On June 11, 2025, NVIDIA announced the launch of the Cosmos Predict-2 world foundation model, along with new tools and services for developers aimed at accelerating the development of the next generation of autonomous vehicle systems. The new releases come as autonomous driving architectures shift from using multiple separate models to unified and integrated systems that execute driving actions directly from sensor data, increasing the need for high-quality, physically grounded sensor data for training, testing, and validation.
Cosmos Predict-2 is part of the NVIDIA Cosmos platform, which provides technologies to address the challenges of developing end-to-end autonomous driving systems. NVIDIA said that Oxa, Plus, and Uber use Cosmos models to scale synthetic data generation and accelerate vehicle development.
Improving Future-State Prediction
Cosmos Predict-2 builds on the Cosmos Predict-1 model, which was designed to predict and generate future world states based on text, image, and visual prompts. According to NVIDIA, the new version has a greater ability to understand context extracted from textual and visual inputs, resulting in fewer hallucinations and the production of detail-rich video clips.
NVIDIA is also using newer optimization techniques to accelerate synthetic data generation on NVIDIA GB200 NVL72 systems and the NVIDIA DGX Cloud service. Post-training Cosmos models on autonomous vehicle data enables the creation of video clips that match existing physical environments and vehicle trajectories, as well as the generation of multi-view video from single-view video, such as dashboard camera recordings.
This capability can transform widely available dashboard camera data into multi-camera data, opening additional sources for training autonomous driving systems. Multi-view clips can also be used in place of real camera data when some sensors are malfunctioning or obstructed.
New Training Data and Developer Tools
NVIDIA’s research team post-trained Cosmos models using 20,000 hours of real-world driving data. The company said that autonomous vehicle-specialized models improved model performance in challenging conditions, such as fog and rain, when used to generate multi-view video.
Alongside Cosmos Predict-2, NVIDIA announced Cosmos Transfer as a preview of an NVIDIA NIM microservice that can be deployed on data center GPUs. The service enhances datasets and creates realistic video clips based on structured inputs or reference simulations from the NVIDIA Omniverse platform. The NuRec Fixer model also helps fill gaps and address defects in neurally reconstructed driving data.
The CARLA platform, an open-source simulator for autonomous vehicles, is scheduled to integrate Cosmos Transfer and NVIDIA NuRec in its latest release. NuRec tools include APIs and technologies for neural reconstruction and rendering. According to NVIDIA, the integration will allow CARLA’s user base, which exceeds 150,000 autonomous vehicle developers, to render scenes, simulations, and synthetic views with a high degree of accuracy, and to create multiple variations in lighting, weather, and terrain using simple prompts.
Developers can try the pipeline using open data from the NVIDIA Physical AI Dataset. Its latest release includes 40,000 clips created by Cosmos, along with exemplary scenes reconstructed for neural rendering. With the latest CARLA release, developers can create new trajectories, reposition sensors, and simulate journeys.
Applications in the Vehicle and Safety Sectors
Plus, which specializes in autonomous trucks and is building its solution on the NVIDIA DRIVE AGX platform, uses Cosmos Predict for post-training on truck data and for creating realistic synthetic driving scenarios to accelerate the large-scale deployment of its autonomous solutions. Autonomous vehicle software company Oxa also uses the model to create highly accurate, temporally consistent multi-camera videos.
These tools are also connected to NVIDIA’s safety efforts through the NVIDIA Halos platform, which the company introduced earlier in 2025 to integrate its automotive safety hardware and software ecosystem with artificial intelligence research related to autonomous vehicle safety. Bosch, Easyrain, and Nuro have joined the Halos platform’s AI Systems Inspection Lab to validate the safe integration of their products with NVIDIA technologies. Continental, Ficosa, OMNIVISION, onsemi, and Sony Semiconductor Solutions were among the members announced as joining earlier in the year.