Factories of the future: Julian Eßer on reinforcement learning in robotics

In a fast-moving world characterized by constant technological progress, transformative changes are also taking place in robotics. Intelligent robots are increasingly able to perform complex tasks with previously unimaginable precision and autonomy. These developments mark the beginning of a new era. Julian Eßer, researcher in the field of robotics, shares his expertise on current research, trends and future prospects in conversation with AI Grid Communications Manager Franziska Peters. He discusses how robots are being trained, what opportunities new AI methods are opening up and what challenges and fears are associated with them.

About JulianJulian Eßer is a researcher at the Fraunhofer Institute for Material Flow and Logistics (IML) in Dortmund, coordinator of the Embodied AI research area at the Lamarr Institute for Machine Learning and Artificial Intelligence and a member of the AI Grid. His research focuses on the behavioral control of dynamic robots and the further development of reinforcement learning.

How did you get into robotics and what fascinates you about AI?

My enthusiasm for robotics arose during my mechanical engineering studies and deepened during my master's degree in robotics and computer science. I am fascinated by how robots find their way around their environment, plan tasks and interact with their surroundings - all made possible by AI. Today, we are moving away from pre-programmed industrial robots towards flexible, intelligent systems that can react to new things and act independently. For me, the appeal lies in integrating AI into the physical world to develop intelligent robots that are not limited to one industry, but can be used in many areas.

"It is important to integrate robots that are able to act in a similar way to humans with 'common sense'."

What progress do you see in robotics, especially in industrial applications?

Robotics, especially in combination with AI, has made considerable progress. At Fraunhofer IML, we are working on automating logistics processes through the use of intelligent robots. This development also extends to production and other branches of industry. The rapid transfer of research results into practice is driven by strong interest from industry. Industry is showing great interest in robotics, which opens up many opportunities. At the same time, we have to deal with the development of new safety standards. It is important to integrate robots that are able to act in a similar way to humans with 'common sense'.

How do you tackle the challenges in real-life applications?

In the real world, there are countless variables that need to be taken into account in robotics. Rule-based, hard-coded processes are not enough when robots interact with humans in dynamic environments. We use machine learning, in particular reinforcement learning, to help robots make decisions. A key aspect of our work is simulation-based AI, which enables us to simulate physically realistic environments and thus prepare robots for unpredictable situations.

"The robots learn through a process of trial and error, reacting to unforeseen events and challenges in these simulated worlds."

Can you describe the training process in the virtual environment in more detail?

The training process of our robots takes place in a sophisticated virtual simulation designed to recreate a variety of realistic scenarios. This simulation continuously generates new, randomized environments - from simple obstacles to complex interactions with humans. The aim is to create an experience that is as close to reality as possible. The robots learn through a process of trial and error, reacting to unforeseen events and challenges in these simulated worlds.

Instead of training the robots on a limited number of predefined scenarios, domain randomization enables the generation of an almost infinite number of situations. In this way, the robots not only learn to master specific tasks, but also to react flexibly and adaptively to new and unexpected situations. This approach helps the robots to develop a deeper understanding and better adaptability, making them more robust and reliable in the real world.

During this training process, the robots gain valuable experience that makes them smarter and enables them to make more effective decisions. This intelligence and learned skills are not limited to a single machine; they can be transferred to other robot platforms, making the entire training process scalable and efficient.

"We are developing a hybrid learning concept in which we integrate expert knowledge, natural common sense and mathematical rules directly into the learning process."

Can the learning process of robots in a virtual environment be compared to the learning process of a small child?

The learning process for robots is actually similar in many ways to trial and error for small children. Children learn to walk by falling and then trying a little differently. It's similar with robots. They learn to move from A to B in the virtual world, pick up packages and hand them over to humans. They try out different methods and receive a reward when they perform their task well. This learning process is an ongoing cycle that enables robots to continuously make better decisions and perform their tasks more efficiently.

"We are moving away from static robots towards more intelligent, versatile models."

How do you see the future development of robotics in terms of costs and integration into everyday life?

Robotics is constantly evolving and becoming increasingly affordable. As it scales and expands to different application areas, prices are falling. We are moving away from static robots towards more intelligent, versatile models. In the near future, robots could become a normal part of our lives, not only in industry but also in everyday life. The technology is becoming more and more accessible.

Thank you, Julian, for the insightful insights into your work and the future of robotics. Are there any final thoughts you would like to share with us?

Thank you for the interview. I believe we are at the beginning of an exciting era in robotics, in which AI will play a central role. The future will bring versatile, intelligent robots that will play a role not only in industry but also in our daily lives. I look forward to seeing how our research continues to drive this development.

AI Grid Ambassadors

AI Grid Ambassador in Israel

AI Grid Ambassador in Great Britain

AI Grid Ambassador in Sweden

AI Grid Ambassador in Switzerland

AI Grid Ambassador in the USA (West Coast)

AI Grid Ambassador in Japan (Tokyo)

AI Grid Ambassador in Spain

AI Grid Ambassador in the USA (East Coast)

AI Grid Ambassador in France

Professor an der Universität TU Dortmund