Digitization, artificial intelligence and virtual reality do not stop at car production either. We spoke to Marc Kamradt, Senior Expert Innovation at BMW in Munich, and had his vision of the car production of the future explained to us.
c’t: With the LabelTool Lite app developed by BMW and published as open source, every employee in production should be able to train their own artificial intelligence (AI) for quality assurance. How can you imagine that?
Marc Kamradt: First of all, the LabelTool Lite helps to create training photos and to provide them with a suitable category. For example, an AI trained by our employees recognizes different types of door sills mounted in the door frame, even if they have protective film on them – and sounds an alarm if the wrong model has been installed. The LabelTool Lite quickly spread throughout the group and was also accepted outside of production. We have therefore expanded the approach and developed a complete AI pipeline so that the individual steps can now run fully automatically if desired.
Once the AI has been trained, our new evaluation tool is used: it tests the network and, if necessary, initiates new training. This step has proven to be particularly important, as the requirements and test criteria can vary greatly depending on the application. Sometimes it depends on the exact position of a part, sometimes just whether it is there at all.
c’t: How long did it take to train this door sill recognition? How many pictures did your employees have to create and label in total?
buddy: The actual training of the AI takes less than an hour. Our employees can use a GPU cluster that is specialized for such tasks. We don’t need more than five images per door sill. Manual categorization takes less than 20 minutes in total. The AI pipeline then optimizes this data and supplements it with suitable synthetically generated images and labels. For the employee, this step runs transparently and without manual effort.
c’t: Can you imagine the LabelTool Lite as a pre-trained image recognition system that your employees can then refine with suitable photos for specific tasks? What material was it pre-trained with?
buddy: Indeed, our system is based on specific, pre-trained neural networks, with the training relying heavily on photorealistic synthetic data. We have a digital twin of the vehicle for every production step. From the CAD data of this twin, we can create an almost infinite number of photo-realistic and already labeled images. Supplemented by real photos, a perfect combination is created with which our employees can train an AI for their own purposes without much effort.
c’t: That means you have an enormous wealth of production-specific data, comparable to what Google or Facebook have set up in the online business?
buddy: Our own synthetic training data covers the entire vehicle production and logistics, including the various installation states of our vehicles. We share part of this with the open source community as a SONDI dataset. SORDI stands for Synthetic Object Recognition Dataset for Industries. It will be the world’s largest and most realistic open source data set for the industrial environment. The use of SORDI will revolutionize the use of AI in the industrial environment and will enable us to do many things in the future that were previously considered unfeasible.
»SORDI will be the world’s largest and most realistic open source data set for the industrial environment.«
c’t: What other tasks, also apart from quality assurance, could your employees delegate to a self-trained AI?
buddy: Quality assurance is of course still a strong focus. Another example from this area would be checking the plug connections of cables or hoses. But thanks to SORDI, we were also able to tackle completely new tasks, especially in logistics. There, for example, the manual effort was extremely high to train neural networks to precisely determine the filling level of transport boxes, containers or shelves. Thousands of photos had to be manually categorized to reflect the incredible number of possible variations. Using SORDI and the AI pipeline, we can now automatically generate hundreds of thousands of such images including labels at the push of a button. Every possible case, every conceivable combination, including different lighting conditions, etc., is taken into account and covered by SORDI. The employee can automatically load this data into LabelTool Lite and start training immediately without any further manual effort.
c’t: You publish the LabelTool Lite and SORDI free of charge as open source. Does BMW benefit from this too?
buddy: The approach of making our tools available as open source has turned out to be very successful. Our community is constantly growing, with large companies such as Microsoft and Google joining us and actively supporting us in development. With the door sills alone, we were able to reduce the time required for the employee to create the AI by more than two thirds thanks to such suggestions for improvement, with greater user-friendliness and even better quality. We will also publish the evaluation tool mentioned at the beginning on our GitHub page and share it with the community.
c’t: The cockpit of the BMW iX was developed using a digital twin, which in turn was generated using AI. What kind of AI was used and what exactly was it trained for?
buddy: An AI-based method kit called “M.OPT” was used for the support structure of the BMW iX. It is characterized by the fact that only a few training examples are required to teach the AI. To achieve this, M.OPT combines deep learning and classic machine learning approaches. With the “M.OPT warpage” method used here, all process parameters as well as many tool and geometry parameters of the component – for example wall thicknesses – can be taken into account in the optimization. This made it possible to reliably produce a sufficiently precise component while still saving a considerable amount of material and time in the approval phase and thus being significantly more sustainable.
c’t: Which part do the engineers determine?
buddy: With M.OPT, the engineers define the basic concept of the component and the variance of the individual parameters. In this way, they have an influence on the constructive design and ensure the feasibility in the planned processes. In this way, the physical properties of the component or complete assemblies can also be optimized in M.OPT. For example, the efficiency of an electric motor can be increased – and at the same time the weight can be reduced, the service life extended and the distortion of the injection molded components minimized.
Gas and oil are getting more and more expensive. In c’t 20/2022 we therefore draw attention to cheap and ecological alternatives with and without replacing the heating system. We’ll also show you how to protect yourself from trackers with the Raspi, test hacker tools, smartphones and graphics cards, and talk to Leica about cameras. You can read that and more in the current issue of c’t.