Danijar Hafner, the 31-year-old researcher behind Google DeepMind's Dreamer agent series, left the lab in the fall of 2025 to launch a stealth robotics startup in San Francisco's SoMa district. The company has no name on the door and one employee besides Hafner on the day of a recent visit, but racks of humanoid robots imported from China hang down the center of the office. Hafner is betting that agents trained inside AI-generated simulations of the physical world can control those robots in environments they have never seen before.
The technical approach is model-based reinforcement learning. Hafner builds world models — neural networks designed to emulate physical reality — and trains agents inside them. The agent treats the model as a real-world simulation, learns to act, and uses those learned dynamics to predict future outcomes. That predictive step is what allows a robot to handle a floor plan, a piece of furniture, or a shove it has never encountered during training.
The commercial pitch is straightforward: if a humanoid is going to enter a stranger's home, real-world trial-and-error training does not scale. Every home is different, and the cost of collecting physical training data across millions of environments is prohibitive. World models sidestep the problem by generating the environments in silico.
Key facts
- 01Danijar Hafner left Google DeepMind in the fall of 2025 to start a stealth robotics company in San Francisco's SoMa district.
- 02Hafner, 31, created the Dreamer agent series, including Dreamer 3, the first agent to solve the Minecraft Diamond challenge.
- 03His startup imports humanoid robots from China and trains them inside AI world models rather than through physical trial and error.
- 04Hafner joined Google Brain as a student researcher in 2015 during his second undergraduate year and held a dozen roles across Google Brain and DeepMind.
Hafner grew up in a rural town in northeastern Germany, the son of two classical musicians, and learned programming from a neighbor before taking online AI courses in high school.
“I was always fascinated with how thinking works”— Danijar Hafner, Founder and former Google DeepMind researcher
In 2015, as a second-year undergraduate at the Hasso Plattner Institute in Potsdam, he won a student researcher role at Google Brain. That opened a decade-long run of roughly a dozen internships and positions across Google Brain and Google DeepMind in the UK, Canada, and the US, including collaborations with Geoffrey Hinton and Ashish Vaswani, a co-author of the 2017 transformer paper that underpins today's large language models.
Timothy Lillicrap, a former manager and co-author at Google DeepMind, ranks Hafner in the top half of 1% of researchers he has worked with.
“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%.”— Timothy Lillicrap, Google DeepMind researcher
The Dreamer track record is what makes the startup worth watching. Hafner's PlaNet was an early world-model agent that planned ahead through learned dynamics. Dreamer 2 was the first world-model agent to reach human-level performance on Atari 2600 games. Dreamer 3 became the first agent to solve the Minecraft Diamond challenge, mining in-game diamonds without hand-coded instructions. Dreamer 4 went further by learning to mine diamonds purely from an offline dataset of recorded gameplay, never interacting with the live game.
The DayDreamer project moved the same algorithm off the screen and onto physical robots, which learned to operate in new environments and recover from disturbances such as being pushed over — without task-specific training.
Hafner will not describe the new company's product, customers, or funding. He calls it a continuation of his work on agents that navigate untrained environments, and the humanoids hanging in the office are the physical embodiment of that thesis. Whether the startup ships a foundation model for robotics, a full humanoid platform, or licensed software to existing hardware makers is unclear.
The competitive frame is crowded. Well-funded humanoid programs at Figure, 1X, Tesla, and Physical Intelligence are all racing to put general-purpose robots into homes and warehouses, and most are pouring capital into large-scale teleoperation data collection — precisely the trial-and-error approach Hafner is trying to skip. If world models can substitute simulated experience for real-world demonstrations, the training economics shift meaningfully in his favor. If they cannot bridge the sim-to-real gap reliably, another year of humanoid demos ends in another year of humanoids that cannot leave the lab.
The bet Hafner is making is that the bottleneck in robotics is not hardware or compute but the environment model itself — and that a researcher who has spent a decade teaching agents to dream is the one to build it. It is a narrow bet with a wide payoff, and it is the kind of founder-market fit that draws capital quickly once stealth ends. Expect the name on the door within twelve months, and expect the first product demo to look less like a chatbot and more like a robot walking into a room it has never seen.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
