Which architecture is best?

It depends on the application. RSSM for sample-efficient RL, diffusion for realistic simulation, autoregressive for connecting to LLM methodology, abstract for decision-making optimization, JEPA for self-supervised physical understanding.

Will transformers replace RSSM-based world models?

Transformers are increasingly used in world models (IRIS, Cosmos), but RSSM remains highly effective for RL. The trend is toward combining transformer scalability with RSSM-like structured state representations.

New: the Timeline is live. Track world model releases, papers, and benchmark updates in real time.

world-models.io

The Knowledge Hub for AI World Models

World Models Architecture

An overview of the major architectural approaches used in AI world models, from the RSSM to transformer-based and diffusion-based designs.

robotics model-based-rl simulation embodied-ai

External Sources

Primary references and official sources surfaced directly in static HTML for crawlers and no-JS readers.

PlaNet (2019)
RSSM architecture
V-JEPA paper
Predictive representation architecture
Cosmos platform
Foundation world model architecture

References

[1] Hafner et al., 2019. Learning Latent Dynamics for Planning from Pixels.
[2] Micheli et al., 2023. Transformers are Sample-Efficient World Models.
[3] NVIDIA, 2024. Cosmos.

This site requires JavaScript for interactive features. Crawlable, page-specific HTML is served for every canonical URL in the sitemap.