NVIDIA Cosmos adds data blueprints and simulation tools for physical AI

NVIDIA has expanded Cosmos with training‑data workflows, model‑adaptation tools and simulation checkpoints for physical AI. Its [October 22 release](https://blogs.nvidia.com/blog/cosmos‑open‑train‑evaluate‑deploy‑world‑models/) describes these as separate parts of the path from raw sensor material to a model that can be evaluated before deployment. The data stage is Cosmos Data Factory, which combines labeling, curation, augmentation and evaluation tools. Cosmos Curator, previously described around video datasets, adds multimodal, multisensory inputs and processing blocks. For a robotics or inspection team, the first choice is therefore the data problem to address, rather than simply which model to run.

The following table groups the blueprints by the task [NVIDIA describes](https://blogs.nvidia.com/blog/cosmos‑open‑train‑evaluate‑deploy‑world‑models/). It is an original summary of the release, not a measured comparison of training quality or cost. | Blueprint | Reported purpose | | --- | --- | | Autonomous Vehicle Dataset Generation | Driving‑data curation and labeling, plus synthetic scenarios | | Image Attribute Augmentation | Changes to object appearance for image‑search data | | Event Video Generation | Synthetic event videos made from still images | | Defect Image Generation | Defect examples for inspection training | | Video Data Augmentation | Weather or lighting variations in existing video | These categories differ in what they change: some add unusual situations, some vary existing footage and others create inspection examples. A synthetic defect image and a weather variation are different inputs even when both extend a training set. Generated examples still need review against the target environment; the list does not show that one blueprint is more effective than another. NVIDIA lists Data Factory availability through GitHub, Microsoft Azure, Nebius and CoreWeave. The announcement does not establish identical features, pricing or access terms for every delivery path.

[The update also extends Cosmos Framework tools](https://blogs.nvidia.com/blog/cosmos‑open‑train‑evaluate‑deploy‑world‑models/) and recipes as agent skills to more open models. NVIDIA names GR00T, Alpamayo, QWEN and Pi0.5 among models developers can adapt, with fine‑tuning and reinforcement‑learning workflows, followed by options such as distillation and quantization. Those are steps in a development process, not interchangeable measures of success. A team can begin from its own architecture or a pretrained model, then decide what behavior to adapt and how to review the result. Exact repository requirements, component licenses and hardware eligibility need checking before implementation; this launch account does not substitute for those instructions.

New Cosmos 3 Nano simulation variants and accompanying code are intended for repeatable multiview, multisensor scenarios. [NVIDIA says](https://blogs.nvidia.com/blog/cosmos‑open‑train‑evaluate‑deploy‑world‑models/) developers can adapt them to train and evaluate autonomous‑vehicle and bimanual‑robot policies in interactive simulations before real‑world testing. Updated Cosmos 3 NIM microservices provide another deployment route for supported models. A simulated scenario is a controlled setting for examining a policy. Passing it does not by itself establish safe performance in every physical setting. The useful evaluation connects the prepared data, adapted model and selected scenarios to the actual task, while keeping simulation results distinct from real‑world evidence. No comparative benchmark or independent safety result is established here. *Photo: Coolcaesar / [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:NVIDIA_Headquarters.jpg), [CC BY‑SA 4.0](https://creativecommons.org/licenses/by‑sa/4.0/). NVIDIA headquarters in Santa Clara, August 2018; context photograph, not a demonstration of the October 2026 software. Native display crop and downscale; image adaptations retain the same license.*