Dreamer 4 mines Minecraft diamonds purely in its imagination

Training an AI through live trial and error is painfully slow and expensive. If you are training a real-world robot, letting it randomly explore will break expensive hardware in minutes. Even in open-ended games, an agent can wander aimlessly for millions of steps without learning a thing.

The holy grail is to let AI build a mental simulation of reality and practice entirely inside its own head. A research team co-led by Danijar Hafner and Wilson Yan just unveiled Dreamer 4, taking this imagination-based training to a wild new level.

Dreamer 4 has become the first AI to obtain diamonds in Minecraft purely from offline data, without ever interacting with the live game environment during training.

Real-time imagination on a single GPU

Until now, world models have hit a wall when handling complex environments. They could generate pretty video frames, but they fell apart when trying to predict intricate object interactions and underlying rules over time.

Dreamer 4 changes that. In the sprawling sandbox of Minecraft, its world model accurately predicts physical interactions and game mechanics, beating previous world models by a wide margin.

Even better, the team paired an efficient transformer architecture with a shortcut forcing objective. This allows the high-fidelity world model to run real-time interactive inference on just a single GPU.

In plain English: the agent can simulate thousands of possibilities inside its imagination at blazingly fast speeds without burning through a massive server cluster.

20,000 steps of pure mental rehearsal

Getting a diamond in Minecraft is a notorious benchmark for long-horizon AI planning.

You start with bare hands, punch trees, craft tools, build a furnace, mine iron, forge pickaxes, and navigate dangerous underground caves while dodging lava. From raw pixels, an agent must execute a sequence of over 20,000 precise mouse and keyboard actions without fatal mistakes.

Past agents relied on millions of live interactions inside the game engine to stumble their way through. Dreamer 4 was given a far tougher constraint: zero real-time interaction, relying solely on offline datasets.

The world model learned general action conditioning from a tiny slice of data, pulling the vast bulk of its knowledge from diverse unlabeled videos. Dreamer 4 then ran reinforcement learning entirely within this mental simulation to master the 20,000-step chain.

Why it matters

This paper isn't really about Minecraft. It is a major proof of concept for the future of robotics and embodied AI.

In the physical world, training a robot through trial and error is practically impossible because hardware breaks, batteries die, and physical resets take hours. If we can instead feed models cheap, passive video footage and let them train inside an imagined world model, the economics of robotics change overnight.

We think Dreamer 4's biggest breakthrough is showing that imagination training can scale to tasks with massive step counts. Achieving real-time inference on a single GPU while handling over 20,000 consecutive actions proves world models are no longer just passive video generators.

The open question is the sim-to-real gap. Video game mechanics, while complex, run on deterministic code. Whether this exact architecture can handle the chaotic, unscripted mess of real-world physics remains to be seen. But as an offline learning blueprint, it looks remarkably sharp.