Robot ChatGPT: NVIDIA Unveils Cosmos World Model Platform

NVIDIA's Cosmos World Model platform offers scalable, open-source video models that generate physics-based synthetic data for AI in robotics and autonomous vehicles, democratizing physical AI.

The next frontier of AI is physics. At yesterday’s CES launch event, NVIDIA CEO Jensen Huang highlighted this theme through a platform called "Cosmos."

In simple terms, Cosmos is a world model platform that includes a series of open-source, open-weight video world models, with parameter sizes ranging from 4B to 14B.

The purpose of these models is clear: they generate large amounts of photo-realistic, physics-based synthetic data for AI systems operating in the physical world, such as robots and autonomous vehicles, to address the severe data shortage in this field.

These models are trained on 20 million hours of video data and are divided into diffusion (continuous token) and autoregressive (discrete token) models, supporting two generation methods: text-to-video and text + video-to-video.