NVIDIA releases four open-weight Cosmos 3 simulation models

NVIDIA has released four open-weight checkpoints for its Cosmos 3 Nano model family, each featuring 16 billion trainable parameters. The release includes specialized models designed for different simulation tasks: Transfer-Auto for synthetic driving data generation, Sim-Auto for streaming driving simulations, Sim-Transfer for rendering color video from physics simulator controls, and Sim-Bimanual for predicting video frames based on robot or camera actions.

The Sim-Bimanual model supports two-arm setups for YAM and AgiBot embodiments and can be operated via joysticks or policies. While NVIDIA provides post-training recipes to adapt these models to specific domains, technical analysis suggests that current evaluation metrics, such as PSNR and SSIM, primarily measure visual resemblance rather than whether the generated motion accurately follows specific action sequences or predicts correct consequences in robotic tasks.