Inside-Out, Outside-In: Two Ways Vision AI Enters the Factory

Physical AI is moving from research labs onto the factory floor along two distinct paths. One begins with digital twins, simulation, and synthetic data before a robot reaches real hardware. The other begins with the cameras already installed across industrial facilities and uses Vision AI agents to turn existing infrastructure into systems for inspection, safety, tracking, and operational insight.

At ImagingNext 2026, Maycon Douglas da Silva Carvalho, Senior Solutions Architect at NVIDIA, presents “Inside-Out, Outside-In: Two Ways Vision AI Enters the Factory” a session on how these two approaches bring physical AI into industrial environments and where they converge at the edge.

About Maycon Douglas da Silva Carvalho

Maycon Douglas da Silva Carvalho is a Senior Solutions Architect with a background in electrical and computer engineering, specializing in robotics and embedded systems. His expertise includes deploying perception AI and multimodal architectures such as Vision Language Models and Vision Language Action Models at the edge. His experience also covers Software-in-the-Loop for robotics and AI systems, advanced sensor simulation, and the hardware integration of industrial sensor and camera systems. Having worked with NVIDIA technologies for more than five years, Maycon Douglas da Silva Carvalho has helped hundreds of customers across the EMEA region realize their technological ambitions.

Why these two paths to physical AI matter

The “inside-out” path builds intelligence from the digital twin outward. Factories, cells, and robots are modeled in NVIDIA Omniverse, while world foundation models such as Cosmos generate the synthetic data and neural simulation needed to train perception and control policies. Robots can then be trained and validated in simulation before they interact with real hardware.

This approach changes the data equation for industrial AI. Robot capabilities are no longer driven only by real-world data collection. Compute, simulation, and validation also contribute to the development process, while cameras support the different phases of physical AI data generation, training, simulation, validation, and deployment.

The “outside-in” path starts from the billions of cameras already installed in industrial environments. Vision AI agents built on NVIDIA Metropolis can turn existing camera networks into systems for automated inspection, worker safety, multi-camera tracking, and operational insight without requiring the facility to be redesigned.

These two paths address different starting points. A digital twin first approach begins with a simulated representation of the factory or robot, while a vision agent approach begins with existing imaging infrastructure. Understanding when to use each approach and how they can work together is an important system design decision for industrial applications.

The session will explore both approaches through real industrial deployments and show how they converge at the edge on Jetson Thor and Holoscan. Multi-camera systems, sensor bridges, and edge compute must work together to support low-latency deployment and connect data capture with trained policies and physical actions.

This convergence also changes the role of the imaging chain. Sensors, optics, ISPs, and sensor bridges are not separate from the physical AI architecture. They are its entry point, capturing the data that supports simulation, training, inference, and deployment.

The requirements of world models, simulation, and edge inference also feed back into sensor, optics, and ISP design. As robotics foundation models and industrial software vendors move toward a common physical AI architecture, sensor and camera manufacturers need to anticipate how these developments will influence future imaging pipelines.

What you’ll take away

• How the digital twin first and vision agent approaches bring physical AI into industrial environments.
• How world foundation models, synthetic data, simulation, and validation contribute to robot development before deployment on real hardware.
• How existing camera networks can support automated inspection, worker safety, multi-camera tracking, and operational insight.
• How sensors, optics, ISPs, sensor bridges, and edge compute fit into the physical AI loop from data capture to deployment..