Danijar Hafner's stealth startup aims to let robots navigate the unknown

Danijar Hafner's career has been marked by a progression from virtual to physical worlds. After early work at Google Brain and Google DeepMind, he developed a series of increasingly capable agents, from PlaNet to Dreamer 4, each trained within world models to plan ahead. His recent DayDreamer project proved these agents could control robots in novel environments, setting the stage for his current venture.

His new startup, formed in the fall of 2025, represents a culmination of this research. The focus on humanoid robots, which he imports from China, signals a move toward physical embodiment. The core technical claim is that his agents do not need real-world trial-and-error training, a departure from conventional robotics that relies on such methods. Instead, they use experiences from simulated world models to anticipate and react to unseen situations.

The stakes are high for the field of embodied AI. If Hafner's robots can successfully navigate the messiness of human spaces, it could accelerate the deployment of robots in areas like home assistance. However, the startup is still in stealth mode, and he has not named specific products. His ambition to 'change the world' suggests he is aiming for a significant impact, but the immediate outcome of this venture remains to be seen.