AI BriefingNVIDIA Releases Cosmos 3 Open Models for Interactive World Simulation
·2026.11.06 04:35
Key point
The 16B parameter checkpoints support multi-sensor driving generation and bimanual robot action prediction under the OpenMDW1.1 license.
Details
- NVIDIA released four open-weight Cosmos 3 Nano post-training models for interactive world simulation, targeting autonomous driving and robotics.
- Cosmos3-Nano-Transfer-Auto generates synchronized video for up to 11 camera views and RGB-LiDAR data using World Scenario Maps and text prompts.
- Cosmos3-Nano-Sim-Auto enables fast streaming simulation via causal distillation, producing small chunks of synchronized camera and LiDAR data for closed-loop interaction.
- Cosmos3-Nano-Sim-Transfer converts simulator physics states into photorealistic RGB observations using spatial controls like depth and segmentation.
- Cosmos3-Nano-Sim-Bimanual predicts visual outcomes for bimanual robot actions, supporting YAM and AgiBot embodiments with 59 control values.
- All models use a 4-stage training recipe including self-forcing to prevent error accumulation and reduce inference costs through distillation and caching.
- Checkpoints are available on Hugging Face with post-training guides for adapting to custom sensor rigs and robot embodiments.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Research Introduces JCA for Joint Video and 3D Motion Generation