Danijar Hafner’s stealth startup targets AI agents that can plan for the unfamiliar
The former Google DeepMind researcher is applying world models and model-based reinforcement learning to humanoid robots and other agents.
AI researcher and entrepreneur Danijar Hafner is developing a stealth startup focused on helping AI agents operate in unfamiliar real-world environments. The venture is exploring humanoid-robot applications, but its name, funding, product, customers and commercial timetable have not been disclosed.
According to MIT Technology Review, Hafner’s approach uses model-based reinforcement learning. His “world models” are designed to emulate physical reality, allowing an agent to simulate possible outcomes and plan actions before acting.
The method is intended to reduce reliance on extensive real-world trial and error. For humanoid robots, that could mean responding to unfamiliar layouts, furniture and other situations without specific training for every environment. The available information does not establish how capable or reliable the robots are in real-world settings.
Hafner left Google DeepMind in autumn 2025 to form the unnamed company. His earlier work included PlaNet, Dreamer 2, Dreamer 3, Dreamer 4 and DayDreamer, projects that applied world models to video games and robots.
Reported milestones include human-level performance in Atari games, autonomous Minecraft diamond mining and robots adapting to novel situations. These claims have not been independently verified in the available material.
Hafner has described the new venture as an effort to solve a problem that could “change the world”, but has not specified its product or plans. For now, the company remains an early-stage project operating in stealth.
