Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
ManufacturingNVIDIA Metropolis (VLMs)67% reduction in error rate
A video language model monitored an 8-step laptop assembly procedure in real-time, providing corrective feedback to operators when they missed steps like locking a cable tab, which reduced the effective error rate by 67%.
“they taught the VLM to recognize it at 99% accuracy. … they decreased their effective error rate by 67% on that line for that. Because instead of going from like assembly station, assembly station, assembly station, and then finally inspection, right, you're now almost have real-time feedback and this feedback is actually corrective…”
ManufacturingNVIDIA Metropolis (Visual AI Agents)
An AI agent performed real-time failure analysis by correlating visual inspection data (e.g., sealant overextrusion) with PLC settings and historical logs to identify the root cause of defects, minimizing downtime and waste compared to traditional end-of-line inspection.
“You can actually prompt the agent and the data to give you this information. So for failer analysis log reporting. But then you can also put an LLM and a rag retriever on it and say, hey, you're telling me there's an overextrusion of sealant. What could be possibly causing this?”
ManufacturingSiemens Industrial Co-pilot
An AI co-pilot deployed on the factory floor acts as a trusted colleague for operators, accessing machine manuals and PLC settings to provide step-by-step corrective actions for error codes, reducing the time needed to resolve issues.
“we're working with Siemens on industrial co-pilots for operations that are actually deployed on the factory floor … it understands the machine. It understands its manuals, it understands that. But he can also talk to the machine. It can go out and look at his PLC settings and all the other things that are happening.”
Customer supportDeepHow (AI Agents)
AI agents captured multimodal tribal knowledge from senior operators via video and documentation to create training materials and provide real-time assistance to new operators, allowing them to query specific steps (e.g.,
“they go in there and they look at the different SOPs, the way things work. They capture that in multimodality with video, and they're able to take that video alongside documentation alongside all these other things, and not only use it to create training material for new operators, but also assistance.”
ManufacturingNVIDIA (Digital Twin/Simulation)
Siemens used simulation and synthetic data generation within a digital twin to train an AI to identify unknown items and determine optimal picking points for pick-and-place robots, enabling end-to-end commissioning without physical prototypes.
“So a perfect example is actually on the show floor at the Siemens booth, where they did end-to-end commissioning of pick AI robots. So it picks, unknown items, so it actually doesn't know what it's going to pick. And they trained an AI to identify these objects and picked the best picking point for it. And that was all done virtually, right, in a digital twin.”