BMW Group Trains Factory Vision Models on SORDI Synthetic Data and NVIDIA DGX, Cutting QA Model Setup Time by Two-Thirds

BMW Group has cut the time its employees need to set up AI-based quality assurance by more than two-thirds by training models on synthetic images generated with NVIDIA DGX systems, according to an NVIDIA customer story. Employees now build simple visual inspections themselves with a no-code tool, and DGX systems that began in research now run in production. "The time it takes for our employees to implement AI automation in quality assurance has been slashed by over two-thirds," a BMW Group IT leader said. BMW has used AI in manufacturing since 2019, but training data was the bottleneck. With highly customizable vehicles, thousands of photos had to be categorized by hand to cover production variations, for tasks from inspecting parts to judging the fill level of transport boxes and shelves. The BMW Group Technology Office in Munich built SORDI (Synthetic Object Recognition Dataset for Industries), an open-source dataset of more than 800,000 photorealistic images in 80 categories, from pallets to forklifts. DGX systems first train the deep-learning models that generate synthetic data from simulated factory scenes, then train the production models for object detection, image segmentation, classification and 6D pose estimation, and finally test and evaluate them; NVIDIA TAO with AutoML scripts handles hyperparameter optimization. On top of this sits LabelTool Lite, a pretrained image-recognition tool that employees refine with their own photos: training a door-sill check takes under an hour with no more than five images, after which the system raises an alarm if the wrong sill is fitted. It also detects missing or wrongly colored stitches in leather. BMW reports that DGX lifted data-scientist productivity eightfold and delivers 4-6x the performance of its earlier systems, with training datasets exceeding 500,000 images. SORDI also stores attributes such as an object's CO2 footprint and energy use, which BMW uses in DGX simulations to optimize energy and CO2 savings in its factories.