­ýÊÿÿôýf¥ÿûbÿÿÿÿÿÿ±ç(ÿÿÿ:ëiÿÿÿ# world-models.io

> The Knowledge Hub for AI World Models. An independent, citation-first reference covering world models across robotics, model-based reinforcement learning, embodied AI, foundation physical AI, simulation engines, autonomous driving and neural game engines.

world-models.io is independently maintained by Bernard Grenat, who leads the editorial process. It maintains a structured database of 48 world models, hundreds of cross-referenced research papers, programmatic comparisons, a real-time release timeline, a glossary and editorial guides. Content is curated from primary sources (arXiv, official lab reports, technical blogs from DeepMind, Meta FAIR, OpenAI, NVIDIA, Wayve and others), reviewed for scope, attribution and factual drift, and updated when stronger evidence changes the answer.

## What is a world model

A world model is an AI system that learns how an environment works: it predicts future states from current observations and actions, supports planning and imagination, and enables sample-efficient learning. World models contrast with model-free reinforcement learning (which learns direct policies) and with passive video generators (which lack explicit action conditioning).

Three converging tracks define the 2025-2026 frontier:
1. Interactive simulation: Genie 3, OASIS, GameNGen
2. Self-supervised understanding: V-JEPA 2, I-JEPA
3. Foundation physical AI: NVIDIA Cosmos, AMI World Model

## Featured world models

- DreamerV3 (Hafner et al., 2023) - General algorithm mastering diverse domains via the RSSM. [https://world-models.io/en/models/dreamer-v3](https://world-models.io/en/models/dreamer-v3)
- Genie 3 (Google DeepMind, 2025) - Real-time interactive 3D world generation at 24 fps with up to one minute of world memory. [https://world-models.io/en/models/genie-3](https://world-models.io/en/models/genie-3)
- V-JEPA 2 (Meta FAIR, 2025) - 1.2B-parameter self-supervised video world model with zero-shot robot transfer. [https://world-models.io/en/models/v-jepa-2](https://world-models.io/en/models/v-jepa-2)
- NVIDIA Cosmos (NVIDIA, 2025) - Open foundation world model platform for physical AI. [https://world-models.io/en/models/nvidia-cosmos](https://world-models.io/en/models/nvidia-cosmos)
- MuZero (Schrittwieser et al., 2020) - Superhuman game play through learned planning without rules. [https://world-models.io/en/models/muzero](https://world-models.io/en/models/muzero)
- Sora (OpenAI, 2024) - Video generation model with emergent world simulation. [https://world-models.io/en/models/sora](https://world-models.io/en/models/sora)
- DIAMOND - Diffusion-based world model achieving SOTA on Atari 100K. [https://world-models.io/en/models/diamond](https://world-models.io/en/models/diamond)
- TD-MPC2 - Scalable world model agent across 104+ tasks. [https://world-models.io/en/models/td-mpc2](https://world-models.io/en/models/td-mpc2)
- IRIS - Autoregressive world model bridging language modeling and RL. [https://world-models.io/en/models/iris](https://world-models.io/en/models/iris)
- GAIA-1 (Wayve) - Generative world model for autonomous driving. [https://world-models.io/en/models/gaia-1](https://world-models.io/en/models/gaia-1)
- OASIS - Open-source real-time neural game engine. [https://world-models.io/en/models/oasis](https://world-models.io/en/models/oasis)

## Primary entry points

- Models database: [https://world-models.io/en/world-models-database](https://world-models.io/en/world-models-database)
- Open World Models Dataset (JSON, CSV, schema, CC0): [https://world-models.io/en/datasets/world-models](https://world-models.io/en/datasets/world-models)
- Leaderboard with Performance Index: [https://world-models.io/en/leaderboard](https://world-models.io/en/leaderboard)
- Research topics: [https://world-models.io/en/research](https://world-models.io/en/research)
- Programmatic comparisons: [https://world-models.io/en/compare](https://world-models.io/en/compare)
- Editorial guides: [https://world-models.io/en/guides](https://world-models.io/en/guides)
- Labs and organizations: [https://world-models.io/en/labs](https://world-models.io/en/labs)
- Release timeline: [https://world-models.io/en/timeline](https://world-models.io/en/timeline)
- Glossary: [https://world-models.io/en/glossary](https://world-models.io/en/glossary)
- Knowledge graph: [https://world-models.io/en/graph](https://world-models.io/en/graph)
- Methodology: [https://world-models.io/en/methodology/performance-index](https://world-models.io/en/methodology/performance-index)
- About world-models.io: [https://world-models.io/en/about](https://world-models.io/en/about)
- Editorial policy: [https://world-models.io/en/editorial-policy](https://world-models.io/en/editorial-policy)
- Contact: [https://world-models.io/en/contact](https://world-models.io/en/contact)

## Canonical reports

- State of World Models 2026 by Bernard GRENAT: canonical publication record, downloadable PDF, HTML edition, figures, bibliography, citation formats, CC BY 4.0 license, and DOI 10.5281/zenodo.21345187. [https://world-models.io/reports/state-of-world-models-2026/](https://world-models.io/reports/state-of-world-models-2026/)

## Featured comparisons

<!-- AUTO:COMPARISONS:START - do not edit by hand. Regenerated by scripts/update-llms.ts -->
- DreamerV3 vs PlayWorld: Two green-index leaders for acting under learned dynamics, but with different centers of gravity. DreamerV3 is the canonical imagination-based general RL agent, while PlayWorld... [https://world-models.io/en/compare/dreamer-v3-vs-playworld](https://world-models.io/en/compare/dreamer-v3-vs-playworld)
- Genie 3 vs NVIDIA Cosmos: Two green-index frontier systems with different ambitions. Genie 3 is a real-time text-to-world interactive generator, while NVIDIA Cosmos is a broad physical-AI platform optimi... [https://world-models.io/en/compare/genie-3-vs-nvidia-cosmos](https://world-models.io/en/compare/genie-3-vs-nvidia-cosmos)
- Genie 3 vs V-JEPA 2: Two green-index leaders that represent different frontier philosophies. Genie 3 is an interactive generative world model that turns text into playable environments, while V-JEPA... [https://world-models.io/en/compare/genie-3-vs-v-jepa-2](https://world-models.io/en/compare/genie-3-vs-v-jepa-2)
- NVIDIA Cosmos vs V-JEPA 2: Two green-index foundation-scale leaders with different views of world modeling. Cosmos emphasizes a platform for physical-AI simulation and generation, while V-JEPA 2 emphasize... [https://world-models.io/en/compare/nvidia-cosmos-vs-v-jepa-2](https://world-models.io/en/compare/nvidia-cosmos-vs-v-jepa-2)
- PlayWorld vs TD-MPC2: Two green-index models for robot decision-making, but with very different operating modes. PlayWorld learns a manipulation-focused world simulator from autonomous play, while TD... [https://world-models.io/en/compare/playworld-vs-td-mpc2](https://world-models.io/en/compare/playworld-vs-td-mpc2)
- PlayWorld vs V-JEPA 2: Two green-index models pushing robotics-relevant world understanding in different ways. PlayWorld is a robot-play simulator for manipulation and policy improvement, while V-JEPA... [https://world-models.io/en/compare/playworld-vs-v-jepa-2](https://world-models.io/en/compare/playworld-vs-v-jepa-2)
- LeWorldModel vs DreamerV3: LeWorldModel revisits LeCun's energy-based JEPA philosophy for control, predicting in latent space without pixel reconstruction. DreamerV3 remains the canonical RSSM-based agent... [https://world-models.io/en/compare/leworldmodel-vs-dreamer-v3](https://world-models.io/en/compare/leworldmodel-vs-dreamer-v3)
- Pandora vs Genie 2: Both generate explorable 3D-feeling environments, but with different control surfaces: Pandora accepts free-form text actions through an LLM backbone, while Genie 2 conditions o... [https://world-models.io/en/compare/pandora-vs-genie-2](https://world-models.io/en/compare/pandora-vs-genie-2)
- PixVerse R1 vs Sora: PixVerse R1 introduces reasoning-trained generation to text-to-video, optimizing for prompt adherence and physical plausibility. Sora remains the reference for cinematic length... [https://world-models.io/en/compare/pixverse-r1-vs-sora](https://world-models.io/en/compare/pixverse-r1-vs-sora)
- Sora vs Genie 3: Two leading generative world models with opposite design goals: Sora targets long, cinematic, non-interactive clips from text, while Genie 3 trades fidelity for real-time contro... [https://world-models.io/en/compare/sora-vs-genie-3](https://world-models.io/en/compare/sora-vs-genie-3)
- V-JEPA 2 vs I-JEPA: Two milestones of the JEPA roadmap: I-JEPA established self-supervised image representation by predicting in latent space; V-JEPA 2 extends the paradigm to video at foundation s... [https://world-models.io/en/compare/v-jepa-2-vs-i-jepa](https://world-models.io/en/compare/v-jepa-2-vs-i-jepa)
- Genie 3 vs Genie 2: Two generations of DeepMind's interactive world model. Genie 2 generates 3D environments from single images; Genie 3 generates them from text prompts in real time at 24fps with... [https://world-models.io/en/compare/genie-3-vs-genie-2](https://world-models.io/en/compare/genie-3-vs-genie-2)
- Full comparisons index: [https://world-models.io/en/compare](https://world-models.io/en/compare)
<!-- AUTO:COMPARISONS:END -->

## Curated collections

- Frontier World Models 2026: [https://world-models.io/en/frontier-world-models-2026](https://world-models.io/en/frontier-world-models-2026)
- Top World Models: [https://world-models.io/en/top-world-models](https://world-models.io/en/top-world-models)
- World Models for Robotics: [https://world-models.io/en/world-models-for-robotics](https://world-models.io/en/world-models-for-robotics)
- Video World Models: [https://world-models.io/en/video-world-models](https://world-models.io/en/video-world-models)
- Autonomous Driving World Models: [https://world-models.io/en/autonomous-driving-world-models](https://world-models.io/en/autonomous-driving-world-models)
- Game World Models: [https://world-models.io/en/game-world-models](https://world-models.io/en/game-world-models)
- World Models Research Papers: [https://world-models.io/en/world-models-research-papers](https://world-models.io/en/world-models-research-papers)

## Feeds

- RSS: [https://world-models.io/rss.xml](https://world-models.io/rss.xml)
- Sitemap: [https://world-models.io/sitemap.xml](https://world-models.io/sitemap.xml)

## Citation policy

All entities cite primary sources (arXiv DOIs, lab technical reports). Editorial guides include lastUpdated timestamps. Performance Index methodology is fully documented at /en/methodology/performance-index. Content is licensed for reference and citation by AI assistants, search engines and human readers.

## Editorial review

- Publisher: world-models.io editorial team
- Coverage: Evidence-backed explainers, comparisons, research maps and release tracking for AI world models
- Review policy: Updated when major papers, benchmarks, or model releases materially change the landscape

<!-- AUTO:BUILD:START -->
Last build: 2026-07-27
<!-- AUTO:BUILD:END -->

<!-- AUTO:GLOSSARY:START - do not edit by hand. Regenerated by scripts/update-llms.ts -->
- Latent Space: A compressed, abstract representation of data learned by a neural network. In world models, the latent space encodes environment states in a compact form tha... [https://world-models.io/en/glossary/latent-space](https://world-models.io/en/glossary/latent-space)
- RSSM (Recurrent State-Space Model): A neural network architecture that combines deterministic recurrent states with stochastic latent variables to model environment dynamics. The RSSM forms the... [https://world-models.io/en/glossary/rssm](https://world-models.io/en/glossary/rssm)
- Imagination-Based Learning: A training paradigm where an agent uses its world model to generate synthetic (imagined) trajectories, then learns policies from these imagined experiences r... [https://world-models.io/en/glossary/imagination-based-learning](https://world-models.io/en/glossary/imagination-based-learning)
- Sample Efficiency: The ability of an algorithm to learn effectively from a small number of real-world interactions. Model-based RL agents achieve high sample efficiency by supp... [https://world-models.io/en/glossary/sample-efficiency](https://world-models.io/en/glossary/sample-efficiency)
- Model-Based Reinforcement Learning: An RL approach where the agent learns a predictive model of the environment (a world model) and uses it to simulate outcomes, plan ahead, and learn from imag... [https://world-models.io/en/glossary/model-based-rl-term](https://world-models.io/en/glossary/model-based-rl-term)
- Latent Dynamics: The learned transition function within a latent space that predicts how hidden states evolve over time in response to actions. Latent dynamics models avoid c... [https://world-models.io/en/glossary/latent-dynamics](https://world-models.io/en/glossary/latent-dynamics)
- Diffusion Model: A generative model that learns to produce data by reversing a gradual noising process. In world models, diffusion-based architectures (used in NVIDIA Cosmos... [https://world-models.io/en/glossary/diffusion-model](https://world-models.io/en/glossary/diffusion-model)
- Autoregressive Model: A model that generates outputs one element at a time, where each output depends on previous elements. IRIS treats world modeling as autoregressive token pred... [https://world-models.io/en/glossary/autoregressive-model](https://world-models.io/en/glossary/autoregressive-model)
- Tokenizer: A component that converts continuous data (images, video frames, states) into discrete tokens that can be processed by transformer-based architectures. In wo... [https://world-models.io/en/glossary/tokenizer](https://world-models.io/en/glossary/tokenizer)
- Sim-to-Real Transfer: The process of transferring policies or skills learned in simulation to real-world robots or environments. World models help narrow the 'reality gap' by lear... [https://world-models.io/en/glossary/sim-to-real](https://world-models.io/en/glossary/sim-to-real)
- Reality Gap: The discrepancy between simulated and real-world environments that causes policies trained in simulation to fail when deployed on physical systems. Learned w... [https://world-models.io/en/glossary/reality-gap](https://world-models.io/en/glossary/reality-gap)
- Foundation Model: A large-scale model trained on broad data that can be adapted to many downstream tasks. Foundation world models (like NVIDIA Cosmos and Genie 2) learn genera... [https://world-models.io/en/glossary/foundation-model](https://world-models.io/en/glossary/foundation-model)
- Full glossary: [https://world-models.io/en/glossary](https://world-models.io/en/glossary)
<!-- AUTO:GLOSSARY:END -->

<!-- AUTO:TIMELINE:START - do not edit by hand. Regenerated by scripts/update-llms.ts -->
- [Lab Update] NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics: Today, we are introducing the next step: Cosmos-H-Dreams, a real-time, action-conditioned generative simulator for surgical robotics. [https://world-models.io/en/timeline/hugging-face-blog-nvidia-cosmos-h-dreams-bringing-real-time-generative-simulatio](https://world-models.io/en/timeline/hugging-face-blog-nvidia-cosmos-h-dreams-bringing-real-time-generative-simulatio)
- [Lab Update] Introducing Cosmos 3 Edge: Today we are releasing NVIDIA Cosmos 3 Edge on Hugging Face's Cosmos 3 repository. It's a 4-billion-parameter open world model that helps robots and vision AI agents understand... [https://world-models.io/en/timeline/hugging-face-blog-introducing-cosmos-3-edge-1784567224835](https://world-models.io/en/timeline/hugging-face-blog-introducing-cosmos-3-edge-1784567224835)
- [Lab Update] Introducing North Mini Code: Cohere's first model for developers Jun 09, 2026 3 min read: North Mini Code's benchmark scores translate to a 33. 4 on the Artificial Analysis Coding Index, a competitive position among similarly sized models. North Mini Code is designed... [https://world-models.io/en/timeline/cohere-research-introducing-north-mini-code-cohere-s-first-model-for-developers-](https://world-models.io/en/timeline/cohere-research-introducing-north-mini-code-cohere-s-first-model-for-developers-)
- [Lab Update] Waabi proves autonomous truck generalization with Volvo Autonomous Solutions: The self-driving company announced in a blog post that its Waabi Driver software, trained exclusively on a Peterbilt 579, took control of a Volvo VNL Autonomous truck and drove... [https://world-models.io/en/timeline/waabi-insights-waabi-proves-autonomous-truck-generalization-with-volvo-autonomou](https://world-models.io/en/timeline/waabi-insights-waabi-proves-autonomous-truck-generalization-with-volvo-autonomou)
- [Paper Published] Concept and architecture of a Business World Model (BWM): Businesses are increasingly adopting AI-enabled tools to improve productivity, reduce costs, and enhance products and services. [https://world-models.io/en/timeline/arxiv-cs-ai-business-world-model-1781064567990](https://world-models.io/en/timeline/arxiv-cs-ai-business-world-model-1781064567990)
- Full timeline: [https://world-models.io/en/timeline](https://world-models.io/en/timeline)
<!-- AUTO:TIMELINE:END -->ÿÿÿ