NVIDIA expands Cosmos tools for world model development

NVIDIA has expanded its open, modular Cosmos platform with tools for preparing training data, adapting world models, evaluating autonomous systems and deploying them. Announced at GTC Berlin, the updates target physical AI systems that need to understand and simulate real-world environments.

Cosmos Data Factory now offers five cloud-native blueprints covering driving datasets, image variations, synthetic event videos, defect images and video augmentation. It is available through GitHub, Microsoft Azure, Nebius and CoreWeave. Cosmos Curator also adds support for multimodal, multisensory inputs. Porsche is using the platform through Voxel51 on Microsoft Azure to develop a shared autonomous-driving data workflow across Volkswagen Group brands, with expected reductions in repeated physical test drives of up to 90% and development cycle times of up to 85%.

Cosmos Framework extends training, optimization and deployment tools to more open models, supporting fine-tuning, reinforcement learning, distillation and quantization. New Cosmos 3 Nano simulation variants let developers evaluate autonomous vehicle and bimanual robot policies in repeatable, multiview, multisensor scenarios before real-world testing.

Updated Cosmos 3 NIM microservices deliver up to 2x faster inference without quality loss in validated configurations. The NVIDIA Metropolis video search and summarization blueprint helps teams monitor 40% more video streams with the same GPU. ADNOC uses Fogsphere’s Cosmos-based platform to detect safety hazards across 65 sites, with synthetic training data supporting detection of rare events.