Building and Scaling AI Factories With Digital Twins and Robotics

In this talk

1. 1:11

   A four-part factory method

   Leo lays out four stages: create a geometry foundation, simulate production, connect enterprise and shop-floor data, then use that data for AI-supported operations. He says this approach also supports energy management and carbon reduction. Before construction, Foxconn uses simulations to assess processes and equipment; after construction, it connects live factory data to the digital twin.

2. 5:22

   Build the geometry, then simulate

   Foxconn converts CAD and BIM assets into USD models for Omniverse, using equipment lists and an internal extension to assemble production lines. For planning, the team uses FlexSim to study cycle time, throughput, and staffing, including a pod-room case where simulation identified four CTUs as the preferred quantity and the exchange station as a bottleneck. It also simulates AMR traffic, humanoid robot behavior, cooling, and camera placement before deployment.

3. 16:23

   Connect factory systems to AI operations

   Foxconn connects equipment IoT, PLC, MES, and shop-floor systems so teams can monitor factory status through an Omniverse portal. It also combines manufacturing records with manuals, maintenance procedures, and quality reports to support machine-level questions and troubleshooting suggestions from an AI agent. The speaker describes MOM as the shared data platform for work orders, materials, equipment status, dashboards, and planning.

4. 23:41

   Train robots and improve them through use

   Foxconn collects robot demonstrations with spatial-computing headsets and supplements them with synthetic data from Omniverse and Isaac Sim. It combines NVIDIA’s GR00T with reinforcement learning and other models, while acknowledging that a basic pick-and-place task can take 10 to 15 seconds compared with about two seconds for a person. The company plans a cloud skill library and a data flywheel; deployed robots still need monitoring, human takeover for failures, and more work for complex assembly.

5. 30:02

   What scaling requires

   Leo summarizes three operating principles: establish accurate geometry, simulate automation before installation, and organize operational data for AI use. In the Q&A, he says Foxconn has about 20 Omniverse software engineers and that useful sim-to-real results require substantial parameter tuning. He also notes that the humanoid work remains early-stage and that supplier access to equipment data depends on the company and project.