Title:
Combining the control of actuators with different time horizons for AI-based satellite attitude control using subnetwork policy
Vereinigung der Steuerung von Aktuatoren mit unterschiedlichen Zeithorizonten für KI-basierte Satelliten-Lageregelung mittels Subnetz-Politik
Authors:
K. Djebko, T. Baumann, E. Dilger, F. Puppe, S. Montenegro
Abstract:
Reliable attitude control is essential for satellite missions. However, the development of conventional controllers is time-consuming and prone to discrepancies between expectations and reality. Deep reinforcement learning (DRL) offers a promising alternative, allowing a controller to learn an adapted control strategy through autonomous interaction with an environment by maximizing a reward function. In practice, reaction wheels and magnetic torquers are often combined, with reaction wheels used for rapid attitude changes and magnetic torquers for slow momentum management. Since both actuators influence each other, a unified AI controller tuned to both actuators is desirable. Training such a unified controller is challenging, however, because the actuators have conflicting requirements for the reward function and the training hyperparameters. Reaction wheels enable rapid attitude changes and require a primary focus on immediate actions during training, whereas magnetic torquers cause slow changes and require long time horizons to be considered. What makes matters worse is that although both actuator types have very different behavioral characteristics, they influence each other. In our work, we combine the control of both actuators by using subnetworks that can be activated, frozen, or deactivated during training. This resolves the training conflict and simplifies the underlying optimization problem. We applied our method to train an AI controller for the InnoCube satellite using a realistic simulator. InnoCube is a 3U CubeSat developed by the Julius-Maximilians-University of Würzburg in cooperation with the Technical University of Berlin and successfully launched in January 2025. The goal is to test the controller directly on board the satellite in space and to iteratively improve it.
Event:
Deutscher Luft- und Raumfahrtkongress 2025, Augsburg
Location, Year:
Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V., Bonn, 2025
Format:
21,0 x 29,7 cm, 10 Pages
Descriptors:
EN: Artificial intelligence, deep reinforcement learning, nanosatellites, CubeSats, attitude control
Citation:
Djebko, K.; Baumann, T.; et al. (2025): Vereinigung der Steuerung von Aktuatoren mit unterschiedlichen Zeithorizonten für KI-basierte Satelliten-Lageregelung mittels Subnetz-Politik. DGLR Published Conference Papers. Volume 4. (2026). ISSN: 2941-3451. Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V.. (Text). https://doi.org/10.25967/650051. urn:nbn:de:101:1-2510241121206.896055490146.
Publishing Date:
24.10.2025