Danijar Hafner, 31, is developing plan-ahead agents for robotics, using model-based reinforcement learning to enable robots to navigate unfamiliar environments. His startup, still in stealth mode, focuses on creating AI models that simulate physical reality to train agents for real-world tasks.
Hafner’s approach relies on world models—AI systems that emulate physical environments—to train agents in simulated settings. These agents then use their simulated experiences to predict future outcomes and navigate new situations in the real world. This method allows robots to handle untested scenarios, such as adapting to unknown floor plans in a home.
"I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%," said Timothy Lillicrap, a former manager and coauthor at Google. Lillicrap highlighted Hafner’s ability to build complex systems independently, often outperforming teams of engineers.
Hafner has tested his approach in video games, with his Dreamer 4 agent learning to mine diamonds from recorded gameplay without direct interaction. He is now transitioning these agents from virtual environments to physical robots, as seen in his DayDreamer project, which allows robots to adapt to new experiences without prior training.
The project is part of Hafner’s new startup, launched in the fall of 2025, which he left Google DeepMind to form. Though he remains coy about the startup’s specifics, he hints at a vision to solve a problem that could change the world.
Hafner did not say what specific problem he aims to solve, and he acknowledges the challenge of applying his methods to real-world robotics. His next steps remain unclear, but his work continues to push the boundaries of AI autonomy.
Source: mittr
