Why cloud AI fails on the factory floor

This article summarizes an episode of Cognitive Revolution’s video series featuring Joseph Nelson, CEO of Roboflow.

Joseph Nelson, CEO of Roboflow/ Photo credit: Joseph Nelson

Analyzing live video in manufacturing environments requires immediate processing power. Joseph Nelson, CEO of Roboflow, a computer vision tools company, warns that relying on centralized servers causes delays that make the technology impractical for factory use.

Roboflow builds software that helps companies turn raw image data into models for tasks like object detection, tracking, and quality inspection.

To eliminate these delays, Nelson advocates for installing smaller programs directly onto local computers. By processing data on-site, companies guarantee fast and dependable reactions in physical spaces.

Rethinking cloud software in physical spaces

Applying standard text-based software to live video feeds quickly overloads central computers. Programs built to read text struggle to interpret the constant stream of raw visual data generated by a factory.

Nelson points out that language is a human invention, while the physical world is much more complex. He explains that a camera captures significantly more varied information in a single day than a text document contains. This makes visual analysis computationally heavier.

Organizing remote environments
Many industrial sites lack the internet capacity to transfer large video files to external servers. Resolving this connection issue requires hardware that operates independently.

Nelson observes that visual tasks are most critical in remote locations without human workers. In these environments, automated systems “need fast reaction times in addition to large-scale reasoning.” Operating completely isolated hardware on-site fulfills both requirements.

Dealing with long transmission delays

Scanning moving production lines using remote software introduces unacceptable lag. Every millisecond spent communicating with an external server increases the risk of missing a defective part.

Nelson highlights the danger of these limitations, noting that factories do not have the luxury of “waiting 40 seconds for a reply from a model.”

He estimates it takes over a year for new cloud technology to be compressed enough to fit on local devices.

Restricting the knowledge base
Engineering teams bypass this waiting period by severely limiting what the local software understands. Developers remove unnecessary data until the camera only recognizes the exact objects required for the specific factory line.

Operating in these restricted settings requires complete ownership of the underlying program. Nelson says, “You need to own the model. You need to have the weights. You need to put it into your environment.”

Condensing large models for local devices

Adding real-world understanding

Expanding automation to personal devices