NVIDIA has announced major updates to its Cosmos platform to streamline the development of physical AI world models. The expansion introduces new training tools, simulation checkpoints, and cloud-native Cosmos Data Factory blueprints, which are now available across platforms including GitHub, Microsoft Azure, Nebius, and CoreWeave. Designed to process multimodal, multisensory inputs, the platform helps developers curate training data, adapt pretrained architectures, and optimize models prior to physical deployment.
Several enterprise partners and research labs are actively integrating the new tools. Porsche is leveraging Cosmos Data Factory via Voxel51 on Azure to build an autonomous vehicle data flywheel across Volkswagen Group brands, aiming to cut physical test drives by up to 90%. Additionally, ANYbotics generated over 1,500 synthetic video clips to train industrial inspection robots, while computer vision firm Ultralytics retrained its YOLO model to enhance defect localization. At GTC Berlin, NVIDIA further extended Cosmos training and deployment tools as agent skills to open models like GR00T and Pi0.5.
Why it matters
Significantly reduces real-world data collection costs by replacing up to 90% of physical test drives with synthetic and curated video data.
Provides flexible deployment pipelines by supporting fine-tuning and distillation of open models directly onto cloud or on-premises NVIDIA GPUs.
Accelerates physical AI and robotics workflows by offering pre-built cloud-native blueprints across major cloud providers like Azure and CoreWeave.
Source: blogs.nvidia.com



