Artificial intelligence (AI) applications are becoming increasingly concrete. The technology can help to calibrate robots and avoid disruptions in supply chains.
Robots are no longer just becoming increasingly popular in automotive production. This is especially true for simple solutions that are easy to configure for users. What is common to all systems, however, is that they have to be calibrated regularly so that they work precisely. This is exactly what should be simplified as part of the research and development project for AI-based robot calibration – Kirk for short. Specifically, machine-based calibration methods are used to develop software-driven calibration methods for industrial robots in order to improve their accuracy in operation.
Why is calibration important?
The robot systems usually have to be recalibrated individually at regular intervals in order to be able to work precisely even after a long period of use. This is costly and time-consuming. For small and medium-sized companies, in particular, this means considerable additional effort. In addition, especially in the case of inexpensive robot arms, due to mechanical reasons, even greater inaccuracies in the positioning potentially occur with increasing duration of use.
According to the Kirk development team, currently available calibration methods can mainly correct geometry errors. However, other factors relevant to the positioning accuracy, such as temperature or load-dependent inaccuracies, could only be compensated for insufficiently. A calibration during operation is currently not possible, although this could optimize the process.
The software takes over calibration tasks
Based on machine learning, the partners of the Kirk joint project now want to develop new software-driven calibration methods for practice. The project initiated in April was initiated by robotics expert ArtiMinds Robotics, the University of Stuttgart and the Baden-Wuerttemberg Cooperative State University in Karlsruhe. ” The possibility of immediately accumulating as well as evaluating information minimizes the effort for the individual and also makes it easier for SMEs, in particular, to develop the skills necessary to make optimal use of a robot system,” explains Darko Katic. He is the technical contact for the Kirk project and team leader for Artificial Intelligence at ArtiMinds. The aim is to increase accuracy through software algorithms so that robots can be used flexibly for a wide range of applications.
AI researcher Marco Huber from the Institute for Industrial Manufacturing and Factory Management (IFF) at the University of Stuttgart relies on machine learning with depth to make the complex relationships between external factors and the time-varying properties of the individual robot manageable and thus increase the positioning accuracy neural networks – so-called deep learning. The IFF is responsible for basic research together with the Robot and Human Motion Lab (Rahm-Lab) of the Baden-Wuerttemberg Cooperative State University in Karlsruhe.
Ultimately, the development team would like to transfer the results to real industrial use cases. In addition, the newly developed methods at the end of the project in spring 2022 are also to be integrated into the programming software Robot Programming Suite from ArtiMinds.
Increase reliability of logistics networks
In contrast, the logistics project Smecs (Smart Event Forecast for Seaports) has recently been completed. In it, the Department of Logistics at the TU Berlin, together with DB Cargo AG and the Kühne Logistics University, examined the use of machine learning with regard to the increasing demands on reliability and efficiency of industrial supply chains in order to be able to better synchronize their individual process sections in the future – especially also in the event of malfunctions. “Machine learning offers great potential for overcoming these challenges by recognizing inefficiencies and conflicts in the logistics networks in advance,” explains project manager Frank Straube, who is also a department head at the TU Berlin.
The project examined intermodal transport chains. Specifically, it was a matter of optimizing the various transport and handling processes of ocean freight containers in combined road-rail transport from the shipper to the seaport. For this purpose, an AI system was first developed that predicts the arrival times of transport orders at important process interfaces before the shipper leaves. Experts speak of the “Estimated Time of Arrival” – ETA for short. At the same time, the software detects malfunctions based on this along the logistical chain and issues suitable actor-specific measures.
Breakdown into sub-problems
In order to be able to process the diverse relationships, the entire chain for machine learning was broken down into various sub-problems, for which individual prediction models were developed. Examples of this are forecast models for road and rail transport as well as models for handling and shunting processes in the logistic hubs. For this purpose, the project team used various methods based on the supervised learning of AI systems, depending on the respective task. In addition, unsupervised learning methods were used in development. These are suitable, for example, to identify common types of faults and corresponding fault patterns.
In addition to the selection of suitable processes, Straube considers the success factor to be the identification and integration of suitable types of data for machine learning. For the use case in the Smecs project, historical data from a total of 15 different IT systems from different actors were included in the forecast for four years. A good 50,000 rail transports and 100,000 road transports came together.
The process-related sub-models developed during the project were finally integrated into an overall system that enables the calculation of arrival time (ETA) for a specific transport order. ” For a total of several days of the marine preparation, the discrepancies of the forecasts from the actual times for many orders amount to double-digit mins – also in the event of delays due to disruptions,” Straube said. This is very promising compared to previous forecasting solutions.
The online demonstrator shows how it works
In addition to the forecast function, the system developed as part of the Smecs project also supports decision-making. Depending on the ETA forecast, connection conflicts of the individual processes are automatically detected and recommendations for optimizing measures made available to the actors involved. The prototype developed in the project was made available to the public in the form of a web-based application. Users can use it to test the potential of machine learning using the example of selected historical, anonymized transport orders.