Deutsch Intern
  • Slider für Innovation Day Mainfranken 2024
  • none

Artificial Intelligence from Würzburg Controls Satellites in Orbit


A team from the University of Würzburg is developing an AI-based attitude control system that will allow satellites to maneuver autonomously. The new technology will be tested in space.

The qualification model of the InnoCube satellite. After training on the ground, the AI-based attitude controller will be uploaded to the flight model in orbit and tested there. (Image: Montenegro/JMU)

At the heart of the new attitude controller is an artificial intelligence that is being trained on the ground and will later be able to change the satellite's attitude in orbit on its own. It is being developed by two computer science departments at the University of Würzburg (JMU) using a process called deep reinforcement learning.

“We let our artificial intelligence interact with a simulator that imitates a satellite in orbit “, explains Kirill Djebko, a researcher at the Chair of Artificial Intelligence and Knowledge Systems (held by Frank Puppe) and one of the two team leaders. “The AI keeps giving this virtual satellite new control signals and learns from the results by trial and error. This will continue until it is able to implement various application scenarios on its own without errors”, adds Sergio Montenegro, Chair of Aerospace Information Technology and one of the team leaders.

New Standards for the Development of Satellite Control Systems

AI-based attitude control based on the Würzburg model could significantly shorten the development time of such systems in the future and thus significantly advance aerospace technology. “The production of an attitude controller currently requires extensive testing and adaptation loops, which take up a lot of time and human resources”, says Djebko. “This effort could be minimized with the help of a self-learning algorithm.”

The Würzburg technology has another advantage: “Sometimes, attitude controllers for satellites in orbit need to be recalibrated because the expected conditions differ from the actual ones or because physical parameters change”, says Montenegro. “With conventional controllers, this is very cumbersome, especially because of the calibration process mentioned above – our AI could also speed this up.”

The researchers have already shown that AI-based attitude controllers are capable of making such adjustments on their own: As part of the “VeriKI” project between the University of Würzburg, the Research Centre for Information Technology (FZI) in Karlsruhe and Gerlich System and Software Engineering (GSSE), the computer science chairs at the University of Würzburg have already developed a simple AI-based attitude controller that was capable of handling variations in a satellite's moments of inertia and evaluated it simulatively. Now an AI-based attitude controller is to be tested in orbit on board a real satellite.

Put to the Test at an Altitude of 500 km

The AI agent trained on the ground will be tested for the first time in 2025 – directly in orbit on a small satellite called InnoCube, developed by the University of Würzburg in cooperation with the Technical University of Berlin and scheduled for launch in October 2024. It will serve as a platform for scientific experiments and technological demonstrations in space and will host the AI agent once it has completed its primary objectives.

The LeLaR project (short for In-Orbit Demonstrator Learning Attitude Control) was launched on 1 July 2024 and is funded with approximately 430,000 euros by the German Federal Ministry for Economic Affairs an Climate on the basis of a decision by the German Bundestag. The project is managed by the German Space Agency at the German Aerospace Center (DLR).


Dr. Kirill Djebko, researcher at the Chair of Computer Science VI – Artificial Intelligence and Knowledge Systems, Tel. +49 931 31-86405,

Prof. Dr. Sergio Montenegro, Chair of Computer Science VIII (Information Technology for Aerospace), Tel. +49 931 31-83715,

Additional images

By Sebastian Hofmann