This AI entrepreneur is developing agents that can plan ahead for the unexpected
2026-09-08 · MIT Technology Review
This AI Entrepreneur is Developing Agents That Can Plan Ahead for the Unexpected
A Stealth Startup Focused on Humanoid Robots
Danijar Hafner, 31, a 2026 Innovators Under 35 honoree, left Google DeepMind in the fall of 2025 to launch a new stealth startup in San Francisco’s SoMa district. While the startup remains unnamed and its office sparsely decorated, the space is filled with humanoid robots of various shapes and sizes hanging from racks. Imported from China, these robots represent the physical embodiment of Hafner’s longtime work: enabling AI to navigate environments it has never encountered during training. This capability is essential for deploying robots into human spaces, such as homes with unfamiliar floor plans and furniture.
Model-Based Reinforcement Learning and World Models
To achieve this adaptability, Hafner relies on model-based reinforcement learning. He develops "world models"—AI models designed to emulate physical reality—and trains agents within them. The agent treats the model as a real-world simulation, learning how to act and using those experiences to make predictions, or "dream" and "imagine," about future outcomes. This technique allows agents and their embedded robots to navigate unfamiliar situations in real life. Crucially, it enables the execution of massively complicated tasks without the traditional real-world trial-and-error training typically required in robotics.
A Standout Career at Google
Raised in a rural town in northeastern Germany by classical musician parents, Hafner learned programming from a neighbor and developed a passion for AI through online high school courses. In 2015, as an undergraduate at the Hasso Plattner Institute, he became a student researcher at Google Brain. He subsequently took on numerous roles at Google Brain and Google DeepMind across the UK, Canada, and the US, collaborating with AI legends like Geoffrey Hinton and Ashish Vaswani.
Timothy Lillicrap, Hafner’s former manager and coauthor at Google, praised him as a standout among standouts. "I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%," Lillicrap noted, adding that Hafner could single-handedly build systems that would normally take entire engineering teams to complete.
From Video Games to Physical Reality
Over the years, Hafner has proven his approach by training agents within his world models to play popular video games. His major breakthroughs include:
- PlaNet: A model that allowed agents to execute actions by planning ahead.
- Dreamer 2: The first agent to achieve human-level performance on Atari 2600 games using a world model.
- Dreamer 3: The first agent to solve the Minecraft Diamond challenge autonomously.
- Dreamer 4: Advanced the technology by learning to mine diamonds from an offline dataset of recorded gameplay videos without ever directly interacting with the game.
More recently, Hafner has migrated his agents from virtual environments into physical reality. His DayDreamer project utilized the Dreamer algorithm to enable robots to operate in novel environments and react to new experiences, such as being pushed over, without any specific prior training.
While Hafner remains coy about the exact next steps for his new startup, his ambitions are clear. "I was interested in solving a problem," he hinted, "that would change the world."