Summary
NVIDIA has added data-generation blueprints, model-training tools and simulation options to its Cosmos platform for developing physical AI. The company also reported customer examples and performance figures for selected tools.
NVIDIA has expanded its Cosmos platform with tools for preparing training data, adapting models and testing autonomous systems in simulation. The company describes Cosmos as a set of open models and development tools for world models: AI systems intended to help physical AI applications interpret and simulate real-world situations.
The announcement brings together three parts of that process: Cosmos Data Factory for preparing and generating data, Cosmos Framework for training and optimizing models, and simulation and inference tools for evaluating and running them. The tools are intended for developers working on applications such as robotics and autonomous vehicles.
New blueprints for training data
Cosmos Data Factory now offers five cloud-native blueprints, available through GitHub, Microsoft Azure, Nebius and CoreWeave. They cover autonomous-vehicle dataset curation and labeling; image-attribute augmentation; synthetic videos of rare or unsafe events; synthetic defect images for inspection; and changes to weather and lighting in existing video while preserving scene structure.
NVIDIA also says Cosmos Curator, its GPU-accelerated video-dataset curation tool, now supports multimodal and multisensory inputs and processing blocks. The aim is to help developers assemble and modify datasets for training physical AI systems, including by generating examples of situations that may be difficult to capture in ordinary data collection.
NVIDIA describes several uses by companies. Porsche is using the blueprints through Voxel51 on Microsoft Azure as part of a shared autonomous-vehicle data project across Volkswagen Group brands. NVIDIA says the project is expected to reduce repeated physical test drives by up to 90% and shorten development cycles by up to 85%; these are projected outcomes, not figures described as already achieved. ANYbotics generated more than 1,500 synthetic video clips for training industrial-inspection robots to recognise liquid spills and other changes. Ultralytics, meanwhile, retrained a YOLO model with real images and targeted synthetic defects, and reported improved defect localization and recall.
Training, simulation and deployment
Cosmos Framework adds training, optimization and deployment tools for open models. Developers can fine-tune models or use reinforcement learning, then apply techniques such as distillation and quantization to adapt them for deployment. NVIDIA lists GR00T, Alpamayo, QWEN and Pi0.5 among the models developers can adapt, alongside compatible models from Hugging Face.
One example from robotics research lab MicroAGI involved post-training Cosmos 3 on videos of people manipulating objects. NVIDIA says MicroAGI distilled the model’s 32-billion-parameter reasoner into an 8-billion-parameter Nano model with no loss in accuracy. The source does not give further evaluation details for that result.
New Cosmos 3 Nano variants for simulation are designed to let developers test autonomous policies in repeatable, multiview and multisensor scenarios, including for autonomous vehicles and bimanual robots, before real-world testing. NVIDIA also says updated Cosmos 3 NIM microservices deliver up to 2× faster inference with no quality loss in validated configurations. Its Metropolis Video Search and Summarization Blueprint 3.3, built on Cosmos NIM, is reported to support monitoring 40% more video streams on the same GPU.
Taken together, the additions position Cosmos as a modular toolkit for moving from dataset preparation to model adaptation, simulation and deployment. The reported performance and customer outcomes come from NVIDIA’s announcement; the stated inference improvement, in particular, applies to validated configurations.