Danijar Hafner Is Building AI Agents That Prepare for the Unexpected

Danijar Hafner is taking his work on AI systems that plan ahead from simulated environments into the physical world. After years of developing agents capable of mastering video games through internal models of their surroundings, the former Google DeepMind researcher is now applying the same principles to humanoid robots.

His new San Francisco startup remains in stealth mode, but its direction is clear: Hafner wants AI-powered machines to operate in places and situations they did not encounter during training. That ability could be essential for robots entering homes and other human environments, where layouts, furniture and unexpected events vary constantly.

Training agents inside world models

Hafner’s approach centers on model-based reinforcement learning. Instead of relying entirely on repeated physical experimentation, he develops “world models”—AI models intended to emulate aspects of physical reality—and trains agents within those environments.

The agent treats a world model as a simulation in which it can learn how actions may affect future outcomes. It can then use those simulated experiences to predict what might happen next and select an action accordingly. Hafner has described this process in terms of dreaming or imagining possible futures.

This differs from robotics methods that depend heavily on real-world trial and error. Hafner’s technique is designed to let agents learn complicated behavior within a model before applying what they have learned outside it.

For robots, the goal is not simply to repeat a fixed sequence. A machine entering an unfamiliar home, for example, would need to respond to a floor plan and objects it had never seen. It may also encounter disruptions that were not included in its training. Planning through a world model could help it adjust rather than stop when conditions change.

From an early interest in AI to Google

Hafner, 31, grew up in a rural town in northeastern Germany. Both of his parents were classical musicians, while he learned programming from a neighbor. During high school, he began taking online AI courses and became interested in using computers to emulate aspects of thought.

In 2015, while in his second year of undergraduate engineering studies at the Hasso Plattner Institute in Potsdam, Hafner secured a student researcher position at Google Brain. He subsequently held about a dozen internships and other roles at Google across the UK, Canada and the US, including positions with Google Brain and Google DeepMind. The two organizations later merged under the DeepMind name.

His work brought him into contact with prominent AI researchers, including Geoffrey Hinton and Ashish Vaswani, a coauthor of the paper “Attention Is All You Need,” which introduced the transformer architecture used in modern large language models.

Timothy Lillicrap, a former manager and coauthor of Hafner’s at Google, placed him “in the top half of 1%” among the researchers he has encountered. Lillicrap also said Hafner would sometimes build by himself systems that would normally require entire engineering teams.

The Dreamer series tested planning in games

Hafner demonstrated the development of his world-model approach through a succession of agents tested on video games. These projects provided controlled environments in which an agent could learn to anticipate consequences and complete increasingly difficult objectives.

Key milestones included:

PlaNet: An early breakthrough that enabled agents to plan ahead when choosing actions.
Dreamer 2: The first world-model agent to reach human-level performance on Atari 2600 games.
Dreamer 3: The first agent to complete the Minecraft Diamond challenge by independently mining in-game diamonds.
Dreamer 4: An agent that learned to mine diamonds from an offline collection of recorded gameplay videos without interacting directly with the game.

The progression from Dreamer 3 to Dreamer 4 was particularly significant for Hafner’s broader method. Dreamer 4 learned from existing recordings rather than gathering experience through direct gameplay, showing how an agent could acquire behavior from offline data.

Moving from software environments to robots

Hafner has more recently begun transferring his agents from games into physical machines. His DayDreamer project applied the Dreamer algorithm to robots, allowing them to operate in new environments and respond to experiences for which they had not received specific training.

One example involved reacting after being pushed over. Rather than requiring an individually programmed response for that event, the system was intended to adjust using the same model-based learning principles developed in virtual settings.

The humanoid robots in Hafner’s new San Francisco office represent the next stage of that work. The machines, imported from China, come in a variety of forms and sizes. They provide physical platforms for testing whether agents trained to predict outcomes can manage the uncertainty of real spaces.

The challenge is substantially different from succeeding in a game. A robot operating around people must encounter unfamiliar arrangements and unexpected physical events. Hafner’s research is focused on giving such systems a way to reason about what could happen before acting.

A stealth startup focused on physical AI

Hafner left Google DeepMind in the fall of 2025 to establish his startup in San Francisco’s SoMa district. The company has not publicly disclosed its name, and Hafner has shared few details about its products or plans.

For now, he characterizes the venture as a continuation of his long-running effort to create AI that can navigate environments outside its training experience. The startup’s collection of humanoid robots suggests that the practical application of world models will be central to its work.

Conclusion

Hafner’s research traces a consistent path: build agents that can model possible futures, validate them in complex virtual environments and then transfer those capabilities into robots. Whether the same approach can reliably handle the unpredictability of human spaces remains the central test for his new company.

Originally reported by revew.