# NVIDIA adds four open Cosmos 3 world models for driving and robotics simulation

published: 2026-11-05T21:05:53.484Z
topic: infrastructure
source: https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation
url: https://aiagentnews.io/stories/nvidia-adds-four-open-cosmos-3-world-models-for-driving-and-robotics-simulation

NVIDIA has added four open-weight checkpoints to its Cosmos 3 model family for interactive world simulation. They cover multi-view driving data, streaming driving simulation, simulator-driven video and bimanual robot action prediction.

NVIDIA has added four open-weight checkpoints to its Cosmos 3 model family, aimed at interactive world simulation and multi-sensor driving scenes. All four are post-trained variants of Cosmos 3 Nano with 16 billion trainable parameters, released under the OpenMDW 1.1 license, which permits commercial and non-commercial use on its terms.

Two checkpoints target autonomous driving. Cosmos3-Nano-Transfer-Auto generates synchronized video across up to 11 camera views, including fisheye cameras, with joint RGB and LiDAR output for offline synthetic data. Cosmos3-Nano-Sim-Auto is a distilled variant that streams small chunks of camera and LiDAR data in a few denoising steps, and closed-loop use requires a simulation runtime and WSM renderer to update controls.

Cosmos3-Nano-Sim-Transfer renders RGB video from simulator controls such as depth and segmentation, plus a text prompt, while physics stays with the simulator. Cosmos3-Nano-Sim-Bimanual predicts the next observations from two-arm robot actions for YAM and AgiBot embodiments. Running the models requires Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit.

## Why it matters if you build agents

Teams building robotics or driving simulators can run these checkpoints locally and adapt them to their own sensor rigs or robot embodiments using the published post-training recipes. The published material says a short successful run verifies execution, not convergence, so developers should evaluate on held-out episodes before using outputs for policy learning. Action timing needs care, since T transitions require T+1 observations.