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

Danijar Hafner, a former Google DeepMind researcher, has launched a stealth-mode startup focused on developing AI agents capable of navigating unfamiliar physical environments. By utilizing model-based reinforcement learning, Hafner creates 'world models' that allow AI to simulate physical reality and 'dream' or imagine future outcomes before taking action. This approach enables robots to handle unpredictable scenarios, such as navigating new home layouts, without requiring extensive real-world trial-and-error training. Hafner, known for his breakthroughs like the Dreamer series—which achieved human-level performance in Atari games and solved complex Minecraft challenges—is now transitioning his research into physical humanoids. His goal is to bridge the gap between virtual simulation and real-world application, potentially revolutionizing how robots interact with human spaces. The startup, currently operating from San Francisco, aims to move beyond traditional training methods to create truly autonomous agents that can plan ahead for the unexpected.