WORLD MODELS ATLAS
Latest source check: Oct 7, 2026Primary-source policyTürkçe
2018 → 2026

Evolution of world models

Selected primary-source milestones from latent environment representations to interactive video, driving, and spatial world generation.

  1. World Models

    The 2018 paper presented agents learning compact environment models for control.

    DemonstratedCurrent
    Primary source

  2. DreamerV3

    DreamerV3 unified latent imagination across diverse reinforcement-learning domains.

    DemonstratedCurrent
    Primary source

  3. GAIA-1

    Wayve published the first GAIA paper framing driving video as a generative world model.

    DemonstratedCurrent
    Primary source

  4. V-JEPA

    Meta introduced the V-JEPA latent video prediction architecture.

    DemonstratedCurrent
    Primary source

  5. Oasis

    Decart and Etched published Oasis for real-time interactive world generation.

    ReportedCurrent
    Primary source

  6. Cosmos 1.0

    NVIDIA announced the Cosmos world foundation model platform for physical AI.

    ReportedCurrent
    Primary source

  7. V-JEPA 2

    Meta connected predictive video representations with robot planning experiments.

    DemonstratedCurrent
    Primary source

  8. Genie 3

    DeepMind introduced a real-time interactive world-model research preview.

    ReportedCurrent
    Primary source

  9. GAIA-3

    Wayve announced a multi-view driving world model.

    ReportedCurrent
    Primary source

  10. GWM-1

    Runway introduced its general world model research line.

    ReportedCurrent
    Primary source

  11. Waymo World Model

    Waymo introduced its autonomous-driving simulation model.

    ReportedCurrent
    Primary source

  12. Cosmos 3

    NVIDIA published the Cosmos 3 technical report and open development artifacts.

    DemonstratedCurrent
    Primary source

  13. Atlas

    World Labs introduced Atlas for persistent multimodal spatial worlds.

    ReportedCurrent
    Primary source

  14. GWM Worlds 2

    Runway previewed a real-time world model generating 720p video and 48 kHz audio.

    ReportedCurrent
    Primary source

  15. Odyssey-3

    Odyssey introduced a foundation world model presented as a shared control backbone.

    ReportedCurrent
    Primary source

  16. Oasis 3

    Decart extended its interactive world model to multi-view physical-AI training.

    ReportedNeeds review
    Primary source