Danijar Hafner, a 31-year-old AI entrepreneur, is developing a stealth-mode startup in San Francisco’s SoMa district aimed at enabling AI agents to navigate unfamiliar environments. The startup’s office, while sparsely furnished, is equipped with various humanoids imported from China, which serve as the physical embodiment of Hafner’s long-term research focus. Hafner’s approach centers on model-based reinforcement learning, where AI models, or “world models,” emulate physical reality, allowing agents to learn and make predictions about future outcomes within these simulated environments. This capability is crucial for deploying robots in human spaces, such as private homes, where they must handle unseen floor plans and furniture configurations.

Hafner’s technique distinguishes itself by enabling AI agents and the robots they control to execute complex tasks without the traditional trial-and-error training in the real world. His research career, marked by significant contributions to the field, includes developing agents like PlaNet, Dreamer 2, Dreamer 3, and Dreamer 4, which have achieved human-level performance in video games and solved challenges like the Minecraft Diamond task. More recently, Hafner has transitioned his work from virtual to physical realms with the DayDreamer project, which allows robots to operate autonomously in novel settings and respond to new experiences without specific training. Hafner left Google DeepMind in the fall of 2025 to found his new venture, with ambitions to address a problem of global significance. Source

This is a summary. Read the original at technologyreview.com.