AI in Intralogistics and Supply Chain
In episode 205 of the Industrial AI Podcast, WITRON CEO Helmut Prieschenk and Franziskos Kyriakopoulos, CEO 7LYTIX were guests. Below is a summary of the most important statements with comments from the editors.15 Sep 2023
LLMs in logistics?
"Of course, we at WITRON have already dealt with LLMs (Large Language Models). However, I plead for a certain serenity. The world will not end through their use - and we are continuously checking whether such tools are suitable to help our customers or our developers in a meaningful way in the implementation of concrete customer requirements," explains Helmut Prieschenk from the Upper Palatinate.
Franziskos Kyriakopoulos agrees with him, but already outlines applications. "LLMs are good at processing sequences - orders, debits, sales or customer communications. That can be used in intralogistics as well." He agrees with Prieschenk, however. "There's a lot of hype, a lot of influencers running around spreading half-truths."
Editor's note: Guests have a point there: Many companies associate LLMs primarily with language. That's true, but the tools can do much more. Code Interpreter from ChatGPT, for example, can parse data sets and takes a step toward AutoML, simplifying the work. Automators use LLMs with special prompt support to make programming easier. Several companies presented such approaches at HANNOVER MESSE 2023. The great success of ChatGPT also results from the good user experience of the application. For the industry, the interplay between AI and UX will be crucial in the coming years. "If I were to bet, it would be on the following: Innovative industrial companies will finetune available LLM's (like LLama from Meta or Alpaca) run on their own data and PoC's in cooperation with research institutions. Of these, many will fail, but those that make the PoC > MVP > Pathway will have an awesome industrial product. Knowledge Bases and multimodality are the keys. And this is exactly the idea we are trying out soon with a digitalization partner and an industrial company," explained Franziskos Kyriakopoulos when asked by the editorial team. Generative AI, such as GPT or the DALL-E tool, will also change industrial product development - the designer receives support from an intelligence. Festo has been working in the area of reinforcement learning for manufacturing processes for several years. The next step is the use of generative algorithms for product development. OpenAI recently published 3D models for DALL-E. The challenge in industry besides the 3D challenge: products must be able to move. And Jan Seyler from Festo Research asks the provocative question: can machines build machines? He is certain: "Many aspects of engineering can be supported by AI methodology. The challenges are mainly semantic, functional understanding and accurate simulation." And Festo CTO Ansgar Kriwet defined the future of a Festo product in simplified terms at a conference: A multipurpose hardware as a software defined product packed with AI.
Exciting, the customer perspective
Prieschenk: "Our customers are not originally looking for a new "tool", they have a problem and need a working solution for it, which optimizes the logistics process in the distribution center or in the supply chain, works stably in practice and can be integrated sensibly into an evolved structure."
But isn't this sobriety holding us back in Germany and Europe? "I already need an ROI," Prieschenk vehemently emphasizes. "LLM developers have a burn rate of $500 million per year and need another few billion," Kyriakopoulos reports. "That would be unthinkable in Germany or here in Austria."
Editor's note: What Helmut Prieschenk describes in the interview applies to many machine builders. AI in industry must deliver ROI and also be easy to integrate into the existing system landscape. That sounds banal, but it is a challenge for many companies. AI in industry is also always a bit more difficult than in a consumer application because existing processes have to be taken into account, because different parameters, 24/7 requirements come up for the companies. Therefore, robustness of algorithms plays a special role. Models must integrate with systems and also take physics into account. Tip: A few weeks ago, scientists at Helmut Schmidt University, Hamburg, analyzed which deep learning models fit cyber-physical systems (more here https://arxiv.org/abs/2306.07737 ).When it comes to AI and machine learning, industry demands reliability and robustness, 24 hours, seven days a week. Boris Scharinger, Senior Innovation Manager at Siemens Digital Industries, coined the term for Siemens: "Industrial-grade AI": this is a challenge that all providers must face. At the Siemens AI Lab in Munich, a team advises customers on AI strategy and possible applications and conducts research on robust algorithms. But Scharinger cautions that you can't do everything alone. There are many examples and approaches for successful collaborations: For example, AI algorithms are undergoing a baptism of fire at Siemens plants, while other startups are relying on Siemens industrial hardware for their solutions. In addition, Siemens can take on the MLOps part in a startup partnership. This means that the Munich-based company will distribute AI models in production, monitor them, and update them as needed. Without an MLOps strategy, machine builders and automation companies will not be able to develop a sustainable AI business model. Even today, many machine builders are struggling to offer digital services to their customers. There is a lack of skilled workers, ideas and infrastructure. It is the hour of the MLOps infrastructure providers, domain knowledge companies.
The difficulty with optimization
7LYTIX developers work with LLMs, but the focus is on demand forceasting. "We can deliver added value, but some companies often don't understand at the beginning what the added value of the model will be. More revenue from better communication to the customer or lost sales? Many can't calculate that. That's where they need help from us," says Kyriakopoulos. Prieschenk adds: "Our WITRON customers can calculate very well and have perfected their business over decades. But I understand what Mr. Kyriakopoulos means: First of all, it must be clarified where to optimize in the first place. The retailer asks himself, do I want to optimize the supply chain network, the warehouse size, be closer to the customer, minimize lead times, change delivery cycles, reduce food waste and stock-out or have less inventory in the warehouse. In this respect, we have learned a lot together with our customers from different parts of the world. Also that the requirements for holidays in Finland are different from holidays in the U.S., or that a Monday holds different requirements than a Thursday." Kyriakopoulos agrees. "We need a requirement first and then an AI tool for it. And we don't need Deep Learning everywhere."
Editor's note: Not everywhere needs Deep Learning - this is a very important statement. This is also our experience from many projects. And: many companies indeed always find it difficult to find a business case for AI in industry. That's why we need application examples, best practices and an exchange of experiences like at HANNOVER MESSE.
How does demand forecasting work?
"In the beginning, we have to get an overview of the data. This is a tedious job for many retail companies. It is not only about stock goods, but also how many goods are in the store, how much was sold, what influencing factors like promotions are there, how many "lost sales" do I have in the store and much more" explains Franziskos Kyriakopoulos. In addition, there are customer cards, seasons, the location of the store or promotions. "And we need to know what's in the distribution center, in the backroom of the store, in the trucks on the road, because optimization doesn't end in the store.
Similarly, it's important to avoid cross-group or cross-departmental restrictions, as well as data lakes. For the most part, much of the necessary data is known, but different departments unfortunately pursue different interests." "The data flows into very simple models" Kyriakopoulos continued. The baseline is people's experience. This is not yet AI. We talk about regressions. Then we ask ourselves, have we gotten better. Time series analysis follows, first machine learning methods. We always have to look at how much accuracy do we achieve through the next stage versus how much is the added value for the customer and user."
Opinion: the podcast illustrates very well the industry requirements for AI and what challenges companies and their customers are facing. Robustness, 24/7 reliability, integration capability, good UX as well as the business case must be solved. Industrial grade AI is the goal.
Interested in news about exhibitors, top offers and trends in the industry?
Your web browser is outdated. Update your browser for more security, speed and optimal presentation of this page.Update Browser