The manufacturing sector generates $2.3 trillion in annual U.S. GDP, yet most shop floors still rely on decades-old automation and paper manuals. That challenge is precisely what MaVila, short for “Manufacturing, Vision and Language”, was built to address. Developed with funding from the U.S. National Science Foundation (NSF) and led by California State University, Northridge professor Bing Bing Lee, the vision-language model can “see” a machine, identify defects, and suggest optimal process parameters in plain English, all in real time.
“Mavila is an advanced vision and language model specifically designed for the smart manufacturing domain as we have been working in this domain for almost ten years,” said Lee in an NSF interview.
What MaVila Does Best
Most industrial AI still relies on rule-based vision systems that choke on new part geometries or glare from workshop lights. MaVila’s novelty is its domain-specific training: thousands of annotated images, manuals and sensor logs from CNC mills, wire-EDM cells and 3-D printers. Shown a titanium bracket fresh off a powder-bed printer, the model can label a micro-crack and immediately suggest new laser power settings. That task would normally require a process engineer on-site.
MaVila can look at a milling machine, spot a flaw and suggest a better cutting speed, all before a human engineer reaches for the stop button.
“Large language models are designed to process to generate the textual context,” Lee explained. “Where the [MaVila]
vision model can understand and process both the visual and the textual data.”
In early lab tests MaVila flagged seeded defects on 3-D-printed parts and generated new print parameters in seconds. The team then mounted the model on a mobile robot that photographed a milling operation, pulled the proper torque spec from a PDF manual and suggested a tool-path tweak, while operating live.
Lee argues that whoever owns specialized data will own the next wave of industrial AI.
“The core of the AI industry, I believe, will be the AI algorithm, the model you have,” he said. “Another big core value is the data. [Whoever] has the data will be the most valuable.”
His group has built one of the first public benchmarks for smart-manufacturing imagery, then fine-tuned MaVila with a retrieval-augmented pipeline so it can pull the right snippet from a user manual in mid-conversation.
Because public datasets rarely cover machine-shop edge cases, Lee’s group has spent years capturing high-resolution images of cutting tools, gathering sensor streams from 3-D printers and scanning operator manuals into a searchable corpus. The result is a retrieval-augmented pipeline that lets MaVila pull the right output while a machine is still running.
AI Increasingly Making Its Way onto the Factory Floor
Siemens is already commercializing a similar product. Its Industrial Copilot, connected to the company’s TIA Portal, can draft PLC code and HMI screens, trimming engineering hours and cutting error rates. The tool goes live this summer.
NVIDIA is helping Foxconn, Pegatron and others build full digital twins of their factories. By simulating layouts before steel is cut, those firms claim faster launches and safer lines. Similarly, BMW says its Virtual Factory now reduces planning costs by up to 30 percent after testing every robot move in a photorealistic model.
Together, these moves show big manufacturers warming to AI copilots, but they also raise the bar for smaller suppliers that lack deep pockets, or data.
Policy Tailwinds, Talent Headwinds
Federal money is flowing to US manufacturing. This spring the NSF earmarked $25.5 million for “future manufacturing” grants spanning digital twins and “recyclofacturing”, a closed-loop approach that turns post-industrial metal scrap straight back into fresh products, eliminating most down-cycling and cutting embodied carbon. Yet a Deloitte study warns of a 2.1 million-worker shortfall in U.S. factories by 2030, one reason Siemens pitches its copilot as a fix for skilled-labor gaps.
However some caution that no amount of clever modeling removes the need for fresh, proprietary data. Mid-size suppliers will have to decide how much of their know-how they’re willing to share with an academic group, or pay to keep private.
Lee concedes the point. His next milestone is signing pilot partnerships with at least three small-to-medium manufacturers. Without real production data, the model could stall in academia.
Whether MaVila scales now hinges on whether factories will share enough images and logs to keep the model sharp, and the deployment ease of the model.
If those hurdles fall, American plants might trade bulky binders for an AI assistant that sees a problem, talks it through and fixes it on the fly. That would turn today’s static automation into living, learning production lines , and perhaps give U.S. manufacturing the edge it has been chasing for years.
