Dreamer 4

A 2025 Google DeepMind world-model agent that mined diamonds in Minecraft using only offline data.

Dreamer 4 is an agent released in September 2025 by Danijar Hafner, Wilson Yan, and Timothy Lillicrap at Google DeepMind, the newest generation in the Dreamer line (following DreamerV3 and earlier versions). The core idea behind the Dreamer approach is “learning in imagination”: first learn a world model, then train the policy with reinforcement learning entirely on experience generated by that model, without having to repeatedly interact with the real environment. Dreamer 4 scales the world model to about 2 billion parameters (a 400-million-parameter video tokenizer plus a 1.6-billion-parameter dynamics model), using a training objective called “shortcut forcing” together with an efficient Transformer to achieve real-time interactive inference on a single GPU. It learns most of its knowledge from video with no action labels, needing only a small amount of action-labeled data to learn action conditioning. It is the first agent to obtain a diamond in Minecraft using purely offline data. Because trial and error is slow and unsafe on real robots, this style of training inside a world model is seen as highly relevant to robotics.

ExampleMining a diamond in Minecraft requires more than 20,000 consecutive mouse and keyboard actions from raw pixels; trained on just 2,541 hours of offline player recordings, Dreamer 4 obtained a diamond in 0.7% of episodes, outperforming OpenAI's offline VPT agent while using about 100 times less data.

Also called

Training Agents Inside of Scalable World Models

Related

World Model · Learning in Imagination · DreamerV3 · Model-Based Reinforcement Learning · VPT · Offline Reinforcement Learning

Sources

Training Agents Inside of Scalable World Models (arXiv 2509.24527)
Dreamer 4 项目页（danijar.com） (Chinese)

As of

2025-09