Danijar Hafner Leaves DeepMind to Bring Dreamer AI to Humanoid Robots

Danijar Hafner, the researcher behind the Dreamer family of AI systems, has left Google DeepMind to form a stealth startup aimed at a harder setting than games: humanoid robots in unfamiliar physical environments. His bet is that robots can anticipate consequences inside a simulated world before taking an action in the real one.

Hafner formed the unnamed company in fall 2025. Its sparse San Francisco office contains humanoid robots imported from China, which he describes as the physical embodiment of work on agents that can navigate situations they did not encounter during training. The machines put a physical test bed at the center of a research program built to plan ahead.

The mechanism behind the bet

The technical foundation is model-based reinforcement learning. Rather than learning only through repeated physical attempts, an agent trains inside a world model—an AI system designed to emulate how an environment changes after an action. It can use those simulated experiences to predict possible outcomes and choose what to do next when it encounters something new.

Hafner’s approach is intended to reduce the real-world trial and error often used to train robots on complicated tasks. His stated rationale is straightforward: a household robot cannot be trained on every floor plan, furniture arrangement, or unexpected event it may meet. It would need to respond when a setting differs from its training experience.

A record in simulated worlds

Hafner established the method in virtual tasks. PlaNet enabled agents to plan actions ahead, while Dreamer 2 reached human-level performance on Atari 2600 games, according to MIT Technology Review. Dreamer 3 completed the Minecraft Diamond challenge, and Dreamer 4 learned diamond mining from recorded gameplay without directly interacting with the game during training.

He has also taken an earlier step into physical robotics. The DayDreamer project used the Dreamer algorithm to let robots operate in novel environments and react to new experiences, including being pushed over, without specific training for that event.

The harder physical test

The startup now shifts that work toward humanoid hardware and the less controlled environments Hafner has in mind. The central question is whether agents trained to imagine outcomes can stay useful when physical spaces, objects, and events differ from what they have previously encountered.