Ex-DeepMind Researcher Launches Stealth Humanoid AI Startup — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Danijar Hafner, 31, founded a stealth startup in San Francisco's SoMa district after leaving Google DeepMind in fall 2025, importing humanoid robots from China to test agents that can handle unfamiliar real-world scenarios.
- Hafner pioneered model-based reinforcement learning, developing 'world models' that emulate physical reality so agents can learn and plan ahead by treating the model as a simulation rather than relying on costly real-world trial and error.
- His Dreamer project lineage includes PlaNet (planning ahead), Dreamer 2 (first agent to reach human-level Atari 2600 play via a world model), Dreamer 3 (first to solve Minecraft's Diamond challenge), and Dreamer 4 (mined diamonds from offline gameplay video without direct game interaction).
- Hafner's DayDreamer project migrated the Dreamer algorithm into physical robots, enabling them to operate in novel environments and recover from unexpected events like being pushed over without task-specific training.
- He began at Google Brain in 2015 as a second-year undergraduate at Hasso Plattner Institute in Potsdam, later working across Google Brain and DeepMind in the UK, Canada, and the US alongside Geoffrey Hinton and transformer coauthor Ashish Vaswani.
- Google DeepMind's Timothy Lillicrap called Hafner 'top half of 1%' among researchers, noting he 'would build, single-handedly, things it would take entire teams of engineers to build.'
- Hafner described his new venture as 'a problem that would change the world,' though he declined to name the company or disclose its funding or launch timeline.
Why it matters: If Hafner's world-model approach scales to physical humanoids, it could cut the enormous real-world training costs that have bottlenecked robotics deployments, letting robots enter unstructured spaces like homes. The startup puts a former DeepMind principal researcher back into the AI race with a technique that already produced firsts in Atari and Minecraft—benchmarks DeepMind itself pursued.
Ask SkimNews
