AI for Laser Welding Robots

How T-Systems teaches industrial robots in automotive manufacturing to learn

February 25 2021Dr. Florian Pinsker

Optimizing quality and efficiency of manufacturing processes

In a joint research project with a premium car manufacturer, we have managed to show that a laser welding robot can recognize its own behavior and independently improve according to indicators it has been provided with.

The next development stage of industrial robots


A robot does what an engineer has tasked it to do. These days, it is usually the case that in practice defined parameter records are produced out of a few test runs which uniquely define the operational behavior of the robot. The best record is selected in accordance with certain KPIs and implemented in the subsequent operation. Currently, the 'right' parameter records are found by carrying out a few test runs with the laser welding robots under the instruction of an engineer.
In everyday operation, however, a robot's environment in a factory is subject to change and the tasks of a robot may vary, too. Imagine if the welding robot autonomously optimized its behavior accordingly so that the engineer only needed to tell the robot what the requirements are for a component or a weld seam? In this scenario, the robots would then find a systematic way to meet the requirements in various situations by themselves.

From AI architecture to the fully autonomous factory


Fig. 1 Architecture for a self-learning laser welding robot

The optimum functionality of robots is unachievable with the current methods, the process of identifying parameters is inefficient, and the robot's actions are not intelligent, so it cannot react to new situations and learn from them.

We therefore want to take a leap forwards. Together with the car manufacturers we have developed and tested an AI architecture which first enables the quality of a weld seam to be detected, and then evaluates it in order to then assess the laser welding process of the robot. This allows us to enable the laser welding robot to identify the parameter automatically, which is more efficient than any previous method, while also enabling a situational understanding of various scenarios. This AI architecture is represented in Fig. 1.

The concept developed also enables robots to learn together and to continue learning during operation. This massively expands the robot's wealth of experience of the weld seam guidance and allows it to integrate and also document this in a systematic way. This functionality gets us one step closer to the fully autonomous and self-improving factory.

How machines learn to learn

The topic of artificial intelligence is viewed differently in different cultures and the perceptions extend from AI as the solution to automatically providing for humanity to concerns about the surrendering of skills and the reliance on systems which humans cannot understand and are superior to them. Even if in our project we have not been able to solve this philosophical question, the challenges which AI brings are an important topic which needs to be kept in the back of our minds. Nevertheless, it can be said that today's systems still only act in a very limited sphere. 

I like the view taken by Japanese popular belief in which humans, animals, objects, and even robots have the opportunity to be animate. Artificial intelligence follows this path by enabling robots to recognize (often realized using Deep Learning), to evaluate and to act, which can be accomplished through reinforcement learning. In this spirit, we at T-Systems are true 'digital enablers' because we enable machines, or more specifically welding robots, to act intelligently, and to continuously improve. Learning new skills and optimizing existing skills is a continuous and iterative process.

Productive cooperation 

This integrated intelligence ultimately allows the laser welding robots to act in a more efficient, fail-safe, cost-effective, and flexible way. One of the reasons for our fruitful cooperation with the car manufacturer, alongside our shared scientific curiosity, is the prospect of increasing productivity while reducing accrued costs.

Finally, I would like to put the following down on record. The work done in collaboration with our partner on the latest AI technologies was exceptional. We were able to show that, in the laser welding use case, the combination of deep learning and reinforcement learning provides laser welding robots with a significantly renewed functionality and thus an improved laser welding process. The new method increases quality, cost efficiency, safety, and productivity and can be scaled up with the number of robots. Finally, one thing that should be highlighted with regard to this project is that T-Systems is available to top clients as an innovative service provider, particularly in the field of AI.

About the author
Dr. Florian Pinsker – Senior Consultant & Data Scientist

Dr. Florian Pinsker

Senior Consultant & Data Scientist, T-Systems International GmbH

Show profile and articles
Do you visit t-systems.com outside of Germany? Visit the local website for more information and offers for your country.