Danijar Hafner is moving his work on planning agents from games into physical robots through a new, still-unnamed startup. He left Google DeepMind in fall 2025 and is now testing the approach on imported humanoid robots in San Francisco.
His method uses model-based reinforcement learning. An AI agent practices inside a “world model,” a learned simulation of an environment, and uses those imagined experiences to predict the consequences of possible actions. The goal is to reduce reliance on costly real-world trial and error when a robot encounters a room, object or event it did not see during training.
Hafner’s earlier Dreamer systems established the approach in games. Dreamer 2 reached human-level performance on Atari 2600 games, Dreamer 3 completed Minecraft’s diamond challenge, and Dreamer 4 learned that task from recorded gameplay without interacting with the game during training. His DayDreamer project later applied the algorithm to robots responding to new experiences, including being pushed over.
The startup remains in stealth, and Hafner has not disclosed a product or release schedule. The important next test is whether results from controlled games and demonstrations transfer reliably to varied human environments.