NVIDIA Expands Cosmos With World Model Training and Simulation Tools

NVIDIA has expanded its Cosmos platform with new tools for training, evaluating and deploying world models, as detailed in a company blog post. The additions include Cosmos Data Factory blueprints, support for more open models and new Cosmos 3 variants for simulation.

Cosmos Data Factory now provides five blueprints for autonomous vehicle datasets, image attribute augmentation, event video generation, defect image generation and video data augmentation. They are available through GitHub, Microsoft Azure, Nebius and CoreWeave. Cosmos Curator also adds support for multimodal and multisensory inputs. NVIDIA says Porsche is using Cosmos Data Factory through Voxel51 on Microsoft Azure to build a shared autonomous driving data pipeline across Volkswagen Group brands, which is expected to reduce repeated physical test drives by up to 90%.

The Cosmos Framework can now adapt models including GR00T, Alpamayo, Qwen and Pi0.5 through fine tuning and reinforcement learning. It also supports distillation, quantization and deployment tools. New Cosmos 3 Nano variants let developers test autonomous vehicle and robot policies in repeatable simulations before physical testing.

Updated Cosmos 3 NIM microservices provide up to twice the inference speed in validated configurations. NVIDIA also released Metropolis Blueprint for video search and summarization 3.3, which uses Cosmos services and can monitor 40% more video streams with the same GPU.