Dreamer 4 becomes first reported agent to obtain Minecraft diamonds using only offline training data

Dreamer 4 is a two-billion-parameter world-model agent developed by Google DeepMind researchers Danijar Hafner, Wilson Yan, and Timothy Lillicrap. Its training has three broad stages: world-model pretraining on recorded video and actions, task-conditioned policy and reward-model training from the dataset, and reinforcement learning on trajectories generated inside the frozen world model. The final policy is then evaluated in the actual Minecraft engine, but the evaluation environment does not supply new training experience.

For the Offline Diamond Challenge, the researchers used 2,541 hours of contractor gameplay containing 360p video, mouse and keyboard actions, and event annotations. Evaluations began in randomly generated worlds with empty inventories, lasted up to 60 minutes, and used raw pixels with the native low-level control interface. The paper estimates that an experienced human needs about 20 minutes, corresponding to approximately 24,000 mouse and keyboard actions, to obtain a diamond.

Dreamer 4 obtained diamonds in 0.7% of 1,000 evaluation episodes—equivalent to seven successful episodes. Every evaluated offline baseline recorded zero diamond successes. Dreamer 4 also obtained iron pickaxes in 29.0% of episodes, compared with 16.9% for the same world-model representation followed only by behavioral cloning and 11.2% for the Gemma 3-based vision-language-action baseline. This ablation supports the narrower claim that imagination training improved the policy beyond imitation from the offline dataset.

Performance was much higher on earlier milestones: Dreamer 4 obtained stone pickaxes in 90.1% of episodes, iron ore in 66.7%, furnaces in 58.1%, iron ingots in 39.5%, and iron pickaxes in 29.0%. The increasing advantage over behavioral-cloning systems on later milestones is consistent with imagination training helping most as the required behavior becomes longer and less common in the dataset.

The world model was also evaluated as an interactive Minecraft simulator. A human operator completed 14 of 16 selected interaction tasks inside Dreamer 4, compared with 5 of 16 in the large Oasis model and none in Lucid-v1. Dreamer 4 generated at 21 frames per second on one H100 GPU with 9.6 seconds of context, allowing real-time interaction while preserving a substantially longer context than the compared models.

“Offline” describes the learning process, not the final evaluation. Dreamer 4 was evaluated in the actual Minecraft engine after training; it did not learn solely by executing inside an imagined game from random initialization. The world model, initial policy, and reward model first learned from recorded gameplay containing video, actions, and event annotations. Reinforcement learning inside the world model was the final improvement stage.

The system also did not receive one unrestricted instruction—such as “find a diamond”—and independently invent the entire plan. It was trained as a 20-task agent and guided during evaluation by a manually specified linear prompt sequence covering gathering wood, crafting successive tools, mining iron, building and using a furnace, and finally mining diamonds. This still requires executing the subtasks under changing procedural conditions, but it limits claims about autonomous task decomposition and open-ended planning.

The diamond success rate was only 0.7%. That is enough to establish a non-zero reported capability across a large evaluation, but it is not reliable task mastery. The result should therefore be interpreted as crossing the offline-training threshold, not as producing a dependable Minecraft agent.

Dreamer 4 provides strong initial evidence that a learned world model can become accurate enough for reinforcement learning inside generated experience to improve a policy on a difficult, long-horizon task and transfer that improvement back into the real evaluation environment. Its key contribution is not that another agent can play Minecraft; earlier systems had already reached diamonds or diamond tools. The frontier change is that the policy reached the final milestone without online trial-and-error learning.

The result remains narrow. It establishes this capability in Minecraft under an authored task hierarchy, using a large action-labelled dataset, substantial training compute, and a low final success rate. It does not establish that the method can yet train reliable general-purpose agents or physical robots entirely through imagination.