NVIDIA Cosmos

2025-01-06 · Publisher : NVIDIA · Domain : Robotics

At CES 2025, NVIDIA launches Cosmos, a platform of world foundation models (WFMs), video tokenizers and data processing, to generate physically plausible video for robots and autonomous vehicles.

What changes

World models generate realistic video from text, images or video.
Models are released under NVIDIA's open model license.
1X, Agility, Figure AI, Uber and XPENG are among early adopters.
Goal: mass-produce synthetic training data.

### Background On January 6, 2025, at CES in Las Vegas, NVIDIA announced Cosmos, a platform built for "physical AI". It bundles world foundation models (WFMs), video tokenizers and an accelerated data processing pipeline, optimized for NVIDIA GPUs. ### Specifications - **World models**: physics-based video generation from text, images or video, with attention to object permanence and to industrial (warehouses, factories) and driving environments. - **License**: models released under NVIDIA's open model license, downloadable from the NGC catalog and Hugging Face. - **Highlighted uses**: searching video for rare scenarios, synthetic data generation, fine-tuning specialized models. - **Early adopters**: 1X, Agile Robots, Agility, Figure AI, Foretellix, Uber, Waabi and XPENG, according to NVIDIA. ### Why it matters Robots lack data. Cosmos proposes to manufacture it: instead of recording thousands of real-world hours, realistic variations of a scene are generated. Jensen Huang presented Cosmos as a way to "democratize physical AI". ### Key takeaway Cosmos became the "data" layer of NVIDIA's robotics stack, later split into Predict, Transfer and Reason, then versions 2.5 and 3.

NVIDIA Cosmos

What changes

Links