Danijar Hafner’s Dreamer AI Powers a Stealth Humanoid Robotics Startup
Danijar Hafner’s stealth startup trains humanoid robots in Dreamer world models to plan ahead and adapt to unfamiliar homes without real-world trial and error.
Summary
As of September 8, 2026, Danijar Hafner, 31, is building an unnamed stealth startup in San Francisco’s SoMa after leaving Google DeepMind in fall 2025. Humanoids imported from China embody his goal: robots that can enter unfamiliar human spaces and cope with unseen layouts, furniture and disruptions. His model-based reinforcement learning trains agents inside world models that emulate physical reality, then uses simulated experience to predict outcomes and plan massively complicated tasks without traditional real-world trial and error.
Hafner grew up in rural northeastern Germany with two classical-musician parents, learned programming from a neighbor and took online AI courses in high school. In 2015, during his second undergraduate year in engineering at Hasso Plattner Institute in Potsdam, he became a Google Brain student researcher. He later held 12 internships and other Google positions across the UK, Canada and US, spanning Google Brain and Google DeepMind before their merger under DeepMind, and worked with Geoffrey Hinton and “Attention Is All You Need” coauthor Ashish Vaswani. Former manager and coauthor Timothy Lillicrap places him in Google research’s top half of 1%, saying he built alone what entire engineering teams might build.
PlaNet first let agents act by planning ahead; Dreamer 2 became the first world-model agent to reach human-level Atari 2600 performance; Dreamer 3 first solved Minecraft’s Diamond challenge autonomously; Dreamer 4 learned diamond mining solely from recorded gameplay, never interacting with the game. DayDreamer moved the algorithm into robots, enabling operation in novel environments and responses to untrained events such as being pushed over. Hafner has not disclosed the startup’s product or next steps, beyond targeting a problem he believes could change the world.
Positives
- Dreamer 4 learned to mine Minecraft diamonds solely from recorded gameplay, without interacting with the game.
- DayDreamer enabled robots to operate in novel environments and respond to being pushed over without specific training.
- Dreamer 2 achieved human-level Atari 2600 performance using a world model, a first for this approach.
- Model-based reinforcement learning could reduce the costly real-world trial and error traditionally required to train robots.
- Hafner left Google DeepMind in fall 2025 to commercialize his research through a new humanoid robotics startup.
Risks & concerns
- Hafner’s startup remains unnamed, with no disclosed product, customer, commercialization timetable or next step.
- Robots entering homes must still cope reliably with floor plans, furniture and disruptions absent from their training.
- Dreamer’s flagship demonstrations center on Atari 2600 and Minecraft, while physical evidence is limited to DayDreamer examples.
