The AI researcher who taught robots to imagine, and just left Google DeepMind to do it full-time
Danijar Hafner spent years training AI agents to plan for situations they have never seen. Now he has a stealth startup, a room full of humanoid robots, and an idea he thinks could change everything.
Key points
- Danijar Hafner, 31, left Google DeepMind in autumn 2025 to found a stealth robotics startup in San Francisco.
- His technique, called model-based reinforcement learning, lets AI agents plan ahead in situations they have never trained on, without real-world trial and error.
- His Dreamer 2 agent was the first to reach human-level performance on Atari 2600 games using a world model; Dreamer 3 solved Minecraft's diamond-mining challenge on its own.
- Hafner imports humanoid robots from China and is testing whether his AI approach can let them handle homes, offices, and other spaces they have never entered before.
Step inside Danijar Hafner's office in San Francisco's SoMa neighbourhood and the first thing you notice is the robots. Humanoid figures of various shapes hang from racks running down the centre of the room like marionettes waiting for the curtain to rise. The space has almost no furniture. There is no company name on the door. His startup is still in stealth mode, meaning it has not yet gone public with its name, product, or funding.
What Hafner will say is this: he is building on a decade of work to let AI agents handle environments they have never encountered before.
What is a world model, and why does it matter for robots?
A world model is an AI program trained to mimic how physical reality works, so that a software agent, a program that can carry out tasks on its own, can practise inside it like a pilot in a flight simulator.
Hafner's agents learn by exploring these simulated worlds, building up predictions about what comes next. That ability to "imagine" future outcomes is what lets a robot walk into a house it has never visited and still figure out how to move around the furniture. Traditional robotics training needs thousands of real-world attempts. Hafner's approach cuts that out almost entirely.
His former manager at Google, Timothy Lillicrap, puts it plainly: "He easily sits in the top half of 1%" of researchers Lillicrap has worked with, as MIT Technology Review first reported. "In many cases he would build, single-handedly, things it would take entire teams of engineers to build."
How did he prove the idea works?
Hafner tested his agents on video games, which provide a clean, measurable challenge.
| Project | What it did | Why it matters |
|---|---|---|
| PlaNet | Planned actions several steps ahead | First proof the approach scaled |
| Dreamer 2 | Hit human-level Atari 2600 scores | First world-model agent to match humans on classic games |
| Dreamer 3 | Mined diamonds in Minecraft alone | Solved a long-standing AI benchmark |
| Dreamer 4 | Learned Minecraft from recorded video, no live play | Proved agents can learn from observation, not just experience |
| DayDreamer | Controlled physical robots in new environments | Moved the method out of games and into the real world |
DayDreamer is the step that matters most for his startup. It showed that a robot guided by a Dreamer-style agent could recover from being pushed over, with no specific training for that situation.
What does this mean for ordinary people?
The honest answer is: not yet very much, but watch this space. If Hafner's approach works at scale, robots that can handle unpredictable real-world spaces become far cheaper and faster to deploy. A care robot that can enter any patient's home without months of custom training would be genuinely useful. A warehouse bot that adapts to a new floor layout overnight would save companies serious money.
None of that is on sale today. Hafner is 31, his startup has no public name, and the road from a room full of hanging humanoids to a product in someone's home is long.
But the underlying research is real, peer-reviewed, and already working in demos.
The one honest takeaway: if your job involves repetitive physical tasks in a fixed, predictable space, pay attention to how fast this field moves. If your work requires adapting to new situations every day, you have more breathing room than the headlines suggest.
