The development
What happened
NVIDIA has described four open-weight Cosmos 3 Nano variants for synthetic driving data and interactive robot observations. Its 5 November walkthrough separates offline generation, streaming driving scenes, simulator-controlled rendering and action-conditioned robot video, rather than treating them as one interchangeable simulator.
Key facts
Transfer-Auto creates offline multi-camera driving data, including joint RGB–LiDAR output. Sim-Auto streams observations but needs a simulation runtime and updated scene controls for closed-loop interaction.
Sim-Transfer renders a physics simulator’s state into RGB observations; Sim-Bimanual instead predicts observations from robot or camera actions.
NVIDIA lists Linux, an NVIDIA GPU, Docker and Container Toolkit. Its one-GB200 Bimanual setup is a tested configuration, explicitly not a minimum-memory requirement.
The context
Why it matters
TideReach assessment: the main choice is which system owns physical state and supplies the next controls. Generated realism is not evidence of reliable policy evaluation; teams still need calibrated inputs and held-out tests, including errors that accumulate through generated history.