NVIDIA releases four open Cosmos 3 models for world simulation

NVIDIA has added four open-weight checkpoints to its Cosmos 3 family of world models for interactive simulation. The Cosmos3-Nano-Transfer-Auto model generates synchronized driving scenes across up to 11 camera views and can produce joint RGB and LiDAR data. Cosmos3-Nano-Sim-Auto creates successive camera and LiDAR segments from driving controls, supporting streaming and longer interactive rollouts.

The other two models target robotics and simulator-based rendering. Cosmos3-Nano-Sim-Transfer converts simulator controls and a text prompt into generated RGB observations, while Cosmos3-Nano-Sim-Bimanual predicts subsequent video from robot or camera actions. The latter supports interactive two-arm simulation for YAM and AgiBot embodiments.

All four checkpoints are post-trained versions of Cosmos 3 Nano and have 16B trainable parameters. They are released under OpenMDW1.1, which permits commercial and non-commercial use under its terms. NVIDIA describes the models as starting points for adapting sensor rigs, driving domains and robotic embodiments, while noting that further work is needed to measure their usefulness for policy learning and evaluation. The published setup requires Linux, an NVIDIA GPU, Docker and NVIDIA Container Toolkit.