Danijar Hafner Is Teaching Robots to Dream Before They Act

Danijar Hafner's new office in San Francisco's SoMa district doesn't have a name on the door yet. It barely has furniture. What it does have is robots, humanoids of various shapes and sizes, imported from China, hanging from racks that run down the middle of the room like a workshop for marionettes.

That's the physical tell for what Hafner, 31, is actually building: a continuation of research he's spent close to a decade refining, now aimed squarely at getting robots to function in spaces they've never seen before. Think unfamiliar floor plans, random furniture, the chaos of an actual home. For a robot to handle that without falling over or freezing up, it needs more than pattern matching. It needs something closer to foresight.

Hafner's answer is model-based reinforcement learning, a technique where an agent doesn't just learn from raw experience, it learns inside a simulated version of reality first. He builds what are called world models: AI systems trained to emulate physical dynamics closely enough that an agent can treat them as a stand-in for the real world. The agent acts inside that simulation, observes outcomes, and uses those observations to predict what happens next. Hafner sometimes calls this "dreaming" or "imagining." It's a reasonable description. The agent is running forward simulations of situations it hasn't physically encountered yet, then using those predictions to decide what to actually do when it hits the real thing.

That's a meaningfully different approach from how most robotics work has historically gone. Traditional reinforcement learning leans on massive amounts of real-world trial and error, which is slow, expensive, and occasionally destructive when you're talking about a several-thousand-dollar humanoid faceplanting on a hardwood floor. Hafner's method lets agents rehearse complicated, multi-step tasks inside the world model first, cutting down dramatically on how much physical trial-and-error is actually needed.

From Atari to Diamond Mining to Real Robots

Hafner didn't arrive at this approach overnight. He grew up in a small town in northeastern Germany, the son of two classical musicians, and learned to program from a neighbor. High school AI courses turned into an obsession. "I was always fascinated with how thinking works," he says, and AI gave him a way to try to replicate that process on a machine.

By his second year studying engineering at Hasso Plattner Institute in Potsdam, he'd landed a student researcher role at Google Brain. That kicked off a run of roughly a dozen internships and positions across Google Brain and Google DeepMind (the two later merged) in the UK, Canada, and the US. Along the way he worked alongside Geoffrey Hinton, one of AI's so-called godfathers, and Ashish Vaswani, coauthor of "Attention Is All You Need," the paper that introduced the transformer architecture underpinning today's large language models.

Timothy Lillicrap, a former manager and coauthor of Hafner's at Google, doesn't hedge when describing his work: "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 says. "In many cases he would build, single-handedly, things it would take entire teams of engineers to build."

That track record shows up clearly across Hafner's project history, each one a step up in difficulty:

• PlaNet was his first breakthrough, letting agents plan ahead and choose actions by simulating consequences before acting.
• Dreamer 2 became the first world-model-based agent to hit human-level performance on Atari 2600 games, a long-standing benchmark for reinforcement learning research.
• Dreamer 3 was the first agent to solve Minecraft's Diamond challenge, autonomously mining an in-game diamond, a notoriously hard sparse-reward problem that requires long chains of sequential actions.
• Dreamer 4 pushed further, learning to mine diamonds from an offline dataset of recorded gameplay footage, without ever directly interacting with the game environment during training.
• DayDreamer took the algorithm off the screen and into hardware, letting real robots operate in novel environments and recover from unexpected events, like getting shoved over, without task-specific training.

That progression, from Atari to Minecraft to physical robots learning from raw video, maps almost exactly onto the bet his new startup appears to be making: that world models trained well enough can generalize past the narrow scenarios they were built on, and do it in the messy, unstructured physical world rather than just inside a game engine.

What to Watch

Hafner left Google DeepMind in the fall of 2025 to start the company, and he's not saying much about it publicly yet. What he will say is that the ambition is large. "I was interested in solving a problem," he hints, "that would change the world."

Reading between the lines isn't hard. The humanoids hanging in his office, the years spent proving world models can handle increasingly unstructured problems, and his stated focus on robots navigating human environments all point toward household or service robotics as the target. The open question is execution: can a world-model approach that dominated video games and simple physical recovery tasks scale to the near-infinite variability of actual homes, with their clutter, pets, stairs, and unpredictable humans.

If Hafner's track record is any signal, betting against him seems unwise. But turning a research technique that's aced Atari and Minecraft into something that reliably works in your living room is a different order of problem, and the robotics industry has a long history of underestimating exactly that gap.