Danijar Hafner brings world models into robotics

Danijar Hafner’s stealth startup in San Francisco’s SoMa district is built around a long-running goal: enabling AI systems to operate in environments they have never encountered during training. The mostly empty office is filled with humanoid robots imported from China, reflecting a shift from virtual benchmarks toward machines that can function in homes and other unpredictable human spaces.

Hafner’s approach centers on model-based reinforcement learning. He builds world models that emulate physical reality, then trains agents inside those simulations so they can predict outcomes and plan actions before confronting real-world situations. The technique is designed to reduce reliance on conventional robotics training built around real-world trial and error.

His research lineage includes PlaNet, which allowed agents to plan ahead, and the Dreamer series, which reached human-level performance on Atari 2600 games, solved the Minecraft Diamond challenge and learned from recorded gameplay without direct interaction. After years at Google Brain and Google DeepMind, including work with Geoffrey Hinton and Ashish Vaswani, Hafner left Google DeepMind to form the startup in the fall of 2025.