Details
- NVIDIA announced new Cosmos simulation recipes for generating training data and testing policies in autonomous vehicles, robots and surgical applications.
- The recipes can transform daytime driving scenes into rainy nighttime conditions while preserving road layouts and vehicle trajectories.
- Cosmos can convert depth video into detailed RGB video, changing lighting, textures and backgrounds without altering scene geometry or motion.
- The Cosmos Framework provides checkpoints and recipes for training and serving world models, including the Cosmos 3 model family.
- NVIDIA added deployment guidance for serving Cosmos models through NVIDIA NIM microservices, including GPU configuration, container startup and inference requests.
- Cosmos 3 can generate camera-controlled views from a single image, including following a subject, rising vertically or moving sideways.
- The release includes simulation options for visual transfer, robot and camera actions, and synchronized multi-view driving scenarios.
- NVIDIA’s interactive simulation examples report real-time performance for selected Cosmos 3 Nano models, including 48.1 frames per second with 166-millisecond latency on a single GB300 GPU.
Impact
The release strengthens NVIDIA’s position in physical AI by connecting world-model training, synthetic data generation and deployment within one stack. Preserving geometry and motion while varying environmental conditions could reduce reliance on costly real-world data collection, particularly for edge cases in autonomous driving and robotics. The inclusion of NIM serving and Framework training recipes also makes Cosmos more operationally accessible to developers, while NVIDIA’s dependence on its GPU ecosystem may limit adoption among teams using competing hardware.