Nvidia unveils Cosmos world models and infrastructure for robotics and physical AI

Nvidia

Nvidia has introduced a new suite of world AI models, libraries, and infrastructure designed to advance robotics development. The highlight is Cosmos Reason, a 7-billion-parameter vision-language model that supports reasoning for physical AI applications and robotics.

Joining the existing Cosmos world models are Cosmos Transfer-2, which accelerates synthetic data generation from 3D simulation scenes or spatial control inputs, and a distilled version of Cosmos Transfer optimised for faster performance.

Announced at the SIGGRAPH conference, these models are intended to generate synthetic text, image, and video datasets to train robots and AI agents. Cosmos Reason stands out for its ability to “reason” by leveraging memory and understanding of physics, enabling it to act as a planning model that predicts the next steps an embodied agent might take. It is suitable for tasks such as data curation, robot planning, and video analytics.

Nvidia also revealed new neural reconstruction libraries, including one that enables developers to simulate real-world environments in 3D using sensor data. This rendering capability is being integrated into CARLA, an open-source simulation platform popular among developers. An update to Nvidia’s Omniverse software development kit was also announced.

On the hardware side, Nvidia launched new servers for robotics workflows. The RTX Pro Blackwell Server provides a unified architecture for robotics development, while Nvidia DGX Cloud offers a cloud-based management platform tailored to these workloads.

These developments align with Nvidia’s strategy to expand its AI GPU applications beyond data centres and into robotics, signalling a push toward the next major growth area.