World Models
World Models are learned internal representations that allow an AI agent to simulate, predict, and reason about how environments evolve over time in response to actions. Rather than interacting with the real world for every decision, an agent equipped with a world model can 'imagine' the consequences of potential actions inside a compressed latent space. The concept spans model-based reinforcement learning, video generation, physical simulation, and robotics.
World models have moved from academic curiosity to strategic priority: NVIDIA (Cosmos), Google DeepMind (Genie 2), Meta (V-JEPA 2), and a wave of well-funded startups are shipping world model systems for robotics, autonomous vehicles, and game simulation. Companies hiring in this space need engineers who understand latent dynamics modeling, recurrent state-space models, and imagination-based policy training—skills that sit at the intersection of deep learning, RL, and generative modeling, making them rare and highly valued in 2026.
🎓 Courses
Deep Reinforcement Learning (CS285)
by Sergey Levine
The definitive graduate-level RL course. Lectures 11-13 cover model-based RL and world models in depth, including RSSM-style architectures. All lectures are free on YouTube.
Deep Reinforcement Learning Course
by Thomas Simonini
Free, self-paced, hands-on course with Colab notebooks. Builds the RL foundations (policy gradients, actor-critic) needed to understand Dreamer-style world model agents.
From Video Generation to World Model
by Speakers from UC Berkeley, Stanford, Google DeepMind, Luma AI, Kuaishou
A full-day tutorial from CVPR 2025 tracing the path from diffusion-based video generation to controllable, physics-grounded world models—exactly the frontier of the field in 2025.
Introducing Dreamer: Scalable Reinforcement Learning Using World Models
by Danijar Hafner (Google DeepMind)
Accessible written walkthrough by the lead author of the Dreamer family, explaining how world models enable imagination-based training. A strong conceptual entry point before reading the papers.
📖 Books
Model-Based Reinforcement Learning: A Survey
Thomas M. Moerland, Joost Broekens, Aske Plaat, Catholijn M. Jonker · 2023
Published in Foundations and Trends in Machine Learning (2023), this is the most comprehensive survey-level reference on model-based RL, covering world model architectures, Dyna-style methods, and latent-space planning in a single structured volume.
🛠️ Tutorials & Guides
World Models Reading List: The Papers You Actually Need in 2025
A practitioner-curated reading map from model-free intuition through RSSM, Dreamer, IRIS, and diffusion-based world models. Useful for structuring a self-study curriculum around primary sources.
World Models (RL Journal Club post)
Traces the lineage from Sutton's Dyna-Q through Ha & Schmidhuber, Dreamer, and IRIS with clear explanations of how each generation improved on the last. Good for understanding historical context and design choices.
Training Agents Inside of Scalable World Models (Dreamer 4)
Danijar Hafner's own project page for Dreamer 4, with demos, code, and a concise summary of how offline world model training advances the frontier. Keeps you current with the lead researcher's latest work.
Learning resources last updated: June 18, 2026