On November 11th, the University of W ü rzburg (JMU) in Germany announced on November 7th that its research team has completed the world's first AI controlled satellite attitude adjustment experiment in orbit, marking an important milestone in the autonomy of space systems.
This test was conducted by the LeLaR (Learning Attitude Control in Orbit Verification Project) team at the school, using a 3U level nanosatellite called InnoCube as the experimental platform.
On October 30, 2025, between 11:40 am and 11:49 am Central European Time, an AI controller successfully used a reaction wheel to adjust the satellite from its initial attitude to the preset target attitude, and the entire process was completely completed autonomously by artificial intelligence in orbit. Subsequently, in multiple tests, the AI was able to stably control the satellite to point towards the target direction.
Project and Technical Background
The LeLaR project, also known as "In Orbit Demonstrator for Learning Attitude Control," aims to develop a new generation of autonomous satellite attitude control systems. The attitude controller is used to stabilize the satellite attitude, prevent it from rolling in orbit, and ensure that its camera, sensor, or antenna is aligned with the designated target.
Unlike traditional control methods that rely on fixed algorithms, the JMU team has adopted Deep Reinforcement Learning (DRL), an AI technology that uses neural networks to autonomously learn optimal control strategies in simulated environments.
Compared to traditional methods, DRL has significant advantages in development efficiency and adaptability, which can significantly reduce AI debugging time and automatically adjust control strategies based on changes in the actual environment, thereby avoiding lengthy recalibration processes.
Experimental breakthroughs and significance
The researchers first trained the AI controller in a high fidelity simulation environment on the ground, and then uploaded it to a satellite for in orbit verification. One major technical challenge in the project is to address the "simulation to reality gap" - ensuring that the control algorithms trained in the simulation can operate effectively in real space environments.
Dr. Kirill Djebko, the project leader, stated, "We have achieved the world's first empirical successful operation of a satellite attitude controller based on deep reinforcement learning in orbit
Team member Tom Baumann added, "This success marks a crucial step forward for future satellite control systems, demonstrating that artificial intelligence can not only perform well in simulations, but also safely execute autonomous maneuvers in actual space environments
Promote the autonomy of space missions
The success of this project demonstrates the reliability of artificial intelligence in aerospace safety critical tasks. The research team believes that this will help increase the acceptance of AI methods in the aerospace industry and lay a foundation of trust for future autonomous space missions.
Team member Professor Frank Puppe pointed out that this achievement will significantly increase the recognition of AI technology in aerospace research. ”
AI control methods have potential application value in deep space exploration missions, especially in environments with communication delays or where human intervention is not possible. Autonomous learning control systems may become key to spacecraft survival and mission execution.
Future prospects
LeLaR project researcher Erik Dilger stated that the team plans to expand the technology to more in orbit scenarios. The InnoCube satellite used in this experiment was jointly developed by the University of W ü rzburg and the Technical University of Berlin (TU Berlin) to provide an in orbit testing platform for innovative space technologies. The satellite is also equipped with a wireless satellite bus system called SKITH (Skip The Harness), which replaces traditional wiring with wireless data transmission to reduce quality and lower potential failure risks.
The research team believes that this achievement lays the foundation for future intelligent, adaptive, and self-learning satellite control systems. Professor Sergio Montenegro, the project leader, summarized, "This is a big step and marks the beginning of a new phase in satellite control systems - intelligent, flexible, and capable of self-learning
