Former DeepMind researcher launches stealth startup to build adaptive robotic agents

The short version

• Former Google DeepMind researcher Danijar Hafner has launched an unnamed stealth startup to develop AI agents capable of handling novel physical settings.[MIT Technology Review]
• The venture applies model-based reinforcement learning and world models to humanoids imported from China to avoid physical trial-and-error training.[MIT Technology Review]
• Details regarding company branding, commercial timelines, and deployment remain undisclosed as development continues in stealth mode.[MIT Technology Review]

Key facts

• Danijar Hafner left Google DeepMind in the fall of 2025 to create a stealth-mode startup.[MIT Technology Review]
• The startup uses model-based reinforcement learning to train autonomous agents without standard real-world trial-and-error robotics training.[MIT Technology Review]
• Humanoids imported from China serve as the physical testbed for the startup's artificial intelligence systems.[MIT Technology Review]
• Google researcher Timothy Lillicrap described Hafner as ranking within the top half of one percent of researchers he interacts with at the company.[MIT Technology Review]

Sources

Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.

• This AI entrepreneur is developing agents that can plan ahead for the unexpectedMIT Technology Review metered