Danijar Hafner builds AI agents for unfamiliar environments
Danijar Hafner is developing a stealth startup in San Francisco focused on helping artificial-intelligence systems operate in environments they did not encounter during training. His mostly empty office contains humanoid robots imported from China, which he describes as the physical continuation of his research. The company’s name and detailed plans have not been disclosed.
Hafner uses model-based reinforcement learning. His world models emulate physical reality, allowing agents to practice inside simulations, predict possible outcomes and select actions before facing unfamiliar situations. He says this could help robots function in human spaces, where homes and furniture may differ from anything in their training data. The approach is intended to reduce reliance on extensive real-world trial and error.
Hafner previously worked at Google Brain and Google DeepMind after joining Google as a student researcher in 2015. His projects include PlaNet and the Dreamer series. Dreamer 2 reached human-level performance on Atari 2600 games, Dreamer 3 solved Minecraft’s Diamond challenge, and Dreamer 4 learned from recorded gameplay without directly interacting with the game. His DayDreamer project applied the method to robots in new environments. He left Google DeepMind to establish the startup in fall 2025.
Written by Sakhanda Wire AI from reporting by MIT Tech Review. Facts, quotes and full context are in the original article below.