Nvidia launches vision AI agent blueprints for industry

Nvidia has introduced Metropolis agent skills and blueprints for vision AI agents, designed to support model development, simulation and deployment across edge and cloud environments.

The launch focuses on reusable workflows for synthetic data generation, video data augmentation, model fine-tuning, and video search and summarisation. Developers can use them with Omniverse for OpenUSD-based simulation and digital twins, and with Metropolis to build and run video AI applications.

Vision AI agents are being adopted in factories, warehouses, transport networks and urban infrastructure as operators seek to turn camera feeds into automated alerts, reporting and process monitoring. Nvidia positioned the new software as a response to a common edge computing problem: large volumes of data are generated near cameras and sensors, but much of it is never turned into action.

Nvidia identified three main obstacles organisations face when building these systems: a lack of representative training data, especially for rare defects or abnormal events; the specialist work needed to fine-tune models after performance gaps emerge; and the engineering effort required to combine video pipelines, models, metadata, search, alerting and system integrations into a working application.

Manufacturing use

In manufacturing, synthetic data can help address a shortage of real-world defect images. Nvidia highlighted work by Roboflow, which is integrating Nvidia's Defect Image Generation skill and Cosmos world foundation models into its platform for customers including Corning.

According to Nvidia, a benchmark with Corning's optical fiber manufacturing engineering team found that a model trained on eight real defect images, combined with synthetic data generated by the Defect Image Generation skill, achieved 95% average precision and perfect recall on the most difficult defect class. It outperformed a baseline model trained only on real data and reduced a project expected to take multiple quarters to a matter of days.

The example underscores one of the main commercial arguments for synthetic data in industrial inspection. Production lines that prevent most defects can struggle to collect enough examples of failures to train the next generation of inspection systems, leaving models weak at detecting infrequent but important anomalies.

City systems

Nvidia also pointed to urban operations as a market for connected video workflows. Linker Vision is using the Nvidia Metropolis Blueprint for video search and summarisation to deploy video reasoning agents across city infrastructure, while using Omniverse digital twins based on OpenUSD to model traffic, weather, emergency events and infrastructure changes.

The system packages tasks such as search, summarisation, alerts, reporting and stream management into workflows that agents can execute. Linker Vision also uses Nvidia Cosmos for video data augmentation and Nvidia TAO for model fine-tuning.

In Kaohsiung, Nvidia said Linker Vision cut development effort by 85% with the video search and summarisation blueprint and reduced incident response times by up to 80%. The company added that the group's newer AI-GRID expansion includes NemoClaw blueprints for secure agentic AI in city and transport settings.

Factory operations

Another example came from industrial workflow monitoring rather than defect detection alone. According to Nvidia, DeepHow's Live Standard Operating Procedure Verification agent, deployed at Foxconn, uses the Metropolis video search and summarisation blueprint to search, summarise and analyse video in operational environments.

The goal is to assess whether work is being carried out correctly, compare actions with standard procedures and identify problems before defects move downstream. Nvidia said Cosmos helps the system interpret sequences of human actions in context, including whether assembly steps are performed in the correct order.

On Nvidia GB300 server production lines, the DeepHow system improved first-pass yield by 3%, achieved 99% task-level accuracy in understanding critical procedure steps and reduced redundant work by identifying issues earlier in the process, according to Nvidia.

Edge push

The broader market context for the launch is a shift in AI processing towards the edge, where data is created rather than sent back to centralised infrastructure. Nvidia cited Gartner forecasts that more than two-thirds of enterprise-managed data will be created and processed outside data centres or the cloud by 2028, and that more than two-thirds of enterprises worldwide will deploy edge AI by 2029, up from 10% in 2025.

Even so, more edge data does not automatically produce more useful insight. Models running close to cameras and machines must work within constraints on latency, power, cost and connectivity, while also adapting to the conditions of each site.

OpenUSD sits at the centre of Nvidia's response because it provides a common way to describe and reuse 3D scenes. Omniverse libraries help teams build simulation, synthetic data and digital twin workflows that expand testing across lighting conditions, weather, traffic patterns, camera angles, occlusion and rare events.

The new package includes the Defect Image Generation skill, the Video Data Augmentation skill, TAO skills for model fine-tuning, and video search and summarisation skills for alerts, reporting and stream management. The aim is to stop developers from rebuilding each part of the workflow from scratch for every deployment.

Those reusable workflows are intended to help developers generate data, improve models and deploy vision AI agents across industrial, transport and city operations.