# Former DeepMind researcher Danijar Hafner launches stealth startup building AI agents for real-world adaptability > Source: > Published: 2026-09-08 10:43:11+00:00 # Former DeepMind researcher Danijar Hafner launches stealth startup building AI agents for real-world adaptability The creator of DreamerV3, the first AI to autonomously mine diamonds in Minecraft, is now putting world models inside humanoid robots Danijar Hafner spent nearly a decade teaching AI systems to navigate worlds they’d never seen before. Now the 31-year-old researcher is betting his career that those same principles can make physical robots useful in unpredictable real-world environments. Hafner, who left [Google](https://cryptobriefing.com/markets/alphabet/) DeepMind on November 3, 2025, has set up shop in San Francisco’s SoMa district with a startup so early-stage it doesn’t even have its name on the door. What it does have: racks of humanoid robots imported from China, hanging like marionettes down the center of an otherwise empty office. ## From Minecraft diamonds to physical robots If you’ve followed AI research over the past few years, you’ve probably encountered Hafner’s work even if you didn’t know his name. His Dreamer series of AI agents represents some of the most significant advances in what researchers call “world models,” systems that let AI build internal simulations of their environment and use those simulations to plan actions. The marquee achievement was DreamerV3, published in *Nature*, which became the first AI agent to autonomously learn how to mine diamonds in Minecraft. No human data. No step-by-step instructions. The system figured out the game’s complex, multi-stage crafting chain entirely on its own. Hafner followed that up with Dreamer 4, released in September 2025, which extended the framework to learn offline from large datasets rather than requiring active interaction with an environment. ## The “brains” bottleneck in humanoid robotics His earlier project, DayDreamer, demonstrated that physical robots could learn efficient control policies end-to-end without relying on simulation environments. Most robotics AI pipelines depend heavily on simulated training, where a digital twin of the robot practices millions of times in a virtual world before transferring its learned behavior to the physical machine. The gap between simulation and reality, often called the “sim-to-real” problem, remains one of the field’s most stubborn obstacles. DayDreamer sidestepped this entirely, training robots directly in the real world by building compact world models that let the agent “imagine” outcomes before committing to actions. The startup’s details remain thin: no public name, no disclosed funding, no announced team beyond Hafner himself and at least one other person. As of early 2026, Hafner’s online portfolio had not been updated to reflect his new role and still listed him at DeepMind. **Disclosure:** This article was edited by Editorial Team. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).