NVIDIA adds blueprints to Cosmos Data Factory to accelerate physical AI development

NVIDIA has expanded its Cosmos platform with new training tools, simulation checkpoints and data factory blueprints to accelerate the development of world models for physical AI systems. These updates aim to help developers simulate and understand real-world environments more effectively, according to a blog post by the chipmaker.

Root cause

The expansion addresses the need for multimodal, multisensory data in building world models. NVIDIA Cosmos Data Factory now includes agentic skills for labeling, curating and evaluating data at scale. Cosmos Curator, a GPU-accelerated solution, has added support for processing blocks and multimodal inputs to handle complex video training datasets.

Attack path

This release is not a security incident but a feature update. It provides developers with recipes to scale training data, adapt models and test behaviour before deployment. The platform ranks in the top five on six open models leaderboards, offering frontier open models alongside optimisation tools for performance tuning.

Affected versions

The updates apply to the NVIDIA Cosmos platform generally. Specific tools mentioned include Cosmos Data Factory, Cosmos Curator and the Cosmos Framework. The framework now supports open models such as GR00T, Alpamayo, QWEN and Pi0.5, along with Cosmos 3 for world model building and deployment.

Mitigation

No mitigation is required as this is a software feature release. Developers are encouraged to review the new blueprints available through GitHub, Microsoft Azure, Nebius and CoreWeave. These include autonomous vehicle dataset generation, image attribute augmentation and video data augmentation for weather and lighting changes.

What it means for organisations

Teams should assess how synthetic data generation fits their current physical AI workflows. Engineering leads need to evaluate whether new blueprints can reduce reliance on costly physical testing or manual data labeling. Procurement should review supplier contracts to see if access to these cloud-native tools is included. Security teams must ensure that synthetic data pipelines maintain integrity and that model outputs are validated against real-world benchmarks before deployment in critical infrastructure.

What to do and how to stay safe: NVIDIA

• Review the five new Cosmos Data Factory blueprints to identify which synthetic data generation tasks align with your current development bottlenecks.
• Evaluate whether your existing GPU infrastructure can support the new Cosmos Curator processing blocks for multimodal inputs.
• Check if your organisation’s current data labeling workflows can be augmented by the new agentic skills to improve efficiency.
• Monitor the performance metrics of fine-tuned models like GR00T or Alpamayo when using the new distillation and quantization recipes.

General security guidance from the TechSec Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What new blueprints are available in Cosmos Data Factory?

Five cloud-native blueprints are available: Autonomous Vehicle Dataset Generation, Image Attribute Augmentation, Event Video Generation, Defect Image Generation and Video Data Augmentation.

Which open models can developers adapt using the updated Cosmos Framework?

Developers can adapt open models including GR00T, Alpamayo, QWEN, Pi0.5 and other compatible Hugging Face models through fine-tuning and reinforcement learning.

Are there examples of companies using these new tools?

Yes, Porsche is using Cosmos Data Factory for autonomous vehicle data, ANYbotics for industrial inspection robots and Ultralytics for retraining its YOLO model with synthetic defects.