Danijar Hafner, a former researcher at Google DeepMind, has launched a stealth-mode startup in San Francisco focused on developing AI agents capable of planning ahead for unexpected scenarios. The new venture aims to bridge the gap between virtual simulation and physical reality by enabling robots to navigate environments they have never encountered before.

What Happened

Hafner left Google DeepMind in the fall of 2025 to establish the startup, which currently operates out of a sparse office in San Francisco’s SoMa district. The space is filled with humanoid robots imported from China, serving as the physical embodiment of his research. Hafner’s approach relies on model-based reinforcement learning, where he develops world models designed to emulate physical reality. Agents are trained within these simulations to predict future outcomes, allowing them to act in unfamiliar situations without the traditional need for extensive real-world trial-and-error training.

Why It Matters

This technology addresses a critical bottleneck in robotics: the ability to operate safely in human spaces with unpredictable layouts. By using world models to simulate interactions, agents can "dream" or imagine potential outcomes, enabling them to handle novel floor plans and furniture configurations. Hafner’s background includes significant milestones in this field, such as Dreamer 3, which solved the Minecraft Diamond challenge, and Dreamer 4, which learned from offline video data without direct game interaction. His recent project, DayDreamer, extended these capabilities to physical robots, allowing them to react to new experiences like being pushed over without specific training. Timothy Lillicrap, a former manager at Google, described Hafner as a standout researcher, noting that he often builds single-handedly what would require entire engineering teams.

The Bottom Line

Hafner’s startup represents a continuation of his efforts to create AI that can navigate unseen environments, moving from video game benchmarks to physical humanoid robots. While the company remains in stealth and its specific name is not yet public, the focus remains on leveraging world models to solve the problem of robotic adaptability in the real world.