This AI entrepreneur is developing agents that can plan ahead for the unexpected
Danijar Hafner, a 31‑year‑old AI researcher, has launched a stealth‑mode startup in San Francisco’s SoMa district that is populating its sparse office with Chinese‑made humanoid robots. The venture builds on Hafner’s long‑standing work on model‑based reinforcement learning, aiming to give robots the ability to navigate environments they have never encountered during training.
Key Takeaways
- Hafner’s new company operates from a largely empty SoMa office, filled mainly with imported humanoid robots.
- The core technology is model‑based reinforcement learning that trains agents inside “world models” – AI simulations of physical reality.
- His prior research produced the PlaNet and Dreamer series, achieving human‑level Atari play, autonomous Minecraft mining, and offline video‑learning.
- The DayDreamer project demonstrated that Dreamer‑based agents can react to real‑world disturbances (e.g., being pushed) without task‑specific training.
- Timothy Lillicrap, a DeepMind senior researcher, rates Hafner among the top‑half of the top 1 % of Google researchers.
| Metric | Value | Why it matters |
|---|---|---|
| Founder | Danijar Hafner, 31 | Leads the venture with a proven AI research track record. |
| Location | San Francisco, SoMa district | Proximity to tech talent and robotics ecosystem. |
| Company status | Stealth mode, no name on the door | Indicates early‑stage development and confidentiality. |
| Spin‑off date | Fall 2025 (left DeepMind) | Marks the transition from research to commercial venture. |
What the startup is building
The company’s physical prototype fleet consists of various humanoid forms suspended from central racks, each serving as a testbed for Hafner’s “world model” agents. By training these agents inside AI‑generated simulations, the robots can “dream” future outcomes and select actions that are likely to succeed even when confronted with floor plans, furniture layouts, or obstacles they have never seen before.
Model‑based reinforcement learning explained
Traditional robotics relies on extensive real‑world trial‑and‑error, which is costly and time‑consuming. Hafner’s approach replaces that with a two‑step loop: first, a neural network learns a predictive model of physics and object dynamics; second, an agent uses the model as a sandbox, planning multiple steps ahead before executing the chosen policy on the actual robot. This “imagination‑based” planning reduces the need for physical data collection while preserving adaptability.
Track record: PlaNet and the Dreamer series
Hafner’s research milestones provide a clear lineage:
- PlaNet – introduced a world‑model that could plan actions in simulated environments.
- Dreamer 2 – reached human‑level performance on Atari 2600 games using only the world model for planning.
- Dreamer 3 – solved the Minecraft “Diamond” challenge, autonomously mining in‑game gems.
- Dreamer 4 – learned to mine diamonds from an offline dataset of gameplay videos, never interacting with the game directly.
These results demonstrate that agents can acquire complex skills without direct environment interaction, a prerequisite for safe deployment in human spaces.
From simulation to physical robots: the DayDreamer project
DayDreamer extended the Dreamer algorithm to real‑world robotics. In laboratory tests, a robot equipped with a Dreamer‑based agent recovered from being knocked over and re‑oriented itself without any task‑specific retraining. The experiment proved that world‑model planning can handle unforeseen physical disturbances, a critical capability for home assistants, warehouse pickers, or rescue bots.
Background of the founder
Hafner grew up in a rural town in northeastern Germany, the son of classical musicians. He learned programming from a neighbor and pursued AI through online courses in high school. In 2015, while a sophomore at the Hasso Plattner Institute in Potsdam, he secured a student researcher position at Google Brain. Subsequent internships and full‑time roles took him to Google Brain and DeepMind locations in the UK, Canada, and the US, where he collaborated with AI pioneers such as Geoffrey Hinton and Ashish Vaswani.
“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%,” said Timothy Lillicrap, a senior researcher at DeepMind. “In many cases he would build, single‑handedly, things it would take entire teams of engineers to build.”
Why it matters
Robots that can anticipate and adapt to novel situations are essential for safe integration into homes, offices, and public spaces. Hafner’s model‑based RL sidesteps the prohibitive data‑collection costs of conventional robotics, potentially accelerating the timeline for commercially viable autonomous assistants. If the startup can scale its simulation‑first pipeline, it could reshape how the industry approaches robot learning, shifting the focus from exhaustive real‑world testing to rapid, imagination‑driven iteration.
Frequently asked questions
Who is behind the new startup?
The company was founded by Danijar Hafner, a 31‑year‑old AI researcher with a history of work at Google Brain and DeepMind.
What technology powers the robots?
The core is model‑based reinforcement learning, where agents train inside AI‑generated world models that simulate physical reality, allowing them to plan ahead and react to unforeseen scenarios.
How does this differ from traditional robot training?
Traditional methods rely on extensive real‑world trial‑and‑error. Hafner’s approach trains agents entirely in simulation, reducing physical wear‑and‑tear and data‑collection costs while still enabling real‑world adaptability.
What are the notable achievements of Hafner’s research?
His work produced PlaNet, Dreamer 2 (human‑level Atari), Dreamer 3 (Minecraft diamond mining), Dreamer 4 (offline video learning), and the DayDreamer project that demonstrated real‑world disturbance recovery.
When did the founder leave DeepMind to start this venture?
Hafner departed DeepMind in the fall of 2025 to focus on the stealth startup.
Bottom Line
Danijar Hafner’s stealth venture is translating a proven model‑based reinforcement learning framework into physical robots capable of planning for the unexpected. Backed by a track record of world‑model breakthroughs and praised by DeepMind leadership, the company could accelerate the deployment of adaptable, safe robots in everyday environments.
We will watch for the startup’s first public demonstrations and any partnership announcements that could signal a move from lab‑scale prototypes to market‑ready products.
Reporting based on: original source.
Comments
Post a Comment