The former DeepMind researcher is moving his Dreamer world models from Minecraft into humanoids, while keeping Embo's product and financing under wraps.
By [RuntimeWire Staff](/author/runtimewire-staff)
· Published
Primary source: [MIT Technology Review](https://www.technologyreview.com/2026/09/08/1142088/danijar-hafner-developing-plan-ahead-agents/)
Why it matters #
Embo is testing whether world models can cut the real-world data and trial-and-error required to train robots, a constraint that has kept general-purpose robotics inside demos.
Danijar Hafner (@danijarh) has filled Embo's sparse San Francisco office with humanoid robots, turning a decade of research on AI agents that imagine future outcomes into a bet on machines that can function outside controlled demonstrations.
Hafner, 31, left Google DeepMind in the fall of 2025 to pursue the venture. MIT Technology Review reported on September 8th that Hafner imports the humanoids from China and is using them as the physical test bed for his work on model-based reinforcement learning. The office remains largely unfurnished and Embo is still operating in stealth.
MIT Technology Review did not name Embo. Earlier reporting by The Information and public company records identify Hafner and former DeepMind researcher Wilson Yan as Embo's co-founders. Those accounts describe Embo as a developer of world models for robotic systems, rather than a manufacturer of proprietary humanoid hardware.
Hafner arrived at the idea through an unusually direct research path. He grew up in rural northeastern Germany with two classical musicians as parents and learned to program from a neighbor. By high school, he was studying AI through online courses. "I was always fascinated with how thinking works," he told MIT Technology Review.
In 2015, while a second-year engineering student at the Hasso Plattner Institute in Potsdam, Hafner joined Google Brain as a student researcher. He went on to work across Google research groups in the UK, Canada and the US, collaborating with researchers including Geoffrey Hinton and Ashish Vaswani. Hafner later earned a computer science PhD at the University of Toronto under Jimmy Ba, completed an MRes at University College London and worked as a visiting student at UC Berkeley with Pieter Abbeel.
Timothy Lillicrap, a former manager and frequent co-author, told MIT Technology Review that Hafner ranked in the top 0.5% of the researchers he encountered at Google. Lillicrap said Hafner could independently build systems that would ordinarily require engineering teams.
From Minecraft diamonds to physical machines
Embo's central idea comes from Hafner's Dreamer research. A world model learns a compressed representation of an environment and predicts how that environment will change after an action. An agent can then practice inside that model, compare possible outcomes and choose an action before moving a physical robot.
That approach can reduce the expensive and occasionally destructive trial and error involved in robot training. A household robot cannot encounter every floor plan, chair placement or dropped object during development. Hafner's bet is that an accurate enough world model can let the robot reason through unfamiliar situations when they occur.
Hafner's early PlaNet system let agents plan ahead from visual inputs. DreamerV2 later reported human-level performance across the Atari 2600 benchmark. DreamerV3 used one configuration across more than 150 tasks and collected diamonds in Minecraft from raw pixels without demonstrations or a hand-built curriculum.
Hafner and Yan then pushed the method toward offline learning. Their Dreamer 4 project trained an agent to obtain Minecraft diamonds using recorded data, without interacting with the game during training. The task required sequences exceeding 20,000 mouse and keyboard actions. The researchers also applied the world model to robotics video and showed that it could generate counterfactual object interactions, an early indication that the method might transfer beyond games.
Hafner had already tested that move with DayDreamer, a 2022 research project conducted with Philipp Wu, Alejandro Escontrela, Ken Goldberg and Pieter Abbeel. The researchers used Dreamer on four physical robots. A quadruped learned to roll over, stand and walk in one hour, then adapted within 10 minutes after researchers pushed it. Robotic arms learned pick-and-place tasks, while a wheeled robot navigated from camera images.
Those experiments were narrow and conducted under research conditions. Embo must make the same underlying method reliable across longer tasks, different robot bodies and environments where prediction errors can compound. A robot that imagines the next second correctly can still fail if its internal forecast drifts over a minute-long sequence. Homes and workplaces also impose a safety standard that an Atari or Minecraft benchmark does not test.
A software bet in a heavily financed robot race
The imported humanoids provide a clue to Embo's strategy. Hafner can test Embo's models across existing hardware instead of spending years designing motors, joints and manufacturing systems. That positions Embo as a potential intelligence layer for multiple robot platforms, assuming Hafner and Yan can make their models generalize across different sensors and bodies.
Other robotics AI developers are chasing the same control layer with different technical language and far larger disclosed war chests. Skild AI said on January 14th, 2026 that it raised $1.4B at a valuation above $14B to build a general-purpose model capable of controlling different robot forms. FieldAI said on August 20th, 2025 that it raised $405M across two rounds for models designed to account for uncertainty and risk in industrial settings. Both figures and their deployment claims come from the companies.
Hafner is also competing with his former employer. Google DeepMind is developing Gemini Robotics models for physical machines, including the system RuntimeWire examined when Apptronik ran Gemini Robotics 2 across three Apollo 2 configurations.
Investor interest reached Embo before a public product did. The Information reported in March that Embo was discussing a seed round exceeding $100M, with Andreessen Horowitz expected to lead and Khosla Ventures, DST Global and Striker Venture Partners also in talks. That report described negotiations, not a completed financing, and no final round has been confirmed in the source material.
A nine-figure seed target would fund the compute, robotics hardware and data collection required to test Hafner's research at commercial scale. It would also price Embo against well-capitalized rivals before Hafner has presented customer deployments or company-specific benchmarks.
Hafner's record gives Embo a coherent technical thesis: agents should learn consequences inside predictive models before acting in the physical world. The humanoids in SoMa mark the point where that thesis leaves games and papers. Embo's next proof must come from robots facing situations that Hafner and Yan did not arrange for them in advance.