Nvidia unveils new Cosmos world models, infra for robotics and physical uses

NVIDIA unveiled a series of AI models, libraries and other infrastructure for robotics developers on Monday, the most remarkable Cosmos Rede, a “reasoning” vision language model of 7 billion parameter for physical AI applications and robots.

Also joining the existing party COSMOS world models are Cosmos transfer 2, which can accelerate synthetic data generation from 3D simulation or spatial controls, and a distilled version of Cosmos transfers that is more optimized for speed.

During the announcement on the Siggraph conference on Monday, Nvidia noted that these models are intended to be used to make synthetic text, image and video datasets for training robots and AI agents.

Cosmos Rede, by Nvidia, enables robots and AI agents to ‘reason’ thanks to the memory and physics comprehension, so that it can ‘serve as a planning model to reason what steps a embodied agent could then take’. The company says it can be used for data management, robot planning and video analysis.

The company also unveiled new neural reconstruction libraries, which includes a rendering technique with which developers can simulate the real world in 3D with sensor data. This rendering option is also integrated into Open Source Simulator Carla, a popular developer platform. There is even an update for the omnifier software development kit.

There are also new servers for robotics workflows. The NVIDIA RTX Pro Blackwell Server offers a single architecture for robot development workload, while Nvidia DGX Cloud is a cloud -based management platform.

These announcements come as the semiconductor giant pushes further into robotics, because it looks at the next large use case for its AI GPUs outside Ai Data Centers.

We always want to evolve, and by giving some insight into your perspective and feedback in WAN, you can help us! Fill in this survey to let us know how we are doing.