NVIDIA Expands Cosmos Platform with New Tools to Train, Evaluate, and Deploy World Models

NVIDIA has announced a major expansion of its Cosmos platform, introducing a suite of developer tools, simulation checkpoints, and cloud-native blueprints designed to accelerate the creation, evaluation, and deployment of world models. Built to help physical AI systems understand and simulate real-world environments, the platform currently ranks in the top five across six different open model leaderboards.

The update provides developers with tools to scale multimodal training datasets, fine-tune and distill models, and simulate robotic and autonomous policies before deploying them into physical environments.

Cloud-Native Blueprints for Data Preparation

World models depend heavily on rich, multisensory data. To address data scarcity and curation bottlenecks, NVIDIA has upgraded Cosmos Curator—its GPU-accelerated video curation engine—to support multimodal and multisensory inputs.

Alongside Curator, NVIDIA launched five cloud-native Cosmos Data Factory blueprints, accessible via GitHub, Microsoft Azure, Nebius, and CoreWeave:

• Autonomous Vehicle Dataset Generation: Curates and labels real driving data while synthesizing rare edge cases.
• Image Attribute Augmentation: Alters object attributes to refine visual search systems.
• Event Video Generation: Transforms still images into synthetic video sequences depicting rare or hazardous scenarios.
• Defect Image Generation: Synthesizes defect imagery on printed circuit boards, glass, and metal for industrial quality inspection.
• Video Data Augmentation: Adjusts lighting and weather conditions in existing video feeds while retaining scene geometry.

Industrial and automotive partners are already adopting these blueprints. Porsche is developing a shared autonomous vehicle data flywheel across Volkswagen Group brands using Cosmos Data Factory through Voxel51 on Microsoft Azure, aiming to cut repeated physical test drives by up to 90% and reduce development cycles by up to 85%. Industrial robotics maker ANYbotics produced more than 1,500 synthetic video clips using Cosmos blueprints and Encord to train inspection robots to identify liquid spills, while computer vision firm Ultralytics used synthetic defect generation to retrain its YOLO model for enhanced localization and recall.

Training, Adaptation, and Model Distillation

Announced at GTC Berlin, the expanded Cosmos Framework opens up training and optimization workflows as agentic skills across Cosmos 3 and compatible open models. Developers can now fine-tune and apply reinforcement learning to models including GR00T, Alpamayo, QWEN, Pi0.5, and other Hugging Face architectures.

The framework supports model optimization techniques such as quantization and distillation. In one implementation, robotics lab MicroAGI post-trained Cosmos 3 on egocentric manipulation footage and distilled its 32-billion-parameter reasoning model into an 8-billion-parameter Nano model without sacrificing accuracy. Developers can also label data and post-train Cosmos 3 using Roboflow, deploying the resulting models directly to edge GPUs in warehouses and factories.

Closed-Loop Simulation and Real-Time Inference

To streamline evaluation, NVIDIA rolled out new Cosmos 3 Nano model checkpoints tailored for interactive simulation. These variants enable developers to evaluate closed-loop policies for autonomous vehicles and bimanual manipulation robots across repeatable, multiview scenarios before real-world deployment.

For production workloads, updated Cosmos 3 NIM microservices deliver up to 2x faster inference performance with no loss in output quality, enabling sub-second alerting for continuous video analysis.

NVIDIA also released the Metropolis Blueprint for video search and summarization (VSS) 3.3. Powered by Cosmos NIM microservices, the blueprint allows teams to monitor 40% more camera streams on the same GPU hardware. Energy group ADNOC has already integrated this architecture via Fogsphere’s MistIQ platform to monitor safety hazards in real time across 65 operational sites, utilizing synthetic data from Cosmos Data Factory to cover dangerous scenarios without staging them in real life.