NVIDIA expanded Cosmos with data-generation blueprints, model-training tools and simulation models, citing customer deployments and performance results.
NVIDIA announced an expansion of its Cosmos platform on October 21, 2026, adding tools for preparing training data, adapting world models and testing robotic or autonomous-driving policies in simulation. The company describes Cosmos as a modular platform for developers building models that help physical AI systems interpret and simulate real-world environments.
The data offering includes five cloud-based blueprints for tasks such as generating autonomous-vehicle datasets, creating synthetic event videos and producing defect imagery for inspection. Cosmos Curator, NVIDIA’s video-data curation tool, now supports multimodal, multisensory inputs. The blueprints are available through GitHub and cloud providers including Microsoft Azure, Nebius and CoreWeave.
NVIDIA cited several customer examples. Porsche is using the tools through Voxel51 on Azure as part of a shared autonomous-vehicle data effort across Volkswagen Group brands. NVIDIA says the approach is expected to cut repeated physical test drives by up to 90% and shorten development cycles by up to 85%. ANYbotics generated more than 1,500 synthetic video clips for training industrial-inspection robots to identify liquid spills and other changes. Ultralytics used synthetic defect examples alongside real images to retrain its YOLO model, reporting improved defect localization and recall.
The expanded Cosmos Framework lets developers fine-tune and use reinforcement learning with models including GR00T, Alpamayo, QWEN and Pi0.5, as well as compatible Hugging Face models. NVIDIA also described a MicroAGI project that distilled Cosmos 3’s 32-billion-parameter reasoner into an 8-billion-parameter model, reporting no loss in accuracy for its object-manipulation task.
For evaluation, new Cosmos 3 Nano variants can generate multiview, multisensor simulations for testing autonomous policies before physical trials. NVIDIA says updated Cosmos 3 NIM microservices can deliver up to twice the inference speed with no quality loss in validated configurations. These additions connect dataset preparation, model adaptation, simulation and deployment within the same platform.