.2 point improvement over FastWAM cannot be attributed specifically to appearance-invariant semantics. The paper needs two controls: (i) the same augmented action stream with K=16 query tokens but no alignment loss, and (ii) alignment to a non-semantic, appearance-sensitive target (for example, the VAE latent of the same future frame) or to a random fixed projection. The current-frame variant is itself a semantic DINOv3 target and therefore does not control for semantics. Because Eq. 6 and the paper's framing make the semantic foresight the load-bearing claim, without these controls the central interpretation is underdetermined.
  2. [Secs. 4.2-4.5, Tables 1-5] Success rates are reported as point estimates without error bars, confidence intervals, or multiple seeds; the real-robot evaluation in Table 5 uses only 25 rollouts per condition. The magnitude of some headline gains (e.g., GE-Act +2.9 on LIBERO-Plus in Table 1 and LingBot-VA +4.6 on RoboTwin random in Table 3) is small relative to what binomial sampling noise would produce at the reported evaluation sizes, and the 'without sacrificing in-distribution performance' claim is likewise unsupported by any variance measure. I ask the authors to report mean and standard deviation over at least three seeds for the simulation benchmarks, or binomial confidence intervals, and to state the number of evaluation episodes per cell and condition.
  3. [Sec. 3.2, Eq. 6 and Sec. 4.4] The paper never verifies that the query outputs actually encode the DINOv3 CLS semantics at inference, when the teacher and the alignment head are dropped. The success-rate improvements are also consistent with the queries acting as extra capacity or as a task-specific regularizer. Reporting the cosine similarity between query outputs and the frozen CLS targets on held-out OOD frames, or ablating the presence of the query tokens at inference, would directly test whether the action DiT has acquired appearance-invariant foresight. This evidence is needed to support the mechanism claimed in the abstract and in Sec. 3.2.
minor comments (4)
  1. [Sec. 4.2, Results] The comparison of GE-Act with Robust-WAM against LDA-1B is used to conclude that semantic representations should enhance rather than replace VGM pretraining, but LDA-1B differs in data scale, model size, and training recipe; this conclusion is illustrative rather than established by a controlled comparison.
  2. [Table 5] The GE-Act row appears to have a formatting error ('247248' instead of '24 72 48 24'); please fix the table alignment so all cells are unambiguous.
  3. [Author list and Related Work] There are several spacing and hyphenation typos, including 'MingMing Y u', 'Y uqiao Du', 'large-scaledynamics', 'latentworld-actionmodel', and 'whatthey align to'; these should be corrected.
  4. [Sec. 4.1 and Eq. 4] Please state explicitly that the frame stride Delta is a positive integer number of control steps, and define the relationship between the action horizon H and the number of action tokens L in the unified WAM case, so that Eq. 4 and the clamp min(j*Delta, L) are unambiguous.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central OOD claims are measured on external benchmarks and the semantic target is a frozen external DINOv3 teacher; only minor, non-load-bearing self-citations are present.

full rationale

Robust-WAM's main result is an empirical success-rate comparison on LIBERO-Plus, RoboTwin, and a real-robot setup, all external to the paper's own equations. The alignment objective in Eq. 6 is a training-time auxiliary loss that supervises query-token outputs against frozen DINOv3 CLS embeddings of ground-truth future frames; at inference the teacher and alignment head are dropped, so the reported OOD gains are not a renamed fit of the training objective. The paper's load-bearing assumptions (VAE latents are appearance-biased, DINOv3 CLS features are appearance-invariant) are supported by external citations and external frozen features, not by the authors' own definitions. The ablation in Table 4 varies the target and query design but does not include a no-alignment parameter-matched control or a non-semantic target; that is a legitimate evidence-strength limitation, but it is a correctness/identification concern, not circularity, because nothing in the derivation defines the outcome in terms of the input. Two related-work citations (FRAPPE, Spatial Forcing) share authors with the present paper, but they are used only as contextual examples of representation alignment and are not load-bearing for any equation, benchmark, or conclusion. The derivation chain is therefore self-contained against external evidence.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are postulated; the learnable query tokens are trainable parameters, not independent entities. The central assumptions are about the usefulness of the frozen DINOv3 teacher and the transfer of alignment to inference.

free parameters (2)
  • alignment loss weight lambda_align = 0.1
    Chosen by hand to balance the auxiliary alignment loss against the WAM objective (Eq. 7); no sensitivity analysis is reported.
  • query token count K (one per future frame and camera) = 16 for LIBERO and real robot, 24 for RoboTwin
    Design choice setting one learnable query per future frame and camera; no ablation varies K, so its influence on the result is unknown.
assumptions (3)
  • domain assumption DINOv3 CLS embeddings of ground-truth future frames are appearance-invariant and carry task-relevant foresight.
    Sec. 3.2 asserts CLS summarizes scene content while discarding texture and illumination; the paper gives no verification on its own domains and no non-semantic control.
  • domain assumption Cosine alignment of query outputs to these targets transfers useful signal to the action stream at inference.
    Eq. 6 is assumed to drive the action tokens to read appearance-invariant future semantics; the ablations change targets and query design but do not isolate alignment from extra parameters or auxiliary loss.
  • domain assumption The LIBERO-Plus, RoboTwin clean-to-random, and real illumination shifts are representative of visual OOD conditions.
    The paper generalizes from these benchmarks to visual OOD conditions; the real-robot test only covers illumination, not camera, background, or sensor noise.

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.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 13 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

    work page 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...

    work page 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 44 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]

    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

  26. [34]

    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

  27. [35]

    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

  28. [36]

    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

  29. [37]

    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

  30. [38]

    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

  31. [39]

    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

  32. [40]

    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

  33. [41]

    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

  34. [42]

    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

  35. [43]

    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

  36. [44]

    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 7, 2026 · model on record in the stance chip above.