If you ask intralogistics expert Andreas Engelmayer about artificial intelligence at CIM, he can't help smiling. “The expectations are definitely high,” he says. And Andreas certainly knows what he’s talking about, since he's currently working together with a team of AI researchers from TUM on building an artificial neural network for intralogistics. “We’ve already had some very exciting results,” he adds cautiously.
The project has been running for about a year now in cooperation with the Chair of Materials Handling, Material Flow, Logistics at TUM. The goal of the collaboration is to design self-learning software based on a neural network model, for use in the field of intralogistics. The algorithm at the heart of the neural network was initially developed by the researchers at TUM. “There are a number of different parameters that need to be defined when designing a neural network,” says Andreas.
The network processes the data based on these parameters – they’re a kind of black box for CIM, in other words. This means that it’s never quite certain how the AI technology will behave in a specific scenario. The scientists at TUM therefore need data as well as multiple test runs to allow them to evaluate the results and adjust the parameters. “The AI performance improves with every test performed,” continues Andreas.
Mass testing as a basis for the initial tests
In order to use AI in the context of an intralogistics software system, a suitable framework had to be developed to feed data into the neural network. “That’s where I came in,” explains Andreas, whose responsibilities at CIM include test automation. “One aspect of our testing strategy is a mass test where the system is put through its paces with a high volume of orders,” he goes on to explain. Andreas adapted CIM’s mass test to provide a framework for the initial AI tests carried by the researchers at TUM. The test framework is capable of generating 100,000 plus orders for processing using AI. For the neural network to work meaningfully with the test framework, there are some additional criteria that needed to be provided. The articles in the generated warehouse have data containing certain probabilities for stock putaway or retrieval. “Let’s take wellington boots and winter hats, for example,” says Andreas. “We specify a pick probability of 70% and 50% in this case so that the neural network can recognise the differences and handle the articles accordingly.” The AI is also familiar with the essential features of the warehouse layout provided to the TUM research group by CIM. “This data is crucial,” comments Andreas.
All warehouse management systems are largely about speeding up goods issue and increasing intralogistics efficiency. “Travel distances in the warehouse are a key variable which needs to be reduced,” he continues. Optimised travel times are pivotal to enhancing the performance of internal logistics. “But it's not always that easy,” admits the CIM expert.
Promising data coupled with significant challenges
The difficulty on the one hand is to adapt the parameters of the artificial neural network algorithm with the goal of increasing efficiency. “That’s what applied research is all about,” says Andreas. “For us at CIM, the practical feasibility is more important,” he continues. Verifying whether neural networks do actually enhance performance involves more than just analysing the results from the AI testing. They also need to be compared with the existing options for optimising intralogistics. “You’ve got to be honest,” says Andreas. “What CIM offers with PROLAG World is pretty much the best on the market at the moment, in my opinion.” For the purposes of comparison, the efficiency analysis already used by PROLAG World in numerous warehouses is fed with the same data as the artificial neural network algorithm.
The neural network clearly delivers an impressive performance with regard to the efficiency analysis. “The potential of the technology is obvious, especially in terms of the correlations between the articles,” states Andreas. However, there are other factors that have a negative impact on the performance of the AI technology. “There’s the issue of storage capacity and neural network response time, for instance,” says Andreas. “Our old algorithm is simply less complex and sophisticated.”
Further tests and research promise new results
Andreas Engelmayer is not ready to speak of a final outcome to the project or even clear interim results. “We’re still very much at the start,” he admits. The coming weeks and months will be very interesting as more complex test data is applied. “So far we've only worked with a warehouse that’s not filled to capacity,” he says. “To get closer to real-life conditions, we’ll be working with high and very high occupancy rates. This will allow us to tell if AI still performs as well in terms of intralogistics efficiency.”
And so it’s a waiting game for now. A test usually takes two weeks and that’s before the data is analysed. Before any preliminary conclusions can be reached, more article data is required so that the results provide a realistic picture of intralogistics managed by AI technology.
If you ask Andreas what sort of results he’s expecting, he cleverly dodges the question and smiles. “It's really exciting to work with this technology. And I’m looking forward to seeing what’s in store for us over the next six months.” The research collaboration between CIM and the Technical University of Munich is set to continue until July 2023. So there’s plenty of time for further testing and optimisation. Watch this space.
