"Karl" moves the ball: Two decommissioned industrial robots demonstrate how machines can learn independently.

University of Paderborn Modernizes Two Systems for Teaching

The University of Paderborn, as part of the project "Articulated Arm Robots in Teaching" (Karl), has modernized two existing industrial robots and put them back into operation for use in university education. Scientific staff and students from the Heinz Nixdorf Institute equipped the robot systems with new sensors and integrated them into simulation and control environments.

The goal was to combine practical applications of robotics with methods of artificial intelligence. According to the university, the systems are particularly intended to support the teaching of machine learning methods.

Kawasaki Robot Receives New Control Interface

The project focused on a six-axis industrial robot from Kawasaki and a seven-axis Franka Emika Panda. Both systems were temporarily unused and were made operational again at the Chair of Control Engineering and Mechatronics. The work included assembly, repair, and commissioning as well as integration into modern software environments.

For the Kawasaki robot, the project team first examined the control in a simulated environment. Subsequently, the participants developed their own interface through which the system could be controlled from external software. Building on this, they were able to plan motion sequences, transfer them to the real robot, and evaluate the executed movements.

The technical revision was carried out with the involvement of scientific staff and student assistants. Three student theses were used to delve deeper into individual aspects of the development. A key project goal was to make existing systems usable for new applications with minimal effort.

Reinforcement Learning for Autonomous Robot Movements

Another focus was on the use of reinforcement learning. In this machine learning method, an algorithm improves its decisions through interaction with an environment. To do this, a learning agent selects certain actions based on available information and then receives feedback on their outcomes. Based on these evaluations, the system gradually adjusts its behavior.

In the "Karl" project, the team examined how such learning methods can be transferred to real industrial robots. The seven-axis Franka Emika Panda served as the test system. One application case dealt with path planning. Here, reinforcement learning was intended to replace the sample-based planning method Rapidly-exploring Random Tree. This method is used to determine possible movement paths in a given space.

In another experiment, the project participants used the Soft-Actor-Critic algorithm to control the robot while balancing a ball. A camera captured the position of the ball on a plate. The learning agent then determined the necessary robot movements to tilt the plate in order to move the ball from the edge towards the center. With these experiments, the participants investigated the practical use of learning-based control methods on real robotic systems.