From Dreamer to Genie: A Curated Path Into World Models for Engineers
World models have quietly become one of the most consequential ideas in machine learning. A system that learns how an environment behaves, rather than just mapping input to output, can predict what happens next, simulate possible outcomes and plan actions before committing to any of them. That capability now shows up in video generation, robotics, embodied AI and physical reasoning, and it sits behind systems such as Dreamer and Genie.
The problem is the on-ramp. There is no single resource that takes a working engineer from reinforcement learning fundamentals all the way to today’s world foundation models. The material exists, but it is scattered across university courses, paper reading lists, project sites and survey repositories, each written for a different audience and a different level of prior knowledge. What follows is a curated learning path assembled from five sources, ordered by what to learn first and why.
Start with reinforcement learning, not the latest paper
The instinct when approaching world models is to open the newest paper and work backward. That is the wrong direction. Many of the fundamental ideas behind world models come from reinforcement learning and model-based planning, and skipping that foundation makes the modern literature read like a collection of unrelated tricks.
UC Berkeley’s CS 285, Deep Reinforcement Learning, taught by Sergey Levine, is the recommended starting point. It covers RL, control and deep learning, with dedicated material on model-based RL. The lectures most relevant to world models are the ones on planning, optimal control, learned dynamics and model-based reinforcement learning, which is precisely where environment prediction and action selection meet. The course ships with lecture recordings, notes and assignments covering both the mathematics and practical implementations, so it rewards you if you want to build intuition by writing code rather than only reading it.
Work through CS 285 first, and terms like latent dynamics or imagined trajectories will land on prepared ground when they appear later.
Get the intuition before the math
Once the RL basics are in place, the next step is a conceptual map of the territory. Stanford’s Introduction to World Models lecture, part of the Spring 2026 CS348K course, is available online with downloadable PDF slides and covers visual and spatial computing workloads. It walks through generative AI-based world creation, latent world representations, non-pixel-based approaches and embodied agentic AI.
This is a starting point for understanding what world models are, not a full course, and it is worth treating it that way. The value is orientation: it shows how the pieces fit together before you commit to the heavier papers. If you have already met related ideas in adjacent fields, you will recognize the framing. Our coverage of Microsoft Research’s Quine: A World Model for Biology That Connects Models, Labs, and Literature makes the same point from a different direction: a world model represents the state of a system, predicts how that state evolves under intervention, and supports reasoning about what to do next. The vocabulary transfers across domains even when the applications do not.
Read the original World Models paper
With the fundamentals and the overview in hand, go to the source. David Ha and Jürgen Schmidhuber’s World Models popularized a deceptively simple idea: an agent can learn a compressed representation of its environment, learn how that representation evolves over time, and train a controller inside that internally simulated environment.
The architecture has three main components, and the key insight is that an agent can sometimes train inside its learned environment instead of interacting with the real environment at every training step. That substitution is what makes the approach practical, because real-world interaction is slow, expensive and sometimes unsafe.
The project website is the best entry point. It includes explanations, diagrams, animations, experiments and links to the implementation, which makes the paper far more approachable than the PDF alone. Modern systems such as Dreamer and large video-based world models are considerably more complex, but they build on the same core intuition, which is exactly why starting here pays off.
Follow the paper lineage
The original work is a foundation, not a destination. Stanford’s world models reading list highlights ten important papers tracing the progression from early neural world models to modern approaches, and it functions as a spine for the field. The list includes World Models, PlaNet, Dreamer, DreamerV2, SimPLe, DayDreamer, Genie and TD-MPC2.
The progression is instructive. The original World Models showed an agent could learn a compact internal simulator. PlaNet explored planning directly inside learned latent dynamics, avoiding the need to reconstruct pixels before deciding what to do. Dreamer shifted toward learning policies using imagined trajectories, and later systems extended similar concepts into robotics and large-scale generative environments. Read in order, the list works less like a set of separate papers and more like one argument developed over several years.
Keep a living reference on hand
Two references are worth bookmarking rather than reading end to end. The Awesome World Model repository was created by researchers associated with Tsinghua University alongside their survey, Understanding World or Predicting Future? A Comprehensive Survey of World Models, which was accepted by ACM Computing Surveys. The repository continuously organizes papers and resources across branches of world-model research, making it the closest thing to a current map of a fast-moving area.
The survey itself does the taxonomic work: it organizes the field’s branches and gives names to distinctions that are easy to miss when reading papers in isolation. Neither resource is designed for linear consumption. Both are for the moment when a new paper appears and you need to know where it fits.
An order of attack, and what to watch
The suggested sequence follows the logic of the sources rather than their prestige. Start with CS 285 for the reinforcement learning and model-based foundations. Move to the Stanford lecture for orientation. Read the original World Models paper alongside its project site. Then work through the Stanford reading list in roughly chronological order, letting PlaNet and Dreamer build on what came before. Keep the Awesome World Model repository and the Tsinghua survey as standing references.
What makes this worth the effort now is convergence. World models are moving out of simulation benchmarks and into systems that act in the physical world, supplying the predictive machinery that both robotics and generative video need. The path from Dreamer to Genie to whatever comes next is not going to get shorter, so build the foundation now and you will read the next wave of papers as a continuation rather than a surprise.
We covered open weight models engineers in more detail elsewhere.
