Ex-Google DeepMind Researcher Danijar Hafner Builds Robots With World Models

AI researcher Danijar Hafner, 31, has left Google DeepMind to start a new company in San Francisco. Founded in the fall of 2025, the venture is still in stealth mode with no public name yet, and it is building humanoid robots, imported from China, around technology Hafner calls world models.

A world model is an AI system built to emulate physical reality, letting an AI agent learn inside a simulation before acting in the real world. Hafner trains his robots using model-based reinforcement learning, a method where a program improves by trying actions and learning from their outcomes, here inside the simulated world model instead of through real-world attempts alone. The agent uses what it learns in simulation to predict outcomes, which Hafner describes as letting it dream, so robots can handle unfamiliar situations, such as a home's floor plan and furniture they have never seen, without being specifically trained for them.

Hafner developed this approach through projects that pitted world-model-trained agents against video games. His first breakthrough, PlaNet, let agents plan ahead by choosing actions based on predictions from a learned model. Dreamer 2 then became the first agent to reach human-level performance playing Atari 2600 games using a world model. Dreamer 3 was the first to solve Minecraft's Diamond challenge, mining in-game diamonds entirely on its own, and Dreamer 4 went further, learning to mine diamonds from an offline dataset of recorded gameplay videos without ever directly interacting with the game.

Hafner then carried the approach into physical machines with a project called DayDreamer, which used the Dreamer algorithm to let real robots operate in new environments and react to experiences such as being pushed over, without any task-specific training.

Timothy Lillicrap, one of Hafner's former managers and coauthors at Google, has called him a standout among the researchers he works with, saying Hafner "easily sits in the top half of 1%."