AI-developed controller directs satellite in pioneering in-orbit maneuver

AI-developed controller directs satellite in pioneering in-orbit maneuver
by Robert Schreiber
Berlin, Germany (SPX) Nov 08, 2025

A team at Julius-Maximilians-Universitat Wurzburg (JMU) has achieved a significant milestone on the road to space autonomy by successfully testing an artificial intelligence-based attitude controller in orbit for the first time. This world-first achievement was carried out using the 3U nanosatellite InnoCube, with direct in-orbit control administered by a system developed at JMU. The test occurred during a window between 11:40 and 11:49 a.m. CET on October 30, 2025, when the AI agent autonomously guided the satellite from its initial orientation to a preselected target configuration using reaction wheels. In subsequent real-world tests, the AI consistently accomplished further maneuvers - safely and precisely aligning the satellite as instructed, repeatedly demonstrating its operational reliability.

The team behind the experiment includes Dr. Kirill Djebko, Tom Baumann, Erik Dilger, Professor Frank Puppe, and Professor Sergio Montenegro. Their initiative, known as LeLaR (In-Orbit Demonstrator for Learning Attitude Control), focuses on developing and testing next-generation autonomous attitude control systems for future space missions. The LeLaR project's core achievement is the design and successful on-orbit deployment of an AI controller on InnoCube. These controllers play a critical role in orienting satellites for their tasks - helping to point cameras, sensors, or antennas, and to stabilize the craft, thereby ensuring mission success and communication reliability.

Unlike conventional controllers that require time-intensive manual adjustment of control parameters and introduce latency in adapting to new mission conditions, the Wurzburg team used a Deep Reinforcement Learning (DRL) approach. Their system, a neural network, was trained terrestrially in a high-fidelity simulator to develop responsive, optimal control strategies, and then uploaded to the satellite for true in-orbit testing.

"This research is a decisive success," emphasized Kirill Djebko. "We have achieved the world's first practical proof that a satellite attitude controller trained using Deep Reinforcement Learning can operate successfully in orbit," he said. Tom Baumann added, "This successful test marks a major step forward in the development of future satellite control systems. It shows that AI can not only perform in simulation but also execute precise, autonomous maneuvers under real conditions." Professor Frank Puppe highlighted that robust simulation models were key to bridging the so-called 'Sim2Real gap' - the challenge of ensuring controllers trained in digital environments remain effective and reliable amidst real, unpredictable conditions of space.

DRL-backed controllers offer a critical advantage in speed and adaptability. Traditional systems may take engineers months or years to recalibrate, but DRL controllers can master their environment far faster, and even adapt dynamically to unexpected differences between theoretical and actual operational scenarios with minimal intervention. This flexibility opens the path for spacecraft to undertake more complex and autonomous missions, critical for deep-space exploration and missions where human response will be limited by extreme distances or signal delay.

Having demonstrated that an AI-based controller can function precisely and safely during mission-critical satellite operations, the Wurzburg team has fortified the reliability and acceptance of artificial intelligence for high-stakes aerospace tasks. Professor Sergio Montenegro notes, "We are at the beginning of a new class of satellite control systems: intelligent, adaptive and self-learning. It's a major step towards full autonomy in space."

Another innovation featured in this mission is InnoCube's SKITH (Skip The Harness) wireless satellite bus. This technology replaces traditional wiring with wireless data transmission, saving both mass and reducing risk of system failures linked to cabling faults. InnoCube itself, built in cooperation with Technische Universitat Berlin, provides a versatile low-Earth-orbit testbed for innovations such as these.

Before launch, the AI controller underwent qualification procedures in specialized test chambers, where it was subjected to simulated space conditions for calibration and verification. The LeLaR project - funded by the German Space Agency at DLR since July 2024, with approximately euro 430,000 provided by the Federal Ministry for Economic Affairs and Energy - represents a growing European commitment to adaptive, self-guided spacecraft and associated technologies. The funding underlines strategic prioritization of technologies that can expedite, adapt, and automate satellite management in future missions.

ADCS box (Attitude Determination and Control System) being installed in the qualification model of the InnoCube satellite.

The LeLaR project's successful completion motivates the team to extend these techniques to further mission scenarios and platforms, with the aim to support more complex satellite and deep-space projects. These results demonstrate the technical feasibility and value of neural network-driven controllers for a broad spectrum of satellite platforms, pointing the way to rapid, cost-effective, and continually improving systems.

In summarizing their achievement, the Wurzburg team emphasizes that these advances represent not just an incremental improvement, but a foundational shift enabling new classes of mission architectures previously impossible with conventional control methods. As simulation and training models mature further, and as confidence in DRL systems grows, JMU's pioneering work sets new benchmarks for reliability and autonomy in the ever-evolving field of space engineering.