Overview

Dreamer 4 - Training Agents Inside Scalable World Models

Dreamer 4 is the fourth generation of the Dreamer algorithm family — a general-purpose reinforcement learning system that learns, plans, and optimizes policies entirely inside its own learned world model. Unlike its predecessors, Dreamer 4 no longer depends on online environment interaction. Instead, it learns from offline video datasets and imagines future trajectories within its internally simulated world to optimize behavior. This advancement closes the gap between data-driven simulation and real-world control, allowing AI agents to master complex, contact-rich environments such as Minecraft or robotic manipulation using only historical or synthetic data.

Dreamer 4 marks a shift from “learning from the world” to “learning inside a world” - training not through external trial and error, but through the accuracy and scalability of its own model of reality.