Danijar Hafner, 31, has founded a new, still-stealth startup whose San Francisco SoMa office is sparsely furnished and lacks an exterior sign. When visited, only one other person was present. The most conspicuous items in the wide-open space are humanoid robots of various shapes and sizes suspended on racks down the center. Hafner imports these humanoids from China.
The goal: robots that handle unexpected situations
Hafner describes the new venture as a continuation of his long-term research to enable AI to function in environments it has not seen during training. The humanoid robots are the physical embodiment of that work: their ability to react to previously untested scenarios is central to deploying robots in human spaces. For example, a robot sent into a person’s home must cope with floor plans and furniture configurations it has never encountered.
Method: model-based reinforcement learning and world models
Hafner’s approach is based on model-based reinforcement learning. He builds "world models"—AI models intended to emulate physical reality—and trains agents inside those models. The agent treats the model as a real-world simulation and learns actions there; it then uses those learned experiences to predict future outcomes (Hafner might call this dreaming or imagining). This lets agents and the robots that contain them act robustly in unfamiliar real-world situations.
This technique contrasts with conventional robotics training, which often requires extensive real-world trial-and-error. Hafner’s method aims to enable agents to perform highly complex tasks without needing to physically rehearse every scenario.
Career and earlier breakthroughs
Hafner grew up in a rural town in northeastern Germany; both of his parents were classical musicians. He learned programming from a neighbor and began taking online AI courses in high school, which developed into a passion. "I was always fascinated with how thinking works," he has said; AI offered a way to emulate thinking on a computer.
In 2015, as a second-year undergraduate engineering student at the Hasso Plattner Institute in Potsdam, he won a student researcher role at Google Brain. He went on to about a dozen internships and roles across Google teams, including Google Brain and Google DeepMind (since merged under DeepMind), in the UK, Canada, and the US. He collaborated with figures such as Geoffrey Hinton and Ashish Vaswani.
Timothy Lillicrap, a former manager and coauthor at Google, praises Hafner as exceptional: "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%." Lillicrap added that Hafner often single-handedly built systems that would otherwise require entire engineering teams.
From PlaNet to Dreamer and the move into the physical world
Hafner refined and validated his approach by training agents inside world models and testing them on popular video games. His first notable project, PlaNet, allowed agents to act by planning ahead. Dreamer 2 was the first agent using a world model to reach human-level performance on Atari 2600 games. Dreamer 3 became the first to solve the Minecraft Diamond challenge—successfully mining in-game gems on its own. Dreamer 4 advanced further by learning to mine diamonds from an offline dataset of recorded gameplay videos, without directly interacting with the game.
More recently, Hafner has begun migrating agents out of virtual environments into the physical world. His DayDreamer project used the Dreamer algorithm to enable robots to operate in novel settings and respond to new experiences (for example, being pushed over) without being specifically trained for those events.
Leaving DeepMind and looking ahead
Hafner left Google DeepMind in the fall of 2025 to form his new startup. He remains reticent about specifics—he won’t disclose the company’s name or detailed product plans—but signals ambitious aims, saying he is interested in solving a problem "that would change the world."
