Static research summary generated from local editorial content.
| Attribute | Value |
|---|---|
| Topic | World Models: A Comprehensive Survey |
| Summary | A survey of AI world models covering taxonomy, leading architectures, landmark systems, open challenges, and future research directions. |
| Related Models | 8 |
| Citations | 4 |
Editorial body section preserved directly in static HTML.
World models span a wide range of AI systems: from latent dynamics models for RL (RSSM, Dreamer) to abstract planning models (Predictron, MuZero), from generative environment simulators (Cosmos, Genie 2) to self-supervised visual models (V-JEPA). This survey organizes them by architecture, learning paradigm, and application domain.
Editorial body section preserved directly in static HTML.
We organize world models along three axes: (1) Architecture: RSSM-based, autoregressive, diffusion-based, abstract; (2) Learning paradigm: RL-based, self-supervised, supervised; (3) Domain: games, robotics, autonomous driving, general-purpose. Each axis captures different trade-offs and design choices.
Editorial body section preserved directly in static HTML.
Foundational (2017-2019): Predictron, I2A, Ha & Schmidhuber's World Models, PlaNet/RSSM. Scaling (2020-2022): MuZero, DreamerV2. Generalization (2023-present): DreamerV3, TD-MPC2, IRIS, UniSim, NVIDIA Cosmos, Genie 2, V-JEPA. Each era brought fundamental shifts in capabilities.
Editorial body section preserved directly in static HTML.
Persistent challenges include: long-horizon prediction stability, sample efficiency in complex real-world domains, bridging the reality gap for robotics, scaling world models to open-ended environments, and integrating world models with language understanding for broader reasoning.
Editorial body section preserved directly in static HTML.
The field is converging toward foundation world models trained at massive scale on video data, combining physical understanding with language reasoning. The integration of world models with LLMs, the development of universal simulators, and the push toward embodied general intelligence are key research frontiers.
| Model | Lab | Category | Index v1.1 |
|---|---|---|---|
| DreamerV3 | Google DeepMind | Model-Based RL | 88/100 |
| MuZero | Google DeepMind | Model-Based RL | 78/100 |
| NVIDIA Cosmos | NVIDIA | Foundation World Model | 87/100 |
| Genie 2 | Google DeepMind | Generative World Model | 79/100 |
| V-JEPA | Meta | Self-Supervised World Model | 70/100 |
| World Models (Ha & Schmidhuber) | Google Brain / IDSIA | Model-Based RL | 48/100 |
| PlaNet | Model-Based RL | 57/100 | |
| Predictron | Google DeepMind | Model-Based RL | 43/100 |
FAQ answers rendered directly into static HTML for extractable responses.
There are dozens of distinct world model architectures, but they cluster into a few families: RSSM-based (Dreamer), search-based (MuZero), autoregressive (IRIS), diffusion-based (Cosmos, UniSim), and joint embedding (V-JEPA). Each family has its own strengths and trade-offs.
Ha & Schmidhuber's 'World Models' (2018) is the most cited and conceptually influential. For practical impact, DreamerV3 (2023) demonstrated the first domain-general world model agent.
Short extractable summary preserved directly in static HTML.
Editorial provenance and refresh policy preserved directly in static HTML.
Published by world-models.io editorial board.
Lead editor Bernard Grenat.
This research page curates topic explanations, linked models, and citations grounded in primary research sources.
Each editorial page is assembled from primary sources, normalized into extractable summaries, checked for factual drift, and reviewed before publication or major refreshes. Last reviewed: 2026-06-21.
Pages are refreshed when a new paper, benchmark, release, architecture update, or stronger primary source materially changes the answer a reader or AI system should retrieve.
Each page links back to relevant primary sources and keeps a stable canonical URL so readers can verify claims, trace context, and reference the most up-to-date version. See the editorial policy.
Primary research citations embedded in static HTML.