AI entrepreneur builds agents that plan for the unexpected

Danijar Hafner, an AI entrepreneur and former Google DeepMind researcher, is developing agents designed to handle unfamiliar situations by planning ahead. His new stealth startup in San Francisco focuses on robots and AI systems that can adapt to environments they were not trained on. Hafner’s approach uses modelbased reinforcement learning, which trains agents inside world models that simulate physical reality. By learning in simulation first, the agents can predict outcomes and act more effectively when they face realworld conditions, including in robotics. Over the years, Hafner has built a series of systems that demonstrated this method in games and robotics, including PlaNet, Dreamer 2, Dreamer 3, Dreamer 4, and DayDreamer. His work has moved from virtual environments toward physical robots, with the goal of enabling machines to operate in human spaces without extensive trialanderror training.