A former Nvidia research scientist who worked on the chipmaker's humanoid robot program has quietly started a company to build world models — the physics-simulating foundation models that a growing number of investors believe will power the next generation of robots.

Joel Jang left Nvidia, where he contributed to Project Groot, the company's own model for humanoid robots, and founded Dream Labs earlier this month. He has not yet made the venture public.

Jang has spoken with investors about raising tens of millions of dollars for the new company, according to The Information. The precise size and structure of the round have not been disclosed.

World models attempt to approximate how objects and humans interact with their physical surroundings. Proponents argue these capabilities are essential for robots that must navigate unpredictable real-world environments rather than follow scripted routines.

Dream Labs enters a field that has already attracted significant capital. Fei-Fei Li's World Labs and Yann LeCun's AMI Labs have together raised billions of dollars over the past year to pursue similar research. The pace of new entrants suggests venture investors see world models as a distinct category worth funding alongside the broader foundation-model ecosystem.

A crowded but early market

Several well-funded startups are already building foundation models for physical-world AI. Physical Intelligence has raised capital to develop general-purpose robot foundation models. Skild AI is working on a universal robot brain. Figure AI is pairing humanoid hardware with learned behaviors.

What separates world-model startups from these robotics-first companies is scope. World models aim to be general-purpose simulators of physics, potentially useful across robotics, autonomous driving, and spatial computing — not tied to a single hardware form factor.

The talent pipeline from large GPU and AI companies into startups continues to accelerate. Nvidia alumni have founded or joined multiple robotics and foundation-model ventures in the past 18 months, drawn by the chance to build products rather than research prototypes.

Jang's background on Project Groot gives Dream Labs a specific edge: direct experience training models that must understand embodied physics at scale, using Nvidia's own simulation and training infrastructure.

Dream Labs has not disclosed a product roadmap or target launch date. The company's fundraising conversations are ongoing.