Article Contents
REVIEW   Open Access     Cite

World models: Toward computationally usable embodied AI

More Information
  • Corresponding author: wang-jx@pku.edu.cn 
  • DownLoad: Full size image
    1. Computational usability asks what to compute, how far to simulate, where to run it, and what value it adds.

      Action-oriented data help models prioritize the states that matter for decisions and safety.

      Risk-adaptive rollouts and device-edge-cloud computing may make world models more deployable.

      Action gain per computational cost measures improvements in decision-making relative to computing and resource costs.

  • World models are evolving from latent dynamics modules in model-based reinforcement learning into critical infrastructure for embodied intelligence, autonomous driving, robot learning, and physical AI. However, current discussions often equate world models with larger-scale video generation or more realistic interactive simulation, overlooking a fundamental constraint: embodied systems must act under constraints on limited data, computation, and communication under real-time safety requirements. This paper frames this challenge as the “computational usability gap.” An actionable world model must not only predict the future, but also determine what to compute, how far to roll out, where to compute, and whether the computation delivers sufficient action value. We review recent progress in representation, training and rollout, deployment, and evaluation, and argue that the focus is shifting from visual fidelity to action utility per unit of computation.
  • 加载中
  • [1] Ha D. and Schmidhuber J. (2018). Recurrent world models facilitate policy evolution. Adv. Neural Inf. Process. Syst. 31:2451−2463. DOI:10.48550/arXiv.1809.01999

    View in Article CrossRef Google Scholar

    [2] Hafner D., Lillicrap T., Fischer I., et al. (2019). Learning latent dynamics for planning from pixels. Proc. Mach. Learn. Res. 97:2555−2565. DOI:10.48550/arXiv.1811.04551

    View in Article CrossRef Google Scholar

    [3] Hafner D., Lillicrap T., Ba J., et al. (2020). Dream to control: Learning behaviors by latent imagination. International Conference on Learning Representations. DOI: 10.48550/arXiv.1912.01603

    View in Article Google Scholar

    [4] Hafner D., Pasukonis J., Ba J., et al. (2025). Mastering diverse control tasks through world models. Nature 640:647−653. DOI:10.1038/s41586-025-08744-2

    View in Article CrossRef Google Scholar

    [5] Brooks T., Peebles B., Holmes C., et al. (2024). Video generation models as world simulators. OpenAI Technical Report

    View in Article Google Scholar

    [6] Assran M., Bardes A., Fan D., et al. (2025). V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. Preprint at arXiv:2506.09985. DOI: 10.48550/arXiv.2506.09985

    View in Article Google Scholar

    [7] World Labs. (2025). Marble: A multimodal world model. World Labs Research Release

    View in Article Google Scholar

    [8] World Labs. (2025). RTFM: A real-time frame model. Research preview

    View in Article Google Scholar

    [9] Parker-Holder J., Ball P., Bruce J., et al. (2024). Genie 2: A large-scale foundation world model. Google DeepMind Research Report

    View in Article Google Scholar

    [10] Parker-Holder J. and Fruchter S. (2025). Genie 3: A new frontier for world models. Google DeepMind Research Report

    View in Article Google Scholar

    [11] NVIDIA. (2025). Cosmos world foundation model platform for physical AI. NVIDIA Technical Report DOI: 10.48550/arXiv.2501.03575

    View in Article Google Scholar

    [12] Bordes F., Garrido Q., Kao J.T., et al. (2025). IntPhys 2: Benchmarking intuitive physics understanding in complex synthetic environments. Preprint at arXiv:2506.09849. DOI: 10.48550/arXiv.2506.09849

    View in Article Google Scholar

    [13] Jiang F., Chen Y., Xu K., et al. (2026). RoboWM-Bench: A benchmark for evaluating world models in robotic manipulation. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recogn Workshops: 4455–4460. DOI: 10.48550/arXiv.2604.19092

    View in Article Google Scholar

    [14] Unterthiner T., van Steenkiste S., Kurach K., et al. (2019). Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717. DOI: 10.48550/arXiv.1812.01717

    View in Article Google Scholar

    [15] Assran M., Duval Q., Misra I., et al. (2023). Self-supervised learning from images with a joint-embedding predictive architecture. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recogn.: 15619–15629. DOI: 10.1109/CVPR52729.2023.01499

    View in Article Google Scholar

    [16] Bardes A., Garrido Q., Ponce J., et al. (2024). Revisiting feature prediction for learning visual representations from video. Trans. Mach. Learn. Res. DOI: 10.48550/arXiv.2404.08471

    View in Article Google Scholar

    [17] Zhang Z., Wang J., Lin C., et al. (2026). TWM: Multimodal data generation engine via traffic world model. Communications in Transportation Research 6(3):9640045. DOI:10.26599/COMMTR.2026.9640045

    View in Article CrossRef Google Scholar

    [18] Tian S., Zheng Y., Zheng Y., et al. (2026). VT-WAM: Visual-tactile world action model for contact-rich manipulation. Preprint at arXiv:2607.02503. DOI: 10.48550/arXiv.2607.02503

    View in Article Google Scholar

    [19] Higuera C., Arnaud S., Boots B., et al. (2026). Visuo-tactile world models. Preprint at arXiv:2602.06001. DOI: 10.48550/arXiv.2602.06001

    View in Article Google Scholar

    [20] Zheng Y., Gu S., Li W., et al. (2026). OmniVTA: Visuo-tactile world modeling for contact-rich robotic manipulation. Preprint at arXiv:2603.19201. DOI: 10.48550/arXiv.2603.19201

    View in Article Google Scholar

    [21] Zhang Z., Zhou P., Zhang K., et al. (2026). ContactWorld: What matters in vision-tactile world models for contact-rich manipulation. Preprint at arXiv:2606.13877. DOI: 10.48550/arXiv.2606.13877

    View in Article Google Scholar

    [22] Wang S., Shi J., Fu Z., et al. (2026). World action models: The next frontier in embodied AI. Preprint at arXiv:2605.12090. DOI: 10.48550/arXiv.2605.12090

    View in Article Google Scholar

    [23] Cen J., Yu C., Yuan H., et al. (2025). WorldVLA: Towards autoregressive action world model. Adv. Neural Inf. Process. Syst (NeurIPS) 2026 DOI: 10.48550/arXiv.2506.21539

    View in Article Google Scholar

    [24] Yang Y., Liu Z., Kou S., et al. (2026). World-language-action model for unified world modeling, language reasoning, and action synthesis. Conference on Robot Learning (CoRL) 2026 DOI: 10.48550/arXiv.2606.05979

    View in Article Google Scholar

    [25] Russell L., Hu A., Bertoni L., et al. (2025). GAIA-2: A controllable multi-view generative world model for autonomous driving. Preprint at arXiv:2503.20523. DOI: 10.48550/arXiv.2503.20523

    View in Article Google Scholar

    [26] Ren X., Lu Y., Cao T., et al. (2025). Cosmos-Drive-Dreams: Scalable synthetic driving data generation with world foundation models. Preprint at arXiv:2506.09042. DOI: 10.48550/arXiv.2506.09042

    View in Article Google Scholar

    [27] NVIDIA. (2026). NVIDIA OmniDreams: Real-time generative world model for closed-loop autonomous vehicle simulation. Preprint at arXiv:2606.03159. DOI: 10.48550/arXiv.2606.03159

    View in Article Google Scholar

    [28] Zhao P., Lin J., Rupprecht T., et al. (2026). PhyWorld: Physics-faithful world model for video generation. Preprint at arXiv:2605.19242. DOI: 10.48550/arXiv.2605.19242

    View in Article Google Scholar

    [29] Oh Y. and Yun S. (2026). Reference-free assessment of physical consistency in world model-based video generation. Preprint at arXiv:2606.22363. DOI: 10.48550/arXiv.2606.22363

    View in Article Google Scholar

    [30] NVIDIA. (2025). Cosmos-Transfer1: Conditional world generation with adaptive multimodal control. Preprint at arXiv:2503.14492. DOI: 10.48550/arXiv.2503.14492

    View in Article Google Scholar

    [31] Kim M.J., Pertsch K., Karamcheti S., et al. (2025). OpenVLA: An open-source vision-language-action model. Proc. Mach. Learn. Res. 270:2679−2713. DOI:10.48550/arXiv.2406.09246

    View in Article CrossRef Google Scholar

    [32] Yue Y., Wang Y., Kang B., et al. (2024). DeeR-VLA: Dynamic inference of multimodal large language models for efficient robot execution. Adv. Neural Inf. Process. Syst (NeurIPS). 37:56619−56643. DOI:10.52202/079017-1803

    View in Article CrossRef Google Scholar

    [33] Kim M.J., Finn C. and Liang P. (2025). Fine-tuning vision-language-action models: Optimizing speed and success. Proc. Robot. Sci. Syst. DOI: 10.15607/RSS.2025.XXI.017

    View in Article Google Scholar

    [34] Liu B., Poe W.Y., Di J., et al. (2026). ComVLA: Communication-aware split inference for VLA models in 6G-connected robotics. Preprint at arXiv:2609.07838. DOI: 10.48550/arXiv.2609.07838

    View in Article Google Scholar

    [35] Yang T., Shen Z., Mi Z., et al. (2026). MiraBench: Evaluating action-conditioned reliability in robotic world models. Preprint at arXiv:2605.29360. DOI: 10.48550/arXiv.2605.29360

    View in Article Google Scholar

    [36] Li J., Guo T., Ye Y., et al. (2026). Efficient-WAM: A 1B-parameter world-action model with low-cost future imagination. Preprint at arXiv:2606.10040. DOI: 10.48550/arXiv.2606.10040

    View in Article Google Scholar

    [37] Bi H., Tan H., Xie S., et al. (2026). Motus: A unified latent action world model. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR): 35101–35113. DOI: 10.48550/arXiv.2512.13030.

    View in Article Google Scholar

  • Cite this article:

    Song Y., Wang J., Liu H., et al. (2026). World models: Toward computationally usable embodied AI. The Innovation Insights 1:100018. https://doi.org/10.59717/j.tiis.2026.100018
    Song Y., Wang J., Liu H., et al. (2026). World models: Toward computationally usable embodied AI. The Innovation Insights 1:100018. https://doi.org/10.59717/j.tiis.2026.100018

Welcome!

To request copyright permission to republish or share portions of our works, please visit Copyright Clearance Center's (CCC) Marketplace website at marketplace.copyright.com.

Figures(2)     Tables(2)

Share

  • Share the QR code with wechat scanning code to friends and circle of friends.

Article Metrics

Article views(788) PDF downloads(597)

Relative Articles

Cited by

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint