NVIDIA is providing reusable agent skills and blueprints to help developers build vision AI systems that process video d
NVIDIA is positioning vision AI agents as a practical way to automatically convert video data from physical environments into operational intelligence for factories, cities, warehouses and transportation systems. This transition to edge processing is accelerating as enterprises recognize that most data creation and processing will move outside traditional data centers. Gartner projects that more than two-thirds of enterprise-managed data will be created and processed outside the data center or cloud by 2028, with two-thirds of all enterprises globally deploying edge AI by 2029, up from just 10% in 2025.
Despite this shift, the infrastructure challenge remains significant. As much as 90% of existing edge data currently goes unprocessed. Converting raw video into actionable intelligence requires vision AI agents that understand video, adapt to real-world conditions and integrate insights into operational workflows. These agents typically run near cameras, machines and sensors in locations with strict latency, power, cost and connectivity constraints, while adapting to site-specific environmental conditions. Developers need repeatable processes for generating training data, fine-tuning models and deploying agentic video applications across both edge and cloud environments.
NVIDIA Metropolis agent skills and blueprints provide developers with reusable workflows throughout this lifecycle. For simulation and synthetic data generation, OpenUSD offers a common framework for describing and composing 3D worlds. NVIDIA Omniverse libraries, built on OpenUSD, help teams create simulation, synthetic data generation and digital twin workflows that model real-world environments and cover varied conditions including lighting, weather, traffic patterns, camera angles, occlusion and rare events.
The company highlights three specific deployment scenarios. In manufacturing, Roboflow integrated NVIDIA's Defect Image Generation skill and Cosmos world foundation models to generate synthetic defect images for customers like Corning when real training data is limited. A benchmark with Corning's optical fiber manufacturing team showed that a model trained on just eight real defect images, augmented with synthetic data, achieved 95% average precision and perfect recall on the most challenging defect class, surpassing models trained solely on real data and compressing a multi-quarter project into just a few days.
In smart cities, Linker Vision built AI systems using the NVIDIA Metropolis Blueprint for VSS to deploy video reasoning agents across city infrastructure, using VSS skills to package common video AI tasks into reusable agent workflows. In Kaohsiung, this approach reduced development effort by 85% and cut incident response times by up to 80%. Linker Vision's newer AI-GRID expansion applies this methodology to autonomous video reasoning across city and transportation environments.
In industrial environments, DeepHow built a Live Standard Operating Procedure Verification agent using the NVIDIA Metropolis VSS blueprint as the workflow layer, with NVIDIA Cosmos providing reasoning capability to interpret complex human activity and work sequences. At Foxconn's GB300 server production lines, the solution improved first-pass yield by 3%, achieved 99% task-level accuracy in understanding critical SOP steps and reduced redundant work by catching problems earlier.