NVIDIA releases four open Cosmos 3 Nano models for simulation in robotics and autonomous driving
NVIDIA has released on Hugging Face four post-trained Cosmos 3 Nano models (16B trainable parameters) for generating interactive simulation scenes for robotics and autonomous driving, under the OpenMDW1.1 license allowing both commercial and non-commercial use.
NVIDIA has expanded the Cosmos 3 model family with four new post-trained variants with open weights, focused on generating interactive simulation scenes for robotics and autonomous driving. The checkpoints, technical report, and sample demonstrations are available for immediate use on Hugging Face.
Each of the four models has a specific purpose: Cosmos3-Nano-Transfer-Auto generates multi-view driving data with varying weather, lighting, and time of day for offline synthesis of training data. Cosmos3-Nano-Sim-Auto is a distilled streaming variant that produces synchronized camera and LiDAR sensor data in small steps for fast interactive simulations. Cosmos3-Nano-Sim-Transfer converts physical simulator controls and a text prompt into photorealistic video. Cosmos3-Nano-Sim-Bimanual predicts video from robot or camera actions and supports interactive simulation of a bimanual robot for the YAM and AgiBot embodiments.
According to the company, all four models have 16B trainable parameters and are available under the OpenMDW1.1 license, which permits both commercial and non-commercial use. According to the company, the goal is to provide researchers and developers with an open foundation for building richer and more responsive simulated environments and a shared starting point for adapting models and benchmarking methods; at the same time, the company states that it remains an open question to what extent the models faithfully support learning and evaluation of control policies.
For running the models, the source states the following requirements: Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit. For the remaining technical details on the individual models and their deployment, see the source article.
Why it matters
The models address a specific problem in the development of autonomous vehicles and robots: the need to generate variable, synchronized sensor data (camera, LiDAR) for training and testing control policies without having to collect real-world data in the field. The open OpenMDW1.1 license allows deployment even in commercial products, which lowers the barrier for companies that would otherwise have to build their own generative models for simulation.
Two audiences, two different impacts
What this means
For individuals
Developers and researchers in robotics and autonomous driving gain four specialized open checkpoints with 16B parameters, which they can immediately download and try on their own sensor rig or dataset, including a technical report and demonstrations.
For a business
Companies developing autonomous vehicles or robotic systems can use these models under the OpenMDW1.1 commercial license to synthetically generate training and test data (multi-view video, LiDAR), which may reduce the cost of real-world data collection, although NVIDIA itself states that the reliability of the models for training control policies still needs to be verified.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.