NVIDIA expands Cosmos with five training-data blueprints
3-line summary
- NVIDIA expanded Cosmos data, training and evaluation tools at GTC Berlin.
- Cosmos Data Factory provides five reference blueprints for tasks including driving data and defect images.
- The public repository documents setup and GPU needs; synthetic outputs still require quality checks.
NVIDIA announced a Cosmos expansion at GTC Berlin on October 21, local time. The update covers data preparation, training and evaluation for physical AI, including robotics and autonomous driving.
Cosmos Data Factory offers five reference blueprints: driving datasets, image attribute augmentation, event videos, defect images and video augmentation. Examples include changing weather or lighting and generating rare-event footage from an image.
The expanded Cosmos Framework supports training and inference. Its repository also describes post-training, reinforcement learning and distillation, giving teams components to connect data generation with model adaptation.
Deployment still needs validation. Data Factory documents GPU requirements by workflow and uses Cosmos Evaluator to check synthetic videos for issues including physical plausibility and object consistency. Teams should test outputs against their own operating conditions.
Why it matters
Rare events and defects are difficult to collect. These tools offer ways to supplement datasets, but generating more data and improving field accuracy remain different goals.
Who it's for
- Developers
- Start with one relevant blueprint and check its GPU, storage and output-quality requirements.
