Former DeepMind Scientist Launches Embo to Build World-Model Robotics

Danijar Hafner, a former Staff Research Scientist at Google DeepMind, has left the company to co-found Embo, a San Francisco-based startup. The venture aims to develop AI agents and humanoid robots that can plan for unexpected scenarios in unfamiliar environments.

Embo's technology leverages model-based reinforcement learning to construct "world models" that emulate physical reality. These models allow agents to effectively "dream" or imagine future outcomes, simulating potential results before executing actions in the physical world. This approach is designed to enable robots to handle complex tasks without the traditional, exhaustive real-world trial-and-error training typically required for autonomous systems.

A Pedigree in Predictive AI

Hafner's transition to entrepreneurship follows a career focused on the evolution of virtual agents into physical embodiments. He previously worked at Google Brain and Google DeepMind under prominent researchers including Geoffrey Hinton and Ashish Vaswani.

His research is anchored by the "Dreamer" series of agents. DreamerV2 achieved human-level performance in Atari 2600 games, while DreamerV3 successfully mastered the Minecraft Diamond challenge. Hafner further bridged the gap between simulation and reality with the DayDreamer project, which demonstrated that world models could allow robots to learn behaviors directly in the real world. These robots showed an ability to operate in novel environments and react to unexpected events—such as being pushed over—without having been specifically trained for those occurrences.

Breaking the Robotics Bottleneck

The ability for a robot to navigate a new physical space, such as a unique home floor plan, without requiring massive amounts of site-specific training is a critical bottleneck for the industry. Currently, the deployment of humanoid robots in consumer and industrial settings is limited by the rigidity of their training and their inability to adapt to the unpredictable nature of human environments.

If Embo successfully scales Hafner's world-model approach, it could significantly accelerate the utility and safety of autonomous physical agents. By allowing a robot to predict the consequences of its actions internally, the risk of physical failure is reduced, and the speed of deployment in diverse settings is increased.

The Path Forward

While Embo remains in stealth mode, the company's focus on foundational models for robotic perception and navigation positions it at the center of the race for general-purpose humanoid AI. The primary challenge remaining is whether these "dreaming" agents can maintain their predictive accuracy as they scale from controlled research environments to the chaotic variables of the open market. For now, the industry is watching to see if Hafner can translate his virtual successes into a commercially viable physical product.