Dreamer 4 becomes first reported agent to obtain Minecraft diamonds using only offline training data
Google DeepMind researchers report that Dreamer 4 obtained diamonds in 0.7% of 1,000 60-minute Minecraft evaluations after learning from a fixed 2,541-hour contractor gameplay dataset and improving its policy through reinforcement learning inside a learned world model, without collecting new Minecraft experience during training.
The Reality Check
Context
Dreamer 4 was trained in 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 imagined trajectories generated by the learned world model. The final policy was then evaluated in the actual Minecraft engine without using that evaluation experience for further training.
For the Offline Diamond Challenge, the researchers used approximately 2,541 hours of contractor gameplay containing 360p video, mouse and keyboard actions, and event annotations. Episodes started in randomly generated worlds with empty inventories, lasted up to 60 minutes, and used raw pixels with Minecraft’s native low-level control interface. Dreamer 4 obtained diamonds in 0.7% of 1,000 episodes, equivalent to seven successful episodes.
The agent was guided through a linear sequence of annotated subtasks covering resource gathering, tool crafting, mining iron, furnace use, and diamond mining. This still required long-horizon control under procedural variation, but it limits claims about open-ended planning. A 0.7% success rate establishes a non-zero reported capability, not dependable task mastery.
THE TAKEAWAY
Dreamer 4 provides strong initial evidence that a learned world model can become useful enough for imagination-based reinforcement learning to improve a policy on a difficult, long-horizon Minecraft task and transfer that improvement to the real evaluation environment. The frontier change is offline policy training for the complete diamond progression, not Minecraft play itself.
The result remains narrow: it depends on a large action-labelled dataset, substantial training, an authored task hierarchy, and a low final success rate. It does not establish a reliable general-purpose agent or a physical robot trained entirely through imagination.
Continue the Thread
World Models for Agent TrainingTracks claims about using learned world models to train or improve agents through simulated or imagined experience.
Sources
Training Agents Inside of Scalable World Models
arXiv / Google DeepMind researchers
Used for: Main Dreamer 4 claim, training procedure, dataset and evaluation setup, milestone success rates, behavioral-cloning ablations, world-model comparison, task prompt sequence, compute details, and disclosed limitations. Because the authors are reporting a result about their own system, this is a Developer / Vendor Claim rather than an Official Claim.
Dreamer 4 project page and uncut evaluations
Dreamer 4 research team
Used for: Public project description, uncut successful evaluation videos, decoded imagination-training sequences, interactive world-model comparisons, and robotics-world-model demonstrations.
Mastering diverse control tasks through world models
Nature
Used for: Previous frontier established by Dreamer 3, including online reinforcement learning from sparse rewards, Minecraft input and action accommodations, and diamond acquisition after environment interaction.
Unofficial Dreamer 4 PyTorch implementation
Nicklas Hansen
Used for: Reproducibility context. The repository explicitly describes itself as an unofficial and incomplete implementation applied to continuous-control tasks rather than a reproduction of the original Minecraft result. It must not be labelled Independent Replication.
Video PreTraining: Learning to Act by Watching Unlabeled Online Videos
NeurIPS 2022
Used for: Earlier native-interface Minecraft frontier, the estimated 24,000-action task horizon, large-scale video pretraining, and diamond-tool performance after online reinforcement-learning fine-tuning.
Last checked Methodology 2.0.0