How to Integrate AI Vision Into an Existing Production Line

To integrate AI vision and robotics into an existing production line, add the system as a module instead of redesigning the line. Define the problem and the current operating costs, assess the factory environment, test the system in offline simulation, pass a factory acceptance test (FAT), run a pilot, then move to production, with the shutdown window planned in advance.

This sequence comes from Lê Thái Sơn, Project Manager at Solomon Technology Corp., who has worked for ten years on AI vision and robotics for factory automation. He discusses it in the Q&A of the podcast episode AI-Powered Robotics: How Physical AI & Embodied AI Will Transform How We Live and Work. This guide organizes his answers into a practical path and notes where the episode gives no detail.

Most businesses do not build new factories from scratch, and Sơn says most AI vision and robotics systems are built as modules. A module is built into the factory's existing production line. It is rarely the case, he says, that a factory completely redesigns its layout or production process to accommodate the system being added.

That shapes the whole approach: the new system has to fit what is already running.

Before choosing technology, Sơn's advice is to start with the customer's actual problem, not with "What can we do with AI?"

Costs come next. For the process being replaced, he suggests looking at what the factory currently spends on:

These figures give the actual current operating cost. Sơn is open about the difficulty: estimating payback is hard because AI and robotics projects come with high costs. The episode gives no payback benchmark or timeline, and this guide does not either.

Sơn recommends looking at real conditions in advance, not relying on lab results. The factors he lists are temperature, humidity, dust, smoke, vibration and surrounding magnetic fields. He adds that there is a lot to consider, and that the more of it a team can anticipate, the more it can reduce the cost and risk of deploying an AI system in a factory.

The reason is that conditions in a factory differ from the lab. In the main talk, he says lighting may change, cameras may vibrate, the environment may be dusty, products may vary, and conveyor speed may not be constant.

Offline simulation, in Sơn's description, means testing the system separately first. It connects to the idea of a digital twin: running the system in a simulation that reflects real-world conditions. Once a problem is solved in simulation, he says, deploying in the real world involves far less risk.

A FAT is a factory acceptance test. Its purpose, per Sơn, is to ensure the system operates reliably before it is integrated into the production line.

Sơn names the pilot and production as the final two stages. He does not describe them in detail, so this guide does not either. What the pilot should measure, and for how long, depends on the line and is best agreed with the team doing the integration. That last point is my suggestion, not his.

Sơn also notes that the time needed to deploy the system on the line has to be estimated, so the factory can plan an appropriate production shutdown. His broader point, in answer to a question about avoiding extended downtime, is that the staged approach above is how a team gets to a predictable deployment window.

Sơn advises starting with small problems that deliver clear value. One example he gives: a process that needs five workers to inspect products could use AI to support inspection of product appearance and other easy-to-detect characteristics, with a low risk of error. Fewer workers, perhaps two, could handle that part, and the rest would take on more detailed inspections. Another is loading products into CNC or cutting machines, where accidents occur frequently. The five-to-two figure is an illustration and not a reported result.

The episode gives no timeline. Sơn says payback is difficult to estimate because these projects cost a lot. His starting point is to calculate current operating costs for the process being replaced, covering labor, electricity, operation and maintenance, and material wear and consumables.

According to Sơn, yes, in most cases. These systems are generally built as modules that fit into the existing line, and factories rarely redesign their layout or process to accommodate them.

He lists offline simulation, a factory acceptance test (FAT), and a pilot before production. Offline simulation tests the system separately, and the FAT confirms reliable operation before integration.

Not yet, in Sơn's view. He recommends continuing to invest in conventional robotics and integrating newer technologies gradually as they develop.

This article is based on Lê Thái Sơn's Q&A and talk in the podcast episode AI-Powered Robotics: How Physical AI & Embodied AI Will Transform How We Live and Work