NVIDIA expanded its Cosmos platform on Oct. 22 with data-generation blueprints, model-training recipes and simulation tools for developers of robots and autonomous vehicles.

The additions are intended to reduce reliance on physical data collection and testing, particularly for rare or hazardous events. NVIDIA said Porsche expects the system to cut repeated test drives by up to 90%; the company’s announcement provides no independent test results.

Cosmos is NVIDIA’s platform for building world models, which the company describes as systems that help physical AI understand and simulate the real world. Training those systems requires varied video, sensor and other data, while some driving or industrial incidents are costly or unsafe to reproduce. NVIDIA says the expanded platform connects data preparation, model adaptation and simulation before deployment.

The five new Cosmos Data Factory blueprints cover autonomous-vehicle datasets, image-attribute changes, synthetic event videos, industrial defects and video variations such as weather and lighting. NVIDIA says the blueprints are available through GitHub, Microsoft Azure, Nebius and CoreWeave1. The GitHub repository2 says workflows can run on a company’s own GPUs or on its cloud provider’s infrastructure; GPU requirements vary by blueprint.

NVIDIA said Porsche is using Data Factory through Voxel51 on Microsoft Azure to develop shared autonomous-vehicle data workflows across Volkswagen Group brands. The company expects the work to shorten development cycles by up to 85%, as well as reduce repeated test drives. NVIDIA did not describe the work as a production deployment.

ANYbotics generated more than 1,500 synthetic video clips to train industrial-inspection robots to recognize liquid spills and other changes between facility visits, NVIDIA said. Ultralytics retrained its YOLO model using real images and synthetic defects generated through the platform; NVIDIA said the work improved defect localization and recall. The announcement gives no independent evaluation of those results.

NVIDIA also extended Cosmos Framework recipes for fine-tuning, reinforcement learning, distillation and quantization to open models including GR00T, Alpamayo, QWEN and Pi0.5. Developers can adapt compatible models hosted on Hugging Face, according to the company, and deploy supported Cosmos models through its NIM microservices.

New Cosmos 3 Nano variants are designed for simulation, where developers can test autonomous-vehicle and bimanual-robot policies in repeatable, multiview and multisensor scenarios that are difficult to capture in the real world. NVIDIA said accompanying code lets developers adapt the models to train and evaluate policies in interactive simulations before physical testing.

Updated Cosmos 3 NIM microservices deliver up to twice as fast inference with no quality loss in configurations NVIDIA validated, the company said. Its announcement does not include pricing.