Nvidia Debuts Cosmos Reason Model and New Infrastructure for Physical AI
Nvidia introduces Cosmos Reason, a 7-billion-parameter vision language model designed to help robots plan and understand physics. The company also reveals new synthetic data tools, neural reconstruction libraries, and dedicated hardware to advance physical AI applications.
Nvidia unveils a suite of new world AI models and infrastructure tailored specifically for robotics developers at the SIGGRAPH conference. The highlight of this release is Cosmos Reason, a 7-billion-parameter vision language model that enables robots and AI agents to reason through tasks using built-in memory and physics understanding. This model acts as a planning system that predicts the next steps an embodied agent should take in the real world.
In addition to Cosmos Reason, Nvidia introduces Cosmos Transfer-2 and a faster, distilled version of the model to accelerate the generation of synthetic text, image, and video training data from 3D simulations. The company also releases new neural reconstruction libraries that allow developers to render highly accurate 3D simulations directly from real-world sensor data, with this rendering capability integrating directly into the popular open-source CARLA simulator platform.
To support these advanced software tools, Nvidia expands its hardware and cloud offerings with the RTX Pro Blackwell Server for unified robotic development workloads and the DGX Cloud management platform. These comprehensive updates highlight Nvidia's strategic push beyond traditional AI data centers as it actively pursues robotics and physical AI as the next major market for its powerful GPU technology.