NVIDIA expands Cosmos tools for world model training and deployment

Highlights

NVIDIA introduced new Cosmos Data Factory blueprints and training tools, enabling large-scale multimodal data processing and synthetic data generation for world model training.
Porsche leverages Cosmos Data Factory to build a shared data flywheel for autonomous vehicles, aiming to reduce physical test driving and accelerate development across its brands.
Cosmos Framework expands support to open models like GR00T, Alpamayo and QWEN, allowing broad industry adoption and customizable fine-tuning for developers before deployment.

New tools for data generation and model training

As reported by NVIDIA Newsroom, the company has added new training tools, simulation checkpoints and Cosmos Data Factory blueprints to help developers build world models for systems that need to understand and simulate real-world environments.

NVIDIA says building world models starts with multimodal, multisensory data, and Cosmos Data Factory is designed to handle labeling, curation, augmentation and evaluation at scale for physical AI training. Cosmos Curator now also supports multimodal multisensory inputs and processing blocks, extending the platform's data preparation capabilities.

The new cloud-native blueprints include tools for autonomous vehicle dataset generation, image attribute augmentation, event video generation, defect image generation and video data augmentation. These workflows are aimed at producing synthetic data, including rare or unsafe scenarios, while preserving scene structure where needed.

Adoption by developers and broader industry impact

Companies are already using the new blueprints to speed up physical AI development across automotive, robotics and industrial inspection. Porsche is using Cosmos Data Factory to build a shared autonomous vehicle data flywheel across its brands, a move expected to reduce physical test driving requirements and shorten development cycles.

ANYbotics is using the platform to generate synthetic video clips for training industrial inspection robots, while Ultralytics has retrained its YOLO model with real and synthetic defect images to improve defect localization. NVIDIA has also expanded Cosmos Framework training and deployment tools to more open models, including GR00T, Alpamayo and QWEN, allowing developers to fine-tune and optimize those models before deployment.

Our earlier coverage of the debate over AI risk management highlighted NVIDIA CEO Jensen Huang’s call for deeper international cooperation, including engagement with Chinese AI companies, to help monitor potentially dangerous uses of advanced systems. We also noted proposals for U.S.-China government channels to handle serious AI incidents and Huang’s view that the UK could serve as a neutral venue for broader AI safety coordination.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.