Pith. sign in

REVIEW 4 major objections 8 minor 32 references

DyPES-VLA: Learning Shared Dynamics Priors and Embodiment-Specific Control for Cross-Embodiment Manipulation

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single checkpoint controls single-arm, dual-arm, and humanoid robots by learning shared dynamics before embodiment-specific actions.

desk verdict A solid cross-embodiment VLA with a genuinely new architecture, but the flagship claim about shared dynamics priors rests on a confounded ablation; still deserves a serious referee. read the letter →

arxiv 2608.06374 v1 pith:IILDG7KQ submitted 2026-08-06 cs.RO

Junfeng Li , Junjie He , Zhide Zhong , Yangyang Zheng , Pingyue Sheng , Jiayu Dong , Ruixin Li , Haodong Yan
show 7 more authors
Jiaguan Zhu Tianran Zhang Runze Yu Wen Chen Liuqing Yang Yuxiang Gao Haoang Li
This is my paper · ORCID
classification cs.RO
keywords cross-embodimentmanipulationvision-language-actionmodelsfuturepredictionmixture-of-expertsshareddynamicspriorsembodiment-specificcontrolrobot
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 proposes DyPES-VLA, a way to train one vision-language-action policy that controls robots with very different bodies—single arm, dual arm, and humanoid—without manually converting their actions into a common format. It argues that future-frame prediction, applied to action-free human and robot videos, makes the shared visual representation carry the dynamics of contact, object motion, and scene change. A second claim is that action generation should not be shared: an embodiment-specific mixture-of-experts head decodes the shared representation directly into each robot's native action space. If correct, this separates the part of manipulation that transfers across embodiments from the part tied to a specific body, allowing a single checkpoint to match or beat both per-benchmark specialists and generalist baselines. The paper reports 98.0% success on a single-arm simulation benchmark, 59.25% on a humanoid benchmark, and 89.02% on a dual-arm benchmark, plus 75.6% average success across three real-world robots.

What carries the argument

The load-bearing object is the shared query state $Z$: a set of learnable query tokens processed by a pretrained vision-language model together with the observation, instruction, and verbalized embodiment metadata. In the first training stage these states are supervised only by a rectified-flow future-frame generation loss, with the generator conditioned solely on $Z$, so predictive regularities must be stored in $Z$ itself. In the second stage the same states condition an embodiment-specific mixture-of-experts action head, a flow-matching diffusion transformer whose attention layers are shared and whose feed-forward experts are chosen by static routing; lightweight per-embodiment encoders and decoders map between the shared width and each native action space. The future-generation head is discarded at inference. This mechanism's job is to keep the representation embodiment-neutral at the level of dynamics while keeping control realization embodied.

What would settle it

Ablate only the human-video portion of the first pretraining stage's video mixture: retrain with the same total video budget drawn entirely from the three robot simulation data sources, keeping every other hyperparameter fixed, and compare the three benchmark scores and the real-world average against the reported numbers. If removing human video leaves performance unchanged, the claim that egocentric human dynamics transfer to robot embodiments fails, even though future prediction itself may still help.

Watch

Extended reading notes

Core claim

The central claim is that cross-embodiment manipulation can be decomposed into shared dynamics priors and embodiment-specific control. DyPES-VLA trains a pretrained vision-language model together with learnable query tokens to predict a future camera frame; because the future-generation head sees only those query states, whatever is needed to synthesize the future must pass through them, and those states become the shared interface between perception and action. A flow-matching action head, with attention layers shared across embodiments and feed-forward experts selected by embodiment metadata, then decodes the same query states directly into action chunks in each robot's native action space. This lets heterogeneous robots be co-trained without pre-aligning their action formats. The paper claims that, with a single co-trained checkpoint, this paradigm outperforms prior specialists and generalists on all three simulation benchmarks and, after joint finetuning on 1,800 real-world demonstrations, averages 75.6% success across three physical robot morphologies.

Load-bearing premise

The load-bearing premise is that teaching the model to predict future frames from action-free videos, including egocentric human hand videos, writes shared dynamics into the representation that truly transfer to robot control; if human-hand dynamics do not transfer to robot arms and humanoids, the first pretraining stage is wasted or harmful and the reported gains would need another explanation.

Editorial extensions

If this is right

  • A single checkpoint can control robots with different morphologies, degrees of freedom, and control frequencies without action-space alignment.
  • Action-free video sources, including egocentric human manipulation videos, can contribute to robot policy learning through future prediction even though they contain no executable actions.
  • Embodiment metadata lets one policy disambiguate which robot is acting, and static routing prevents heterogeneous control spaces from interfering during joint training.
  • Because the future-generation head is removed at inference, the cross-embodiment transfer gains come without additional test-time video generation cost.
  • Future supervision makes contact onset and release linearly decodable from the query states, evidence that the representation tracks interaction dynamics rather than only visible current contact.

Reading between the lines

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

  • The separation of shared dynamics from embodiment-specific control suggests a scaling recipe: add a new robot family by adding one expert encoder-decoder pair to the action head and adding its data to future-prediction pretraining, without re-engineering a common action representation.
  • A strong test the paper leaves implicit is whether a genuinely novel embodiment, absent from the second-stage co-training, can be controlled after only expert initialization plus a small amount of its own data; the real-world finetuning reuses experts already trained on the matching simulation embodiment, so it does not demonstrate true zero-shot body transfer.
  • A concrete testable extension is to ablate the source mixture of the first-stage videos, comparing simulation-only, human-video-only, and mixed pretraining; the current ablation removes the whole future-prediction objective, not the human-video component, so it does not isolate where the transfer gain comes from.
  • If the dynamics-prior story is right, the same future-supervised query states could seed other downstream heads, such as subgoal proposal or grasp-point prediction, without retraining the shared backbone.
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

4 major / 8 minor

Summary. The paper proposes DyPES-VLA, a two-stage cross-embodiment VLA that learns a shared query representation through future-frame prediction and then decodes it into native action spaces through an embodiment-specific Mixture-of-Experts action head. Stage 1 pretrains the VLM, query tokens, and a SANA future-generation head on action-free human and simulation videos; Stage 2 jointly optimizes the future-prediction and action objectives on action-labeled demonstrations from RoboTwin 2.0, RoboCasa-GR1, and LIBERO. The resulting single checkpoint is evaluated on the three simulation benchmarks and, after joint finetuning, on three physical robots (FR3, COBOT Magic, G1), with reported success rates of 89.02%, 59.25%, and 98.0% in simulation and a 75.6% average in the real-world study. Ablations target the future-prediction objective, Stage-1 pretraining, the MoE head, and embodiment metadata, and a linear probe tests contact decodability.

Significance. If the results hold, the paper makes a useful architectural contribution: separating a future-supervised shared representation from embodiment-specific action decoding is a clean way to avoid manual action-space alignment, and the two-stage training protocol is clearly specified. The paper also has genuine strengths: the future-prediction objective is an independent training signal rather than a circular byproduct of action prediction, the linear probe is carefully designed to avoid trivial prediction by conditioning on the current contact state, and the evaluation spans three morphologies in both simulation and real hardware, with per-task breakdowns in the supplementary material. However, the central claim that action-free human/robot video pretraining produces transferable shared dynamics priors is not yet established, because the Stage-1 ablation is confounded, and the headline state-of-the-art claims rest on margins that are often below what can be resolved without reported variance. With cleaner ablations and error bars, this would be a solid contribution.

major comments (4)
  1. [§4.4, Table 5] The 'w/o Stage-1 pretraining' ablation cannot support the paper's central causal claim about shared dynamics priors. Stage 1 adds 100,000 optimization steps, changes the initialization of the VLM/query tokens, and its video mixture contains 50% simulation clips rendered from exactly the three evaluation benchmarks (RoboTwin 2.0, RoboCasa-GR1, and LIBERO, per Section 4.1). The reported drop under this ablation is therefore consistent with additional compute or in-distribution visual pretraining; the contribution of EgoDex human video is never isolated. I request an ablation that matches total compute and initialization while varying only the presence of EgoDex, for example by comparing Stage-1 pretraining on simulation videos alone versus simulation videos plus EgoDex, or by training a 300k-step no-Stage-1 baseline.
  2. [§4.2, Tables 1-3] No variance or number of seeds is reported for any simulation result, despite several headline margins being very small: 0.95 points over ABot-M0 in RoboCasa-GR1 (Table 2), 0.4 points over Fast-WAM in LIBERO (Table 3), and 0.1 points below the best finetuned X-VLA in LIBERO. Without seed-to-seed or rollout variance, the 'state-of-the-art' claim is not statistically supported. Please report means with standard deviations or confidence intervals over at least three training seeds, and state the evaluation rollout counts in the table captions.
  3. [§4.3, Table 4] The real-world comparison reports point estimates from 25 rollouts per cell with no variance, confidence intervals, or protocol details for ensuring identical evaluation conditions across methods. The G1 rows for ACT are 0% on two tasks, and the 16.0-point margin over GR00T-N1.6 would be much more convincing with a per-rollout breakdown and a statement of whether object placements, water amounts, and success criteria were matched across all compared policies.
  4. [§4.5, Table 6] The linear probe demonstrates only that future supervision improves decodability of contact transitions on LIBERO. It does not measure cross-embodiment transfer, nor does it test whether EgoDex human video is the source of the improvement. Since the claimed mechanism is that shared dynamics priors are learned from action-free human and robot video and then transfer across embodiments, the probe should either be run on held-out embodiments or be complemented by an ablation that isolates the EgoDex contribution to downstream control.
minor comments (8)
  1. [Table 5 and §4.4] The row labels 'w/o future prediction' and 'w/o Stage-1 pretraining' are easy to misread. Please state explicitly whether the former removes the future loss only from Stage 2 or from both stages, and whether Stage 1 is still run in that variant.
  2. [§4.1] The text says 'We use the full EgoDex corpus' and also that the video mixture is '50% EgoDex, 20% RoboTwin 2.0, 20% RoboCasa-GR1, and 10% LIBERO.' Please clarify how the 'full corpus' interacts with the 50% mixing ratio, and report the resulting number of EgoDex frames or sequences actually used.
  3. [§4.1] Please state the total scale of the Stage-1 and Stage-2 training data (number of videos, frames, or action sequences), since the compute-versus-data tradeoff is central to interpreting the Stage-1 ablation.
  4. [§3.4] The Beta schedule is introduced only in prose as τ = s(1-u) with u ~ Beta(1.5, 1.0) and s = 0.999; please define s and u explicitly near Eq. (4) and state whether the same schedule is used for the future-generation head in Eq. (2).
  5. [§3.3] The future offset Δ_e is said to match the action horizon H_e. Please clarify how Δ_e is defined for the action-free Stage-1 videos, where there is no action horizon, and whether the prediction target is exactly H_e frames ahead or a different offset.
  6. [Figure 2] Figure 2 is visually dense; labeling the Stage-1 and Stage-2 branches directly in the figure, or splitting it into two panels, would help the reader map the textual description to the architecture.
  7. [Abstract and §4.2] The 59.25% versus 58.3% comparison with ABot-M0 is a 0.95-point margin on a point estimate; please qualify this as within expected noise until variance is reported.
  8. [General] The manuscript does not state whether code, model weights, or evaluation scripts will be released; please add an availability statement describing what will be made public.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the future-prediction and action objectives are independent training signals and the benchmark results are externally evaluated; the Stage-1 ablation confound is an experimental-validity issue, not a circular reduction.

full rationale

The derivation chain is self-contained against external evidence. The future-prediction loss in Eq. (2) supervises query states with a rectified-flow objective on the next frame; the action loss in Eq. (5) supervises native action chunks through the embodiment-routed MoE head. These are separate training signals, and neither is defined in terms of the reported success rates. The benchmarks (LIBERO, RoboCasa-GR1, RoboTwin 2.0) are external evaluation suites; the reported numbers come from rollouts, not from the training objective itself. No fitted parameter is renamed as a prediction, and no equation reduces to another by construction. The linear-probe study in Sec. 4.5 is explicitly conditioned on the current contact state so that echoing the state cannot explain the AUROC/AUPRC gains, making it a genuine representation diagnostic. The related-work self-citations (e.g., Zhong et al. 2025/2026, Song et al. 2025/2026, Li et al. 2026b, Yan et al. 2026, Chen et al. 2026) are contextual examples of existing VLA designs; none is used as a load-bearing premise, and no uniqueness theorem or ansatz is imported from them. The one notable weakness is that the 'w/o Stage-1 pretraining' ablation in Table 5 changes total training steps (200k vs. 300k) and removes in-distribution simulation video from the same benchmarks used for evaluation, so the reported gain cannot be cleanly attributed to transferable dynamics priors. This is a confound in experimental attribution, not a circular reduction: the full and ablated models are not equivalent by construction, and the comparison is made on held-out external rollouts. Therefore no circular step is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central contribution is architectural and empirical, so the ledger consists mostly of training hyperparameters and domain-transfer assumptions rather than invented theoretical entities. The future-supervised query state is a learned representation, not an independent postulate; the paper provides a probe that partially checks its content. The most load-bearing assumptions are the transfer from human video to robots and the sufficiency of the query interface.

free parameters (6)
  • future loss weight lambda_w = 0.05
    Hand-chosen balance between future prediction and action loss in Eq. (6); no sensitivity analysis is reported.
  • number of shared query tokens N = 96
    Set without sweep; determines the capacity of the shared interface in Section 3.2.
  • action horizons H_e (and future offsets Delta_e) = 8 for single-arm, 16 for humanoid, 50 for dual-arm
    Chosen per embodiment to match native action chunking; directly couples future-prediction target length to action horizon.
  • number of MoE experts K = 3
    Matches the three embodiment families; no comparison with other K values.
  • Stage 1 and Stage 2 training steps = 100,000 and 200,000
    Fixed budget without convergence analysis.
  • flow-matching Beta schedule parameters = Beta(1.5, 1.0), s=0.999
    Borrowed from GR00T-N1 for the action head; treated as given rather than tuned for this model.
assumptions (5)
  • domain assumption Future-frame prediction with a SANA generator is a valid surrogate for dynamics priors that transfer to action control.
    The Stage-1 objective and the Q2 ablation rest on this; the linear probe is indirect evidence, not proof.
  • domain assumption Egocentric human videos (EgoDex) share enough dynamics with robot embodiments for pretraining to improve robot control.
    Stage 1 mixes 50% EgoDex with simulation; cross-species transfer is assumed rather than isolated in an experiment.
  • domain assumption A pretrained Qwen3-VL-2B with 96 appended query tokens provides a sufficient shared interface for both future prediction and action generation.
    Section 3.2 assumes last-layer query hidden states carry all information needed by both heads.
  • domain assumption Flow-matching with 4 Euler steps at inference yields actions of sufficient quality.
    The paper reports no sweep over inference steps.
  • standard math Rectified-flow objectives for future frames and actions are standard and correctly applied.
    Equations (2) and (5) use standard flow-matching losses from Lipman et al.; no derivation is needed here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DyPES-VLA: Learning Shared Dynamics Priors and Embodiment-Specific Control for Cross-Embodiment Manipulation." pith.science (2026). https://pith.science/paper/IILDG7KQ

@misc{pith2026260806374,
  author       = {Pith},
  title        = {Pith review of: DyPES-VLA: Learning Shared Dynamics Priors and Embodiment-Specific Control for Cross-Embodiment Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IILDG7KQ}},
  note         = {Machine review of arXiv:2608.06374}
}
read the original abstract

Vision-Language-Action (VLA) models have become a powerful paradigm for robot manipulation, but training a single generalist policy for heterogeneous robot embodiments remains an open problem. Existing methods have two main limitations. First, they underuse dynamics priors shared across diverse visual and interaction data, limiting cross-embodiment transfer. Second, they require extensive manual preprocessing to convert embodiment-specific actions into a common format. To overcome these limitations, we propose DyPES-VLA, a cross-embodiment VLA that learns shared Dynamics Priors and Embodiment-Specific control. First, we learn shared dynamics priors by training the vision-language model (VLM) with a future-prediction objective on cross-embodiment data, driving the shared query representation to capture object motion, contact, and interaction-induced scene changes. Second, an embodiment-specific Mixture-of-Experts (MoE) action head translates these shared dynamics priors into executable controls directly in each embodiment's native action space, without manually pre-aligning heterogeneous actions into a common format. This head shares attention layers to capture common temporal action structures, while its embodiment-specific feed-forward experts resolve the unique kinematic constraints and control semantics of distinct embodiments. As a generalist policy, our \ourmethod achieves state-of-the-art performance across simulation and real-world evaluations, reaching 98.0% success on LIBERO, 59.25% on RoboCasa-GR1, and 89.02% on RoboTwin~2.0.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 canonical work pages

  1. [1]

    Qwen3-VL technical report.arXiv preprint arXiv:2511.21631,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, et al. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631,

    arXiv

  2. [3]

    GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    Johan Bjorck, Fernando Castañeda, 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,

    arXiv

  3. [4]

    π0.5: a vision-language-action model with open-world generalization

    Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, et al. π0.5: a vision-language-action model with open-world generalization. In9th Annual Conference on Robot Learning, 2025a. Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, et al.π0: A vision- language-action ...

    doi:10.15607/rss.2025

  4. [6]

    Worldvla: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539,

    Jun Cen, Chaohui Yu, Hangjie Yuan, et al. Worldvla: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539,

    arXiv

  5. [7]

    Unified diffusion vla: Vision-language-action model via joint discrete denosing diffusion process

    Jiayi Chen, Wenxuan Song, Pengxiang Ding, Ziyang Zhou, Han Zhao, Barrett Tang, Donglin Wang, and Haoang Li. Unified diffusion vla: Vision-language-action model via joint discrete denosing diffusion process. InInternational Conference on Learning Representations, volume 2026, pages 139291–139311,

    work page 2026

  6. [10]

    Gemini robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342,

    Gemini Robotics Team. Gemini robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342,

    arXiv

  7. [11]

    Egodex: Learning dexterous manipulation from large-scale egocentric video.arXiv preprint arXiv:2505.11709,

    Ryan Hoque, Peide Huang, David J Yoon, Mouli Sivapurapu, and Jian Zhang. Egodex: Learning dexterous manipulation from large-scale egocentric video.arXiv preprint arXiv:2505.11709,

    arXiv

  8. [12]

    X-DiffVLA: X-Embodied Diffusion Action Heads for Vision-Language-Action Models

    doi: 10.15607/RSS. 2025.XXI.017. Boyu Li, Chaoyi Xu, Haoqi Yuan, Xinrun Xu, Börje F. Karlsson, Dongbin Zhao, Haoran Li, and Zongqing Lu. X-diffvla: X-embodied diffusion action heads for vision-language-action models.arXiv preprint arXiv:2605.25044, 2026a. Fuhao Li, Wenxuan Song, Han Zhao, Jingbo Wang, Pengxiang Ding, Donglin Wang, Long Zeng, and Haoang Li...

    work page Pith review arXiv 2025

Show all 32 references
  1. [13]

    Genie envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint arXiv:2508.05635,

    Yue Liao, Pengfei Zhou, Siyuan Huang, et al. Genie envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint arXiv:2508.05635,

    arXiv

  2. [14]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    arXiv

  3. [15]

    Rdt-1b: A diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864,

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: A diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864,

    arXiv

  4. [16]

    Rdt2: Exploring the scaling limit of umi data towards zero-shot cross- embodiment generalization.arXiv preprint arXiv:2602.03310,

    Songming Liu, Bangguo Li, Kai Ma, et al. Rdt2: Exploring the scaling limit of umi data towards zero-shot cross- embodiment generalization.arXiv preprint arXiv:2602.03310,

  5. [17]

    Being-h0: Vision-language-action pretraining from large-scale human videos.arXiv preprint arXiv:2507.15597,

    Hao Luo, Yicheng Feng, Wanpeng Zhang, et al. Being-h0: Vision-language-action pretraining from large-scale human videos.arXiv preprint arXiv:2507.15597,

    arXiv

  6. [18]

    F1: A vision-language-action model bridging understanding and generation to actions.arXiv preprint arXiv:2509.06951,

    Qi Lv, Weijie Kong, Hao Li, Jia Zeng, Zherui Qiu, Delin Qu, Haoming Song, Qizhi Chen, Xiang Deng, and Jiangmiao Pang. F1: A vision-language-action model bridging understanding and generation to actions.arXiv preprint arXiv:2509.06951,

    arXiv

  7. [19]

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

    arXiv

  8. [20]

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

    Teli Ma, Jia Zheng, Zifan Wang, Chunli Jiang, Andy Cui, Junwei Liang, and Shuo Yang. Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control.arXiv preprint arXiv:2603.10448,

  9. [21]

    Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213,

    Octo Model Team, Dibya Ghosh, Homer Walke, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213,

    arXiv

  10. [22]

    Open x-embodiment: Robotic learning datasets and rt-x models.arXiv preprint arXiv:2310.08864,

    Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, et al. Open x-embodiment: Robotic learning datasets and rt-x models.arXiv preprint arXiv:2310.08864,

    arXiv

  11. [23]

    Qwen Team

    doi: 10.15607/RSS.2025.XXI.012. Qwen Team. Qwen-vla: Unifying vision-language-action modeling across tasks, environments, and robot embodiments. arXiv preprint arXiv:2605.30280,

    2025 arXiv

  12. [24]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629,

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629,

    arXiv

  13. [25]

    S-VAM: Shortcut video-action model by self-distilling geometric and semantic foresight.arXiv preprint arXiv:2603.16195,

    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:2603.16195,

  14. [26]

    Egovla: Learning vision-language-action models from egocentric human videos.arXiv preprint arXiv:2507.12440,

    Ruihan Yang, Qinxi Yu, Yecheng Wu, et al. Egovla: Learning vision-language-action models from egocentric human videos.arXiv preprint arXiv:2507.12440,

    arXiv

  15. [27]

    Abot-m0: Vla foundation model for robotic manipulation with action manifold learning.arXiv preprint arXiv:2602.11236, 2026a

    Yandan Yang, Shuang Zeng, Tong Lin, et al. Abot-m0: Vla foundation model for robotic manipulation with action manifold learning.arXiv preprint arXiv:2602.11236, 2026a. Yi Yang, Xueqi Li, Yiyang Chen, et al. Mantis: A versatile vision-language-action model with disentangled vis...

    arXiv

  16. [28]

    Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666, 2026.https://arxiv.org/abs/2603.16666

    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.https://arxiv.org/abs/2603.16666. Wenyao Zhang, Hongsi Liu, Zekun Qi, Yunnan Wang, XinQiang Yu, Jiazhao Zhang, Runpei...

    2026 arXiv

  17. [29]

    Universal actions for enhanced embodied foundation models

    Jinliang Zheng, Jianxiong Li, Dongxiu Liu, et al. Universal actions for enhanced embodied foundation models. In CVPR, 2025a. Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-VLA: Soft-promp...

    arXiv

  18. [30]

    Dualcot-vla: Visual-linguistic chain of thought via parallel reasoning for vision-language-action models.arXiv preprint arXiv:2603.22280,

    Zhide Zhong, Junfeng Li, Junjie He, et al. Dualcot-vla: Visual-linguistic chain of thought via parallel reasoning for vision-language-action models.arXiv preprint arXiv:2603.22280,

  19. [31]

    Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets.arXiv preprint arXiv:2504.02792,

    Chuning Zhu, Raymond Yu, Siyuan Feng, Benjamin Burchfiel, Paarth Shah, and Abhishek Gupta. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets.arXiv preprint arXiv:2504.02792,

    arXiv

  20. [32]

    Pour water from the cup into the bowl

    14 Supplementary Material A Per-Task Results on RoboCasa-GR1 Table S1 reports the per-task success rates (SR) of our single co-trained checkpoint on all 24 RoboCasa- GR1 Nasiriany et al. (2024); Bjorck et al. (2025) tabletop tasks, evaluated with 50 rollouts per task (1,200 ro...

    2024

  21. [2023]

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

    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,

    arXiv

  22. [2024]

    Himoe-vla: Hierarchical mixture-of-experts for generalist vision-language-action policies

    Zhiying Du, Bei Liu, Yaobo Liang, Yichao Shen, Haidong Cao, Xiangyu Zheng, Zhiyuan Feng, Zuxuan Wu, Jiaolong Yang, and Yu-Gang Jiang. Himoe-vla: Hierarchical mixture-of-experts for generalist vision-language-action policies. arXiv preprint arXiv:2512.05693,

    arXiv

  23. [2025]

    Motus: A unified latent action world model.arXiv preprint arXiv:2512.13030,

    Hongzhe Bi, Hengkai Tan, Shenghao Xie, et al. Motus: A unified latent action world model.arXiv preprint arXiv:2512.13030,

    arXiv

  24. [2026]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprint arXiv:2506.18088, 2025a

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Qiwei Liang, Zixuan Li, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprint ar...

    arXiv

Pith tools

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