The Physical AI Frontier: NVIDIA Cosmos, Action World Models & the Sim-to-Real Paradigm

NVIDIA VP of Research Ming-Yu Liu presented NVIDIA Cosmos on Machine Learning Street Talk, describing a foundation model architecture that treats action as a first-class model token and unifies forward and inverse dynamics under an information bottleneck to simulate interactive physical worlds. Liu said Cosmos provides the foundational software substrate for embodied general intelligence, with ranking-preserved neural verification supporting the sim-to-real paradigm.

In an epoch-defining conversation on Machine Learning Street Talk, NVIDIA VP of Research Ming-Yu Liu presents NVIDIA Cosmos—a foundation model architecture transforming visual generation into interactive physical world simulation. By treating action as a first-class model token, unifying forward and inverse dynamics under an information bottleneck, and establishing ranking-preserved neural verification, Cosmos provides the foundational software substrate for embodied general intelligence.