This AI entrepreneur is developing agents that can plan ahead for the unexpected

Danijar Hafner's San Francisco office is currently empty, save for a few humanoid robots hanging from racks lining the middle of his spacious workspace. The 31-year-old entrepreneur is developing an AI-powered system that can plan ahead for unexpected scenarios. This venture is an extension of Hafner's past work on enabling AI to navigate unfamiliar environments.

Hafner's approach relies on a technique called model-based reinforcement learning, which involves creating "world models" to emulate physical reality and training agents within them. By acting within these simulated environments, the agents learn to predict future outcomes and make decisions accordingly. This method allows robots and AI agents to handle complex tasks without the need for extensive real-world trial-and-error, which has traditionally been common in robotics.

Hafner's expertise in this field was honed during his time at Google DeepMind, where he collaborated with several renowned AI researchers. His standout achievements include developing PlaNet, Dreamer 2, Dreamer 3, and Dreamer 4 – all of which utilized world models to teach agents how to act in various scenarios. For instance, Dreamer 3 successfully solved the Minecraft Diamond challenge, while Dreamer 4 learned to mine diamonds from an offline data set of recorded game-play videos.

In recent years, Hafner has shifted his focus from virtual to physical environments. His DayDreamer project demonstrated the effectiveness of his approach by enabling robots to operate in novel settings and respond to new experiences, such as being pushed over, without prior training. Hafner is currently working on bringing his innovative AI technology to market through his newly formed startup, a project he left Google DeepMind to pursue in the fall of 2025.