英伟达开源Cosmos 3模型，面向驾驶与机器人仿真
英伟达发布四款Cosmos 3开放权重模型，生成带相机、激光雷达和动作反馈的交互式世界模拟数据。
NVIDIA opens a broader world-model toolkit
NVIDIA released four open-weight checkpoints in the Cosmos 3 family for interactive world simulation. The models target two demanding areas: autonomous driving and robotics. All four are post-trained versions of Cosmos 3 Nano with 16 billion trainable parameters and are released under the OpenMDW 1.1 license for commercial and non-commercial use.
The driving models generate synchronized observations across multiple cameras and LiDAR. Cosmos3-Nano-Transfer-Auto can vary weather, lighting, traffic, and road layouts, while Cosmos3-Nano-Sim-Auto produces streaming camera and LiDAR observations from driving controls. The robotics checkpoint, Cosmos3-Nano-Sim-Bimanual, predicts how a scene changes after two-arm robot or camera actions.
From synthetic video to action-conditioned environments
The release is more significant than a conventional video-generation model because it connects observations to controls. A simulator can update physical state, then Cosmos can render the resulting visual view. In the bimanual model, actions themselves become inputs for predicting the next observation.
That design lets researchers build closed-loop tests: a policy acts, the environment responds, and the next observation feeds back into the policy. NVIDIA also provides post-training recipes, data contracts, and inference skills so developers can adapt the checkpoints to different sensor rigs or robot embodiments.
The limitations are material. Running the released examples requires Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit; the published setup was tested on one GB200. The models also do not automatically provide a trustworthy substitute for physical testing. NVIDIA explicitly says researchers still need to measure where the simulations support policy learning and where they fail.
Why it matters
Open access to action-conditioned driving and manipulation models lowers the cost of experimentation for robotics and autonomous-vehicle teams. More importantly, it shifts world-model competition toward reusable simulation infrastructure, sensor alignment, and evaluation—not only prettier generated frames.
Uncle Cat take: Cosmos 3’s differentiator is synchronized camera-LiDAR output, but its real adoption risk is the GB200-class setup required by the reference workflow.