Danijar Hafner founded a stealth startup developing AI agents and humanoid robots using world models
Danijar Hafner, a former researcher at Google Brain and Google DeepMind, left Google DeepMind in autumn 2025 and founded a stealth startup in San Francisco. He is developing AI agents using model-based reinforcement learning and world models, tested on humanoid robots imported from China.
Danijar Hafner, a former researcher at Google Brain and Google DeepMind, left Google DeepMind in autumn 2025 and founded a stealth startup in San Francisco. According to the article, the company is based in an empty office in the SoMa district, where humanoid robots imported from China hang from stands; Hafner has not yet disclosed the name of the startup or further details.
Hafner has long focused on enabling AI agents to operate in environments they have never seen during training. He uses model-based reinforcement learning to do this, creating so-called world models: AI models that emulate physical reality, in which agents train as they would in a simulation and use these experiences to predict future outcomes. According to the company, this approach enables agents and robots to handle even very complex tasks without the extensive trial-and-error training common in traditional robotics. According to Hafner, the goal is to bring robots into human spaces, such as homes, where they must handle unfamiliar floor plans and furniture.
Hafner developed and tested this approach over several years of research at Google: the PlaNet model enabled agents to plan ahead, Dreamer 2 was the first world-model agent to achieve human-level performance in Atari 2600 games, Dreamer 3 was the first to solve the Minecraft Diamond challenge, and Dreamer 4 learned to mine diamonds solely from an offline dataset of gameplay recordings without direct interaction. The DayDreamer project then transferred the same principle to physical robots, which were able to adapt to new situations (e.g. being tipped over) without specific training. His former supervisor at Google DeepMind, Timothy Lillicrap, describes him as an exceptional researcher who can single-handedly create what would otherwise require an entire team of engineers.
The source provides no details about the name of the startup, its funding or a specific product.
Why it matters
The approach using world models and model-based reinforcement learning could reduce the need for extensive physical trial-and-error training, a longstanding obstacle in robotics. This is relevant to researchers and companies tracking how quickly robots can enter everyday human environments such as homes. The departure of a respected researcher from Google DeepMind to found his own company also signals growing market interest in embodied AI and humanoid robots.
Two audiences, two different impacts
What this means
For individuals
For people working with AI and robotics, this signals that model-based reinforcement learning and world models are considered a promising alternative to traditional trial-and-error training and are worth watching as a direction for further development.
More practical updates →For a business
The launch of the startup signals a new player in embodied AI and world models and the departure of an experienced researcher from Google DeepMind, which is relevant to companies tracking competition and investment trends in humanoid robotics.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.