Through simulation, Google's AI agent becomes proficient in Minecraft

Dreamer 4, the first AI to gather Minecraft diamonds using only offline data and without interacting with the game itself, was revealed by Google DeepMind researchers. Dreamer 4 learns to perform tasks in video games by training in its own mental simulation. The specifics: Dreamer 4 trains by doing more than 20,000 actions based on visual input in a predictive environment model that replicates the physics of Minecraft in real time. The training is divided into phases: watching videos to learn Minecraft, developing decision-making skills, and becoming better via practice—all without actually playing the game. With testers finishing 14/16 tasks in Dreamer 4's simulation as opposed to 5 in competing models like Oasis, the world model set new accuracy records. Dreamer also outperformed systems based on Gemma vision-language models and defeated OpenAI's Minecraft VPT agent despite learning from 100 times less data. While it's still cool to see games like Minecraft being used to test advanced agentic training and capabilities, Dreamer 4's capabilities go far beyond gaming. Learning through simulation can replace expensive and frequently risky IRL testing, creating safer and more effective development pathways for robots.