Nvidia has expanded its Cosmos platform with five cloud-native blueprints for preparing synthetic and real-world training data for robots, vehicles and industrial vision systems. The tools are available through GitHub, Microsoft Azure, Nebius and CoreWeave.
The blueprints cover autonomous-driving datasets, object-appearance changes, synthetic videos of rare events, defect images for industrial inspection, and weather or lighting changes in existing footage. Cosmos Curator now handles multimodal and multisensory inputs, while agent tools can assist with labeling, curation, augmentation and evaluation.
Nvidia also extended its open training framework to models including GR00T, Alpamayo, Qwen and Pi0.5. Developers can fine-tune them, apply reinforcement learning, distill or quantize them, and deploy supported Cosmos models through NIM microservices. New Cosmos 3 Nano simulation variants support repeatable tests of vehicle and two-handed robot policies before physical trials.
Company examples include 1,500 synthetic clips created by ANYbotics to train inspection robots and a Porsche project intended to reduce repeated test drives. Those outcome estimates come from Nvidia and partners rather than independent comparisons. Synthetic data can cover dangerous or rare events, but teams still need real-world validation before trusting a policy on physical equipment.