Pith. sign in

REVIEW 2 major objections 5 minor 68 references

Robot policies can train on future-video prediction and then drop video generation at runtime, reaching 85 ms action-only inference on a local GPU while keeping or improving real-robot success rates.

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

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

T0 review · deepseek-v4-flash

2026-08-02 03:11 UTC pith:PJGPCP5P

load-bearing objection A useful incremental technical report on efficient WAM inference, but the 'stronger' claim rests on experiments that do not isolate the proposed mechanism. the 2 major comments →

arxiv 2607.13960 v3 pith:PJGPCP5P submitted 2026-07-15 cs.RO

GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

GigaWorld Team , Angen Ye , Angyuan Ma , Boyuan Wang , Chaojun Ni , Fangzheng Ye , Guan Huang , Guo Li
show 21 more authors
Guosheng Zhao Haodong Yan Hengtao Li Jiwen Lu Kai Wang Mingming Yu Qitang Hu Qiuping Deng Songling Liu Xiaoyu Tian Xiaofeng Wang Xinyu Zhou Xiuwei Xu Xinze Chen Yang Wang Yejun Zeng Yifan Chang Yun Ye Zhenyu Wu Zhanqian Wu Zheng Zhu
This is my paper
classification cs.RO
keywords World Action Modelaction-centered policyMixture-of-Transformersfuture visual dynamicsinference-time video generationrobot manipulationflow matchingreal-time closed-loop control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

GigaWorld-Policy-0.5 makes a specific bet: future visual dynamics are a training-time teacher, not a runtime partner. The model predicts actions directly from the current observation, state, and instruction, while future scene tokens are predicted only during training and conditioned on those actions. This action-centered causal design lets the policy skip the expensive video-generation branch entirely at deployment. A Mixture-of-Transformers splits the network into a heavy visual expert and a slim action expert, so the action-only pathway runs at 85 ms per action on an RTX 4090. The paper reports real-robot success rates of 0.85 on fruit picking, 0.89 on object placement, and 0.80 on long-horizon tasks, each above the compared baselines, and argues that the future-dynamics supervision is what transfers into better policies.

Core claim

The central claim is that coupling action prediction with future visual prediction during training yields better robot policies, and that this coupling can be completely severed at inference time without losing the benefit. The model uses an action-centered causal mask: action tokens attend to the current context but never to future visual tokens, while future visual tokens attend to the actions. That asymmetry makes the future-video branch optional at deployment. The paper demonstrates this with a Mixture-of-Transformers where a visual expert (hidden size 3072) and a much smaller action expert (hidden size 1024) are trained jointly with flow matching; at runtime only the action expert's pat

What carries the argument

The load-bearing piece is the action-centered causal mask combined with a Mixture-of-Transformers. The mask enforces that action tokens can only attend to current observations, state, and language, never to future visual tokens; future visual tokens may attend to the actions. This is what makes future-video prediction a train-only auxiliary branch that can be dropped at inference. The Mixture-of-Transformers then separates the network into a visual expert (initialized from a large pretrained world model and carrying a general scene-dynamics prior) and a smaller action expert; because only the action expert needs to run during deployment, the cost of the visual expert is paid once at training

Load-bearing premise

The 'stronger' claim depends on success rates measured from only 10 real-robot trials per condition, reported without confidence intervals or significance tests; the gap between 0.60 and 0.80 on long-horizon tasks is within binomial sampling noise at that sample size.

What would settle it

Run the real-robot comparisons with at least 100 trials per condition and check whether the success-rate gaps over baselines persist; if they shrink to zero, the 85 ms latency result stands but the comparative 'stronger' claim fails.

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

If this is right

  • World-action-model training becomes compatible with real-time closed-loop control: the dense supervision of future visual dynamics can be kept without paying video-generation latency at runtime.
  • The action-expert/visual-expert split lets a policy scale its world-model capacity (and param count) without proportionally increasing the latency of the deployed action-only pathway.
  • The mixed AC-WM and WAM pretraining recipe, as ablated in the paper, reaches higher success and converges earlier than WAM pretraining alone, pointing to a concrete training recipe for downstream policy learning.
  • The reported long-horizon gains (0.80 vs 0.60 for the strongest baseline) suggest future-dynamics supervision is particularly valuable for multi-step manipulation, where intermediate state transitions must be maintained.

Where Pith is reading between the lines

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

  • Editorial extension: the causal-mask design, not the specific MoT, is the transferable idea; a reader would expect the same train-only-future trick to work with other generative backbones or with non-pixel auxiliary predictions (e.g., latent subgoals), as long as the mask prevents leakage into actions.
  • Editorial extension: the paper's ablation isolates AC-WM+WAM vs WAM pretraining but does not compare against a policy trained with no future-visual objective at all; a direct test of whether the gains come specifically from future-dynamics supervision, rather than from the strong world-model initialization, would pin down the mechanism.
  • Editorial extension: the 85 ms figure is for a specific RTX 4090 with a C++ runtime and KV caching; a natural follow-up is measuring action-only latency on weaker edge hardware and across longer action horizons, where autoregressive decoding may become the bottleneck.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes GigaWorld-Policy-0.5, an action-centered World Action Model (WAM) built on GigaWorld-Policy. It introduces a Mixture-of-Transformers (MoT) architecture with a lightweight action expert and a large visual expert, a mixed Action-Conditioned World Modeling (AC-WM) and WAM pretraining stage, and an agent-based AutoResearch pipeline for hyperparameter search. At inference, the model uses action-only decoding, omitting future-video generation, and achieves 85 ms latency on an RTX 4090 with a C++ runtime. The paper reports real-robot success rates on fruit picking (0.85 average), object placement (0.89), and long-horizon tasks (0.80), claiming improvements over VLA and WAM baselines. The flow-matching formulation (Eqs. 4–8) is standard and self-consistent.

Significance. If the central claims hold, the paper makes a useful engineering contribution by showing that WAMs can be deployed with low latency while retaining world-model supervision. The latency measurements are concrete and the architectural separation of experts is a plausible design. However, the 'stronger' claim is currently not well supported: no experiment ablates the future-visual loss, and all success-rate comparisons rest on 10-trial averages without error bars or significance tests. The paper also does not release code or checkpoints, limiting reproducibility. The AutoResearch pipeline is a nice tool but is not the main novelty.

major comments (2)
  1. [§4.2, Fig. 5; Eq. (8)] The abstract's claim that GigaWorld-Policy-0.5 'preserves the training benefits of future visual dynamics' is not directly tested. The only training ablation (Fig. 5) contrasts mixed AC-WM+WAM pretraining with WAM-only pretraining; both include the future-visual flow-matching loss in Eq. (8). There is no condition trained with action supervision alone or without the visual loss. The success-rate gains (e.g., Table 1: 0.85 vs 0.80) could therefore be driven entirely by the GigaWorld-1 visual-expert initialization (Sec. 3.2) or the AC-WM stage, with the post-training future-visual objective inert. To support the central claim, add an ablation that removes the future-visual loss during post-training while keeping all other components fixed.
  2. [§4.1, Tables 1–3] Success rates are averages over 10 real-robot trials per condition, with no confidence intervals or significance tests. This is statistically insufficient for the headline 'stronger' claim. In Table 3, the 0.20 absolute gain (0.60 vs 0.80) for long-horizon tasks corresponds to, for example, 6/10 versus 8/10 successes; with n=10 this difference is within binomial sampling noise and not significant at conventional levels. The per-task gains in Tables 1–2 face the same issue. The authors should report error bars, per-trial breakdowns, or a significance test, and ideally increase the number of trials for the key comparisons, before claiming superiority over baselines.
minor comments (5)
  1. [§3.2] When initializing the action expert by taking the leading n dimensions of the visual-expert weights, clarify how the remaining (unmatched) parameters are initialized: zero, random, or otherwise.
  2. [Table 4] The row 'GigaWorld-Policy-0.5 189 110 0.85; w/ C++ deployment 140 85 0.85' is confusing. Split into two rows, one for the torch.compile setting and one for the C++ runtime, and state the success rate corresponding to each configuration.
  3. [Figure 2] The 'Optional Future' path in the inference panel is ambiguous. Please indicate clearly which expert branches and token sequences are active during action-only decoding.
  4. [§4.2, Fig. 5] Clarify whether the success-rate curves in Fig. 5 come from real-robot trials or a validation set. If real-robot, the small number of trials per checkpoint makes these curves noisy; consider reporting confidence bands or smoothing.
  5. [§4.2] The batch-size sweep is described only in text; no table or figure is provided. Report the tested batch sizes and their validation metrics for transparency.

Circularity Check

0 steps flagged

No significant circularity; the paper is an empirical incremental report. Main weakness is a missing no-visual-loss ablation, not a circular derivation.

full rationale

The paper is an incremental technical report that inherits the action-centered WAM formulation from the authors' prior GigaWorld-Policy [55] and measures real-robot success rates and latencies. I checked each alleged reduction. (1) The loss Eq. (8) is a standard flow-matching regression; no output is defined in terms of the success rates it later reports. (2) The AutoResearch section tunes LR/batch size and selects a 30K checkpoint on a validation split of the pick-the-fruit task, then Table 1 reports real-robot SR for that same task. This is model selection on a related validation set, not a fitted parameter that by construction equals the reported SR; the SR is measured on 10 live trials, not computed from validation action MSE. (3) The MoT inference latency numbers are direct measurements. (4) Self-citations to [55] and [39] are present and load-bearing as prior building blocks, but they are not used to forbid alternatives or to define the reported quantities; external baselines (Motus, FastWAM, pi0.5) are also compared. The closest thing to a scientific defect is that no ablation removes the future-visual objective, so the claim that future visual dynamics 'preserve training benefits' is under-supported; that is a missing-control/correctness issue, not a circular reduction. No step makes X derive Y where X is defined by Y or where a fitted quantity is renamed as a prediction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical entities or forces are introduced; the 'experts' are architectural modules, and AutoResearch is an external tool. The paper's novelty is an engineering combination, so the ledger is dominated by hand-chosen and AutoResearch-selected hyperparameters plus domain assumptions about transfer and evaluation.

free parameters (6)
  • Action/visual expert dimensions (action 1024/4096, visual 3072/14336) = 1024/4096; 3072/14336
    Manual architecture choices that trade capacity vs action-only inference latency; not swept in reported ablations.
  • Flow-shift factors γa, γv = not reported
    Eq. 4 introduces modality-specific timestep shifts; their values affect the training objective but are not given.
  • Action chunk length p and future-observation stride Δ = not reported ('settings adopted in GigaWorld-Policy')
    Determine horizon and number of future frames K=⌊p/Δ⌋; central to action/future supervision but not specified here.
  • Learning rate = 6×10⁻⁵
    Selected by AutoResearch 1K-step pilot sweep on pick-the-fruit validation (Table 5) and used for final model.
  • Batch size = 16
    Retained after AutoResearch batch-size sweep because alternatives did not outperform the original setting (§4.2).
  • Number of training steps / checkpoint = 30K steps
    Checkpoint chosen as best validation action MSE on pick-the-fruit (Fig. 6), then used for real-robot evaluation.
axioms (6)
  • standard math Rectified-flow interpolation between Gaussian noise and latent action/visual data is a valid training objective.
    Background from Lipman et al. [23] and Esser et al. [13]; used without proof in §3.1.
  • domain assumption Action-centered causal mask (future visual tokens may attend to actions; action tokens may not attend to future visuals) prevents leakage and provides beneficial dense supervision.
    §3.1–3.2 states this; no experiment varies the mask to test leakage/stability.
  • domain assumption Initialization of the visual expert from GigaWorld-1 transfers a useful world-model prior to robot data.
    §3.2–3.3 relies on this; no ablation replaces or removes the initialization.
  • domain assumption Mixed AC-WM + WAM pretraining on 2K hours of filtered data transfers better action representations than WAM-only pretraining.
    Tested in one ablation on pick-the-fruit (Fig. 5); the effect on other tasks is assumed.
  • domain assumption Ten real-robot trials per condition are sufficient to rank policies.
    §4.1 averages 10 trials and reports no confidence intervals or significance tests.
  • domain assumption Flow-matching velocity regression on joint action and future visual tokens induces physically grounded action generation.
    Eq. 8 is the core objective; its link to downstream control success is assumed from prior WAM work.

pith-pipeline@v1.3.0-alltime-deepseek · 15488 in / 17364 out tokens · 151330 ms · 2026-08-02T03:11:49.701628+00:00 · methodology

0 comments
read the original abstract

World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead and hindering real-time closed-loop deployment. GigaWorld-Policy addresses this issue with an action-centered formulation, where future visual dynamics are used during training while action-only decoding is used at inference time. Building upon this framework, we present GigaWorld-Policy-0.5, an enhanced action-centered WAM designed for more efficient robot control. During pretraining, GigaWorld-Policy-0.5 adopts a mixed Action-Conditioned World Modeling (AC-WM) and WAM training strategy. This strengthens the coupling between visual dynamics and robot actions and improves the transferability of action representations for downstream policy learning. For efficient inference, GigaWorld-Policy-0.5 introduces a Mixture-of-Transformers architecture that separates visual dynamics modeling and action generation into specialized experts, reducing active computation during action-only inference and achieving 85 ms inference latency on a local RTX 4090 setup. In addition, we employ an agent-based AutoResearch pipeline to systematically search training configurations, enabling more efficient identification of optimal experimental setups while reducing the time and manual intervention required for hyperparameter tuning. Experiments and ablations show that GigaWorld-Policy-0.5 preserves the training benefits of future visual dynamics while improving inference efficiency for robot control.

Figures

Figures reproduced from arXiv: 2607.13960 by Angen Ye, Angyuan Ma, Boyuan Wang, Chaojun Ni, Fangzheng Ye, GigaWorld Team, Guan Huang, Guo Li, Guosheng Zhao, Haodong Yan, Hengtao Li, Jiwen Lu, Kai Wang, Mingming Yu, Qitang Hu, Qiuping Deng, Songling Liu, Xiaofeng Wang, Xiaoyu Tian, Xinyu Zhou, Xinze Chen, Xiuwei Xu, Yang Wang, Yejun Zeng, Yifan Chang, Yun Ye, Zhanqian Wu, Zheng Zhu, Zhenyu Wu.

Figure 1
Figure 1. Figure 1: Comparison of GigaWorld-Policy-0.5 with baselines on inference frequency and success rate across hardware platforms in real-world settings. World models have recently made significant progress in learning predictive representations of the physical world from large-scale visual data [1, 2, 20, 24, 38, 41, 45, 46, 62, 68]. By modeling how scenes evolve over time, these models capture rich spatiotemporal prio… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GigaWorld-Policy-0.5, an MoT-based action-centered World Action Model. The model consists of a visual expert and an action expert: the visual expert specializes in processing video tokens, while the action expert focuses on action-token modeling. The two experts are connected through multi-modal self￾attention, which follows the same causal masking strategy as GigaWorld-Policy [55] to preserve … view at source ↗
Figure 3
Figure 3. Figure 3: Real-world demonstration of GigaWorld-Policy-0.5 on the Tableware Arrangement task. relationship between actions and visual dynamics, we also incorporate action-conditioned world-model training during this stage, where future visual evolution is predicted under robot-relevant observations and actions. This encourages the model to learn how robot actions affect scene changes before target-domain specializat… view at source ↗
Figure 4
Figure 4. Figure 4: Real-world demonstration of GigaWorld-Policy-0.5 on the Food Heating task. follows the same binary definition as in simulation, where SR = 1 only if the full task sequence is completed and SR = 0 otherwise. Baselines. We compare GigaWorld-Policy-0.5 with several representative baselines spanning two major paradigms. For VLM-based VLA methods, we include 𝜋0.5 [15], which uses a large vision-language back￾bo… view at source ↗
Figure 5
Figure 5. Figure 5: Success rates at different training steps in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: AutoResearch hyperparameter search and training progression on the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

68 extracted references · 32 linked inside Pith

  1. [1]

    Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575, 2025

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575, 2025. 1

    Pith/arXiv arXiv 2025

  2. [2]

    Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800, 2026

    Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800, 2026. 1

    Pith/arXiv arXiv 2026

  3. [3]

    Cosmos-transfer1: Conditional world generation with adaptive multimodal control.arXiv preprint arXiv:2503.14492, 2025

    Hassan Abu Alhaija, Jose Alvarez, Maciej Bala, Tiffany Cai, Tianshi Cao, Liz Cha, Joshua Chen, Mike Chen, Francesco Ferroni, Sanja Fidler, et al. Cosmos-transfer1: Conditional world generation with adaptive multimodal control.arXiv preprint arXiv:2503.14492, 2025. 3

    Pith/arXiv arXiv 2025

  4. [4]

    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, et al. Motus: A unified latent action world model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 35101–35113, 2026. 2, 3, 5, 8, 9, 10

    2026

  5. [5]

    Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    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, 2025. 2

    Pith/arXiv arXiv 2025

  6. [6]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.𝜋0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024. 2

    Pith/arXiv arXiv 2024

  7. [7]

    Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025

    Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025. 6

    Pith/arXiv arXiv 2025

  8. [8]

    Worldvla: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539, 2025

    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. 1

    Pith/arXiv arXiv 2025

  9. [9]

    Lawam: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026

    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. 4

    arXiv 2026

  10. [10]

    Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining

    Hyung Won Chung, Xavier Garcia, Adam Roberts, Yi Tay, Orhan Firat, Sharan Narang, and Noah Constant. Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining. InThe Eleventh International Conference on Learning Representations, 2023. 6

    2023

  11. [11]

    Emma: Generalizing real-world robot manipulation via generative visual transfer.arXiv preprint arXiv:2509.22407, 2025

    Zhehao Dong, Xiaofeng Wang, Zheng Zhu, Yirui Wang, Yang Wang, Yukun Zhou, Boyuan Wang, Chaojun Ni, Runqi Ouyang, Wenkang Qin, et al. Emma: Generalizing real-world robot manipulation via generative visual transfer.arXiv preprint arXiv:2509.22407, 2025. 3

    arXiv 2025

  12. [12]

    Learning universal policies via text-guided video generation.Advances in neural information processing systems, 36:9156–9172, 2023

    Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation.Advances in neural information processing systems, 36:9156–9172, 2023. 3

    2023

  13. [13]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024. 5 13 GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by...

    2024

  14. [14]

    arXiv preprint arXiv:2511.14759, 2025

    Physical Intelligence, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Kevin Black, Ken Conley, Grace Connors, James Darpinian, Karan Dhabalia, Jared DiCarlo, et al.𝜋0.6: a vla that learns from experience. arXiv preprint arXiv:2511.14759, 2025. 2

    Pith/arXiv arXiv 2025

  15. [15]

    2, 8, 9, 10

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.𝜋0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025. 2, 8, 9, 10

    Pith/arXiv arXiv 2025

  16. [16]

    Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, Thomas Charbonnier, et al.𝜋0.7: a steerable generalist robotic foundation model with emergent capabilities.arXiv preprint arXiv:2604.15483, 2026. 1

    Pith/arXiv arXiv 2026

  17. [17]

    Wovr: World models as reliable simulators for post-training vla policies with rl.arXiv preprint arXiv:2602.13977, 2026

    Zhennan Jiang, Shangqing Zhou, Yutong Jiang, Zefang Huang, Mingjie Wei, Yuhui Chen, Tianxing Zhou, Zhen Guo, Hao Lin, Quanlu Zhang, et al. Wovr: World models as reliable simulators for post-training vla policies with rl.arXiv preprint arXiv:2602.13977, 2026. 3

    Pith/arXiv arXiv 2026

  18. [18]

    autoresearch, 2026

    Andrej Karpathy. autoresearch, 2026. URLhttps://github.com/karpathy/autoresearch. 3, 9, 10

    2026

  19. [19]

    Freeaction: Training-free techniques for enhanced fidelity of trajectory-to-video generation.arXiv preprint arXiv:2509.24241, 2025

    Seungwook Kim, Seunghyeon Lee, and Minsu Cho. Freeaction: Training-free techniques for enhanced fidelity of trajectory-to-video generation.arXiv preprint arXiv:2509.24241, 2025. 3

    arXiv 2025

  20. [20]

    A path towards autonomous machine intelligence version 0.9

    Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27.Open Review,

    2022

  21. [21]

    Mimicdreamer: Aligning human and robot demonstrations for scalable vla training.arXiv preprint arXiv:2509.22199, 2025

    Haoyun Li, Ivan Zhang, Runqi Ouyang, Xiaofeng Wang, Zheng Zhu, Zhiqin Yang, Zhentao Zhang, Boyuan Wang, Chaojun Ni, Wenkang Qin, et al. Mimicdreamer: Aligning human and robot demonstrations for scalable vla training.arXiv preprint arXiv:2509.22199, 2025. 3

    arXiv 2025

  22. [22]

    Causal world modeling for robot control.arXiv preprint arXiv:2601.21998, 2026

    Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control.arXiv preprint arXiv:2601.21998, 2026. 2, 3, 4

    Pith/arXiv arXiv 2026

  23. [23]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022. 5

    Pith/arXiv arXiv 2022

  24. [24]

    Timestep embedding tells: It’s time to cache for video diffusion model

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7353–7363, 2025. 1

    2025

  25. [25]

    Robotransfer: Controllable geometry-consistent video diffusion for manipulation policy transfer

    Liu Liu, Xiaofeng Wang, Guosheng Zhao, Keyu Li, Wenkang Qin, Jiagang Zhu, Jiaxiong Qiu, Guan Huang, and Zhizhong Su. Robotransfer: Controllable geometry-consistent video diffusion for manipulation policy transfer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1410–1420, 2026. 3

    2026

  26. [26]

    Rdt-1b: a diffusion foundation model for bimanual manipulation

    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. InInternational Conference on Learning Representations, volume 2025, pages 29982–30009, 2025. 6

    2025

  27. [27]

    Being-h0

    Hao Luo, Wanpeng Zhang, Yicheng Feng, Sipeng Zheng, Haiweng Xu, Chaoyi Xu, Ziheng Xi, Yuhui Fu, and Zongqing Lu. Being-h0. 7: A latent world-action model from egocentric videos.arXiv preprint arXiv:2605.00078, 2026. 1

    Pith/arXiv arXiv 2026

  28. [28]

    Viva: A video-generative value model for robot reinforcement learning.arXiv preprint arXiv:2604.08168, 2026

    Jindi Lv, Hao Li, Jie Li, Fankun Kong, Yang Wang, Pengfei Yi, Yifei Nie, Xiaofeng Wang, Zheng Zhu, Chaojun Ni, et al. Viva: A video-generative value model for robot reinforcement learning.arXiv preprint arXiv:2604.08168, 2026. 2 14 GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

    Pith/arXiv arXiv 2026

  29. [29]

    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,

    arXiv

  30. [30]

    Recondreamer-rl: Enhancing reinforcement learning via diffusion-based scene reconstruction.arXiv preprint arXiv:2508.08170, 2025

    Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Xinze Chen, Guanghong Jia, Guan Huang, and Wenjun Mei. Recondreamer-rl: Enhancing reinforcement learning via diffusion-based scene reconstruction.arXiv preprint arXiv:2508.08170, 2025. 3

    Pith/arXiv arXiv 2025

  31. [31]

    Swiftvla: Unlocking spatiotemporal dynamics for lightweight vla models at minimal overhead

    Chaojun Ni, Cheng Chen, Xiaofeng Wang, Zheng Zhu, Wenzhao Zheng, Boyuan Wang, Tianrun Chen, Guosheng Zhao, Haoyun Li, Zhehao Dong, et al. Swiftvla: Unlocking spatiotemporal dynamics for lightweight vla models at minimal overhead. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13474–13485, 2026. 2, 3

    2026

  32. [32]

    mimic- video: Video-action models for generalizable robot control beyond vlas.arXiv preprint arXiv:2512.15692,

    Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava. mimic- video: Video-action models for generalizable robot control beyond vlas.arXiv preprint arXiv:2512.15692,

    Pith/arXiv arXiv

  33. [33]

    Videovla: Video generators can be generalizable robot manipulators.Advances in neural information processing systems, 38:95597–95621, 2026

    Yichao Shen, Fangyun Wei, Zhiying Du, Yaobo Liang, Yan Lu, Jiaolong Yang, Nanning Zheng, and Baining Guo. Videovla: Video generators can be generalizable robot manipulators.Advances in neural information processing systems, 38:95597–95621, 2026. 2, 3

    2026

  34. [34]

    World guidance: World modeling in condition space for action generation.arXiv preprint arXiv:2602.22010, 2026

    Yue Su, Sijin Chen, Haixin Shi, Mingyu Liu, Zhengshen Zhang, Ningyuan Huang, Weiheng Zhong, Zhengbang Zhu, Yuxiao Liu, and Xihui Liu. World guidance: World modeling in condition space for action generation.arXiv preprint arXiv:2602.22010, 2026. 1

    arXiv 2026

  35. [35]

    Anypos: Automated task-agnostic actions for bimanual manipulation.arXiv preprint arXiv:2507.12768,

    Hengkai Tan, Yao Feng, Xinyi Mao, Shuhe Huang, Guodong Liu, Zhongkai Hao, Hang Su, and Jun Zhu. Anypos: Automated task-agnostic actions for bimanual manipulation.arXiv preprint arXiv:2507.12768,

    Pith/arXiv arXiv

  36. [36]

    Gigabrain-0: A world model-powered vision-language-action model.arXiv preprint arXiv:2510.19430, 2025

    GigaBrain Team, Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Haoyun Li, Jie Li, Jiagang Zhu, Lv Feng, et al. Gigabrain-0: A world model-powered vision-language-action model.arXiv preprint arXiv:2510.19430, 2025. 1, 2

    arXiv 2025

  37. [37]

    Gigabrain-0.5 m*: a vla that learns from world model-based reinforcement learning

    GigaBrain Team, Boyuan Wang, Bohan Li, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigabrain-0.5 m*: a vla that learns from world model-based reinforcement learning. arXiv preprint arXiv:2602.12099, 2026. 2

    arXiv 2026

  38. [38]

    Gigaworld-0: World models as data engine to empower embodied ai

    GigaWorld Team, Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Haoyun Li, Jiagang Zhu, Kerui Li, Mengyuan Xu, et al. Gigaworld-0: World models as data engine to empower embodied ai. arXiv preprint arXiv:2511.19861, 2025. 1, 3

    arXiv 2025

  39. [39]

    Gigaworld-1: A roadmap to build world models for robot policy evaluation

    GigaWorld Team, Angyuan Ma, Boyuan Wang, Bohan Li, Chaojun Ni, Guo Li, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, et al. Gigaworld-1: A roadmap to build world models for robot policy evaluation. arXiv preprint arXiv:2607.02642, 2026. 6, 8

    Pith/arXiv arXiv 2026

  40. [40]

    Motubrain: An advanced world action model for robot control.arXiv preprint arXiv:2604.27792, 2026

    MotuBrain Team, Chendong Xiang, Fan Bao, Haitian Liu, Hengkai Tan, Hongzhe Bi, James Li, Jiabao Liu, Jingrui Pang, Kiro Jing, et al. Motubrain: An advanced world action model for robot control.arXiv preprint arXiv:2604.27792, 2026. 2, 4

    Pith/arXiv arXiv 2026

  41. [41]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 1, 6

    Pith/arXiv arXiv 2025

  42. [42]

    Embodiedreamer: Advancing real2sim2real transfer for policy training via embodied world modeling.arXiv preprint arXiv:2507.05198, 2025

    Boyuan Wang, Xinpan Meng, Xiaofeng Wang, Zheng Zhu, Angen Ye, Yang Wang, Zhiqin Yang, Chaojun 15 GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch Ni, Guan Huang, and Xingang Wang. Embodiedreamer: Advancing real2sim2real transfer for policy training via embodied world modeling.arXiv preprint arXiv:2507.05198, 2025. 3

    Pith/arXiv arXiv 2025

  43. [43]

    Humandreamer-x: Photorealistic single-image human avatars reconstruction via gaussian restoration.arXiv preprint arXiv:2504.03536, 2025

    Boyuan Wang, Runqi Ouyang, Xiaofeng Wang, Zheng Zhu, Guosheng Zhao, Chaojun Ni, Xiaopei Zhang, Guan Huang, Yijie Ren, Lihong Liu, et al. Humandreamer-x: Photorealistic single-image human avatars reconstruction via gaussian restoration.arXiv preprint arXiv:2504.03536, 2025. 3

    arXiv 2025

  44. [44]

    Reconphys: Reconstruct appearance and physical attributes from single video.arXiv preprint arXiv:2604.07882, 2026

    Boyuan Wang, Xiaofeng Wang, Yongkang Li, Zheng Zhu, Yifan Chang, Angen Ye, Guosheng Zhao, Chaojun Ni, Guan Huang, Yijie Ren, et al. Reconphys: Reconstruct appearance and physical attributes from single video.arXiv preprint arXiv:2604.07882, 2026. 3

    Pith/arXiv arXiv 2026

  45. [45]

    Drivedreamer: Towards real-world-drive world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. InEuropean conference on computer vision, pages 55–72. Springer, 2024. 1

    2024

  46. [46]

    Worlddreamer: Towards general world models for video generation via predicting masked tokens.arXiv preprint arXiv:2401.09985, 2024

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens.arXiv preprint arXiv:2401.09985, 2024. 1

    Pith/arXiv arXiv 2024

  47. [47]

    Egovid-5m: A large-scale video-action dataset for egocentric videos generation.Advances in Neural Information Processing Systems, 38, 2026

    Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, and Yingya Zhang. Egovid-5m: A large-scale video-action dataset for egocentric videos generation.Advances in Neural Information Processing Systems, 38, 2026. 3

    2026

  48. [48]

    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, volume 2024, pages 10641–10662, 2024. 3

    2024

  49. [49]

    Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation.arXiv preprint arXiv:2412.13877, 2024

    Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation.arXiv preprint arXiv:2412.13877, 2024. 6

    Pith/arXiv arXiv 2024

  50. [50]

    A pragmatic vla foundation model.arXiv preprint arXiv:2601.18692, 2026

    Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, Fangjing Wang, Qian Zhu, He Sun, Yong Wang, Shuailei Ma, et al. A pragmatic vla foundation model.arXiv preprint arXiv:2601.18692, 2026. 2

    Pith/arXiv arXiv 2026

  51. [51]

    Pandora: Towards general world model with natural language actions and video states.arXiv preprint arXiv:2406.09455, 2024

    Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states.arXiv preprint arXiv:2406.09455, 2024. 3

    Pith/arXiv arXiv 2024

  52. [52]

    S-vam: Shortcut video-action model by self-distilling geometric and semantic foresight.arXiv preprint arXiv:2603.16195, 2026

    Haodong Yan, Zhide Zhong, Jiaguan Zhu, Junjie He, Weilin Yuan, Wenxuan Song, Xin Gong, Yingjie Cai, Guanyi Zhao, Xu Yan, et al. S-vam: Shortcut video-action model by self-distilling geometric and semantic foresight.arXiv preprint arXiv:2603.16195, 2026. 4

    arXiv 2026

  53. [53]

    Vla-r1: Enhancing reasoning in vision-language-action models.arXiv preprint arXiv:2510.01623, 2025

    Angen Ye, Zeyu Zhang, Boyuan Wang, Xiaofeng Wang, Dapeng Zhang, and Zheng Zhu. Vla-r1: Enhancing reasoning in vision-language-action models.arXiv preprint arXiv:2510.01623, 2025. 2

    arXiv 2025

  54. [54]

    Halo-wa: Hybrid-attention latent-guided online reinforcement learning for world-action models.arXiv preprint arXiv:2607.04265, 2026

    Angen Ye, Weijie Ke, Xiaofeng Wang, Xinze Chen, Chaojun Ni, Guosheng Zhao, Boyuan Wang, Zheng Zhu, Junjie Xie, and Dapeng Zhang. Halo-wa: Hybrid-attention latent-guided online reinforcement learning for world-action models.arXiv preprint arXiv:2607.04265, 2026. 2

    Pith/arXiv arXiv 2026

  55. [55]

    Gigaworld-policy: An efficient action-centered world–action model.arXiv preprint arXiv:2603.17240, 2026

    Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigaworld-policy: An efficient action-centered world–action model.arXiv preprint arXiv:2603.17240, 2026. 2, 3, 4, 5, 6, 8, 9, 10

    arXiv 2026

  56. [56]

    World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, 2026

    Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, 2026. 2 16 GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

    Pith/arXiv arXiv 2026

  57. [57]

    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. 2, 4, 8, 9, 10

    Pith/arXiv arXiv 2026

  58. [58]

    Igniting vlms toward the embodied space.arXiv preprint arXiv:2509.11766, 2025

    Andy Zhai, Brae Liu, Bruno Fang, Chalse Cai, Ellie Ma, Ethan Yin, Hao Wang, Hugo Zhou, James Wang, Lights Shi, et al. Igniting vlms toward the embodied space.arXiv preprint arXiv:2509.11766, 2025. 2

    arXiv 2025

  59. [59]

    Qwen-robotworld technical report: Unifying embodied world modeling through language-conditioned video generation.arXiv preprint arXiv:2606.17030, 2026

    Jie Zhang, Xiaoyue Chen, Anzhe Chen, Chenxu Lv, Deqing Li, Gengze Zhou, Hang Yin, Haoqi Yuan, Haoyang Li, Jiahao Li, et al. Qwen-robotworld technical report: Unifying embodied world modeling through language-conditioned video generation.arXiv preprint arXiv:2606.17030, 2026. 3

    Pith/arXiv arXiv 2026

  60. [60]

    Drivedreamer4d: World models are effective data machines for 4d driving scene representation

    Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, et al. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. InProceedings of the computer vision and pattern recognition conference, pages 12015–12026, 2025. 3

    2025

  61. [61]

    Recondreamer++: Harmonizing generative and reconstructive models for driving scene representation

    Guosheng Zhao, Xiaofeng Wang, Chaojun Ni, Zheng Zhu, Wenkang Qin, Guan Huang, and Xingang Wang. Recondreamer++: Harmonizing generative and reconstructive models for driving scene representation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 26718–26728, 2025. 3

    2025

  62. [62]

    Drivedreamer-2: Llm-enhanced world models for diverse driving video generation

    Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 10412–10420, 2025. 1

    2025

  63. [63]

    Unidrivedreamer: A single-stage multimodal world model for autonomous driving.arXiv preprint arXiv:2602.02002, 2026

    Guosheng Zhao, Yaozeng Wang, Xiaofeng Wang, Zheng Zhu, Tingdong Yu, Guan Huang, Yongchen Zai, Ji Jiao, Changliang Xue, Xiaole Wang, et al. Unidrivedreamer: A single-stage multimodal world model for autonomous driving.arXiv preprint arXiv:2602.02002, 2026. 3

    arXiv 2026

  64. [64]

    Robodreamer: learning compositional world models for robot imagination

    Siyuan Zhou, Yilun Du, Jiaben Chen, Yandong Li, Dit-Yan Yeung, and Chuang Gan. Robodreamer: learning compositional world models for robot imagination. InProceedings of the 41st International Conference on Machine Learning, pages 61885–61896, 2024. 3

    2024

  65. [65]

    Drivedreamer-policy: A geometry-grounded world-action model for unified generation and planning.arXiv preprint arXiv:2604.01765, 2026

    Yang Zhou, Xiaofeng Wang, Hao Shao, Letian Wang, Guosheng Zhao, Jiangnan Shao, Jiagang Zhu, Tingdong Yu, Zheng Zhu, Guan Huang, et al. Drivedreamer-policy: A geometry-grounded world-action model for unified generation and planning.arXiv preprint arXiv:2604.01765, 2026. 3

    arXiv 2026

  66. [66]

    Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets.arXiv preprint arXiv:2504.02792, 2025

    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, 2025. 3

    Pith/arXiv arXiv 2025

  67. [67]

    Aether: Geometric-aware unified world modeling

    Haoyi Zhu, Yifan Wang, Jianjun Zhou, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Chunhua Shen, Jiangmiao Pang, and Tong He. Aether: Geometric-aware unified world modeling. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8535–8546, 2025. 3

    2025

  68. [68]

    Is sora a world simulator? a comprehensive survey on general world models and beyond.arXiv preprint arXiv:2405.03520, 2024

    Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Bohan Li, Nianchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond.arXiv preprint arXiv:2405.03520, 2024. 1 17

    arXiv 2024