31-year-old Danijar Hafner is running a stealth San Francisco startup that pairs imported humanoids with AI agents designed to handle situations they were never explicitly trained on.
His approach uses model-based reinforcement learning: world models simulate physical reality, letting agents plan, "dream" and predict outcomes before acting in unfamiliar real-world settings.
That method aims to cut the trial-and-error burden common in robotics, a key hurdle for deploying robots in homes and other human spaces where layouts and obstacles constantly vary.
Hafner built the approach through a string of systems including PlaNet and Dreamer 2, 3 and 4, then extended it to physical robots in DayDreamer before leaving Google DeepMind in fall 2025.
Former DeepMind manager Timothy Lillicrap called Hafner "top half of 1%" among Google researchers, underscoring why his new venture is drawing attention despite remaining in stealth.
How will Danijar Hafner's stealth startup overcome the compounding errors that cause AI world models to hallucinate in real-world robotics?
Can Hafner's world models outpace heavily funded giants like World Labs in the race to bring spatial intelligence to physical robots?
If simulated robots drift into over-optimistic predictions, how can we trust these AI-driven machines in unpredictable human homes without catastrophic failures?