This AI entrepreneur is developing agents that can plan ahead for the unexpected

Details

Danijar Hafner, 31, who spent years building world model architectures at Google Brain and Google DeepMind, has launched a stealth startup in San Francisco's SoMa district. The office already houses multiple humanoid robots imported from China as physical testbeds. The company has no disclosed name, funding, or commercial timeline. Hafner's technical approach is model-based reinforcement learning: build a world model that emulates physical reality, train an agent inside that simulation, and let the agent "dream" — predict forward — to navigate situations never encountered in the real world. The goal is robots that handle novel home environments — new floor plans, unfamiliar furniture — without needing real-world trial-and-error for each new location. Hafner worked at Google Brain and Google DeepMind across multiple stints in the UK, Canada, and the US, collaborating with Geoffrey Hinton and Ashish Vaswani (coauthor of "Attention Is All You Need"). His former manager Timothy Lillicrap placed him "in the top half of 1%" of researchers encountered at Google. The startup represents a continuation of the same research lineage, now aimed at embodied hardware rather than simulation benchmarks.