Danijar Hafner: the researcher teaching robots to imagine the future before acting

Danijar Hafner, 31, is the subject of a profile in MIT Technology Review as part of its 2026 'Innovators Under 35' list. The article portrays him at the office of his new startup, located in San Francisco's SoMa neighborhood and still in stealth mode: no name on the door, barely any furniture and just one other person present on the day of the visit. What that space does have in abundance are humanoid robots of various sizes, hanging from frames like marionettes, imported from China. Hafner is guarded about the details of his new project, but describes it as a continuation of years of work: getting artificial intelligence to operate in environments it has never seen during training.

Those humanoids represent, according to the article, the physical evolution of that research. The logic is simple to state though complex to solve: if you want to send a robot into someone's home, that robot has to be able to handle a layout and furniture it has never encountered before. To achieve this, Hafner relies on a technique called model-based reinforcement learning. His method consists of developing 'world models', that is, AI systems designed to emulate physical reality, and training agents inside them. The agent treats that model as if it were a simulation of the real world and learns to act within it; it then uses that experience to predict -or, as Hafner would say, 'dream' or 'imagine'- future outcomes. That process is what allows agents, and the robots that incorporate them, to navigate unfamiliar situations in real life. Unlike other approaches in robotics, Hafner's technique would make it possible to carry out highly complex tasks without depending on traditional trial-and-error training in the physical world.

The profile also reviews his personal trajectory. Hafner grew up in a rural village in northeastern Germany, the son of two classical musicians, and learned to program thanks to a neighbor; in high school he began taking online courses on artificial intelligence, an interest he himself attributes to a fascination with understanding 'how thinking works'. In 2015, as a second-year engineering student at the Hasso Plattner Institute in Potsdam, he landed a position as a student researcher at Google Brain. From there he strung together a dozen internships and other posts at the company, with stints at Google Brain and Google DeepMind (the two teams would eventually merge) in the United Kingdom, Canada and the United States. Along the way he worked alongside well-known figures in the field, such as Geoffrey Hinton -usually described as one of the 'godfathers' of AI- and Ashish Vaswani, co-author of the influential paper 'Attention Is All You Need', the work that introduced the transformer architecture on which today's large language models are based.

One of his former managers and co-authors at Google, Timothy Lillicrap of Google DeepMind, describes him in the article as someone exceptional even among top-tier researchers: he says that among all the smart people he interacts with in research at Google, Hafner easily ranks in the top 0.5%, and adds that in many cases he has been able to build on his own things that would normally require entire teams of engineers.

The article also runs through the track record with which Hafner refined and demonstrated his approach, using popular video games as a testbed for agents trained inside his world models. His first major breakthrough was PlaNet, a model that allowed agents to execute actions by planning ahead. Then came Dreamer 2, the first agent to reach human-level performance playing Atari 2600 titles by means of a world model. Dreamer 3 was the first to solve the so-called 'Minecraft Diamond challenge', that is, to mine gems within the game autonomously. And Dreamer 4 went a step further: it learned to mine diamonds from an offline dataset of gameplay recordings, without ever needing to interact directly with the game.

More recently, Hafner began moving his agents from the virtual world into physical reality. His DayDreamer project used the Dreamer algorithm to let robots operate themselves in new environments and react to unforeseen experiences -such as being pushed- without having received specific training for those situations. That work connects directly with what he is now pursuing at his startup, formed after he left Google DeepMind in the fall of 2025.

About that new company, Hafner himself offers barely any clues beyond a line that closes the profile: he says he was interested in solving a problem that would change the world. The article does not reveal the company's name, its funding, its timelines or what specific product he plans to launch; it merely describes the physical setting -the China-imported humanoid robots hanging in the office- as the most tangible hint of where his ambition is headed. Taken together, the portrait casts Hafner as one of the researchers who has pushed hardest on the idea that world models, trained inside simulations and video games, can become the basis for giving physical robots the ability to improvise in the face of the unexpected, a requirement the text itself identifies as decisive if these machines are ever to operate in spaces as variable and unpredictable as a home.