| Intended Application & Domain |
World generation and simulation for Physical AI, specifically sim-to-real visual transfer for robotic policy training and evaluation. |
| Model Type |
Transformer-based Mixture-of-Transformers (MoT) model, derived from Cosmos3-Nano through controlled video-generation post-training and causal distillation. |
| Intended Users |
Physical AI developers, researchers, and practitioners integrating generated visual observations into robotics simulation, policy-training, and evaluation workflows. |
| Output |
Photorealistic RGB video in MP4 format or streamed video frames, conditioned on simulator-rendered controls. |
| Tools used to evaluate datasets to identify synthetic data and ensure data authenticity. |
Dataset provenance analysis, metadata validation, watermark and artifact detection, embedding-based clustering, heuristic quality checks, and model-assisted data validation pipelines are used to identify synthetic content patterns, assess dataset authenticity, and improve data quality during dataset curation. |
| Describe how the model works |
The model generates video conditioned on simulator-rendered visual controls and text. It uses cached history and incoming controls, while the simulator maintains the underlying physics and scene state. |
| List the groups or group characteristics for which this was tested to produce comparable outcomes |
None. |
| Technical Limitations & Mitigation |
The model may not follow input conditions accurately in challenging cases, especially those involving complex scene composition, unusual camera motion, multiple interacting agents, low lighting, high motion blur, or fine-grained physical interactions. Generated outputs may contain temporal inconsistency, object morphing, inaccurate 3D structure, or implausible physical dynamics. Users should validate output quality and adherence to input conditions for their intended use cases. |
| Verified to have met prescribed NVIDIA quality standards |
Yes. |
| Performance Metrics |
Evaluation uses PAI-Bench-C and Isaac-generated sequences. Metrics include DOVER, Vision-Language Model (VLM) video quality, depth SI-RMSE, segmentation mIoU, blur SSIM, edge F1, prompt following, quality drift, and temporal Sampson error. Control-specific metrics apply to the corresponding evaluation tracks. These metrics are listed in the overview; please see the overview for numerical results. |
| Potential Known Risks |
This model can generate synthetic media and may produce content that is offensive, unsafe, misleading, indecent, or unsuitable for a target deployment. Users should implement robust safety guardrails — including content filtering, abuse monitoring, and access controls — to reduce the risk of harmful outputs. Users are responsible for ensuring that their use of the model complies with all applicable laws and regulations, and for regularly reviewing and updating their guardrails as risks evolve. |
| Licensing |
OpenMDW1.1 |
