AI controls satellite attitude in orbit for first time

As a real step on the path to autonomous space systems, a research team from the Julius-Maximilians-University of Würzburg (JMU) has successfully tested an AI-based attitude controller for satellites directly in orbit – a world first. The test was carried out on board the 3U InnoCube nanosatellite.

During the satellite’s passage between 11:40 and 11:49 CET on October 30, 2025, the AI ​​agent developed at JMU performed a complete attitude maneuver in orbit, entirely controlled by artificial intelligence. Using reaction wheels, the AI ​​brought the satellite from its current initial attitude to a specified target attitude. The AI ​​then had several more opportunities to prove its capabilities: In subsequent tests, it also managed to safely control the satellite in the desired attitude.

The LeLaR research team—Dr. Kirill Djebko, Tom Baumann, Erik Dilger, Professor Frank Puppe and Professor Sergio Montenegro have thus taken a decisive step towards spatial autonomy.

The LeLaR project

The In-Orbit Demonstrator for Learning Attitude Control (LeLaR; German: In-Orbit Demonstrator Lernende Lageregelung) project aims to develop the next generation of autonomous attitude control systems. Its central focus was the design, training and on-orbit testing of an AI-based attitude controller aboard the InnoCube nanosatellite.

Attitude controllers stabilize satellites in orbit and prevent them from tipping. They are also used to steer the spacecraft in a desired direction. For example, to align cameras, sensors or antennas to a specific target.

What makes this work special is that the Würzburg controller was not built using traditional fixed algorithms. Instead, the researchers applied a deep reinforcement learning (DRL) approach, a branch of machine learning in which a neural network autonomously learns the optimal control strategy in a simulated environment.

Fast and adaptive

The main advantage of the DRL approach is its speed and flexibility compared to traditional control development. Traditional attitude controllers often require lengthy manual parameter adjustments by engineers, sometimes taking months or even years.

The DRL method automates this process. Additionally, it provides the ability to create controllers that automatically adapt to differences between expected and actual conditions, eliminating the need for time-consuming manual recalibration.

Overcoming the Sim2Real Gap

Before deployment, the AI ​​controller was trained on Earth in a high-fidelity simulation and then uploaded to the orbiting satellite’s flight model. One of the biggest challenges was bridging the Sim2Real gap, namely ensuring that a simulation-trained controller was also operational on the real satellite in space.

“A truly decisive success,” emphasizes JMU’s Djebko. “We have obtained the world’s first practical proof that a satellite attitude controller trained using deep reinforcement learning can operate successfully in orbit,” he adds.

Baumann explains: “This successful test marks a major step forward in the development of future satellite control systems. It shows that AI can not only operate in simulation, but also perform precise and autonomous maneuvers in real-world conditions. »

Acceptance and trust in AI for space applications

By successfully demonstrating an AI-based controller in orbit, the Würzburg team showed that artificial intelligence can be reliably applied in safety-critical space missions.

Puppe is convinced that “this will significantly increase the acceptance of AI methods in aeronautical and space research,” emphasizing the important role of the simulation model.

Growing confidence in such technology is a crucial step toward future autonomous missions. For example, interplanetary or deep space missions, where human intervention is impossible due to great distances or communication delays. The AI-based approach could therefore become vital for the survival of spacecraft.

A significant contribution to spatial autonomy

With this experience, the Würzburg team achieved a major goal of the LeLaR project.

“This success motivates us to extend the technology to new scenarios,” says Erik Dilger. The test was carried out on board InnoCube, a satellite developed in cooperation with Technische Universität Berlin (TU Berlin). InnoCube serves as a platform for innovative space technologies, giving researchers the opportunity to test new concepts directly in orbit.

One of these innovations is the SKITH (Skip The Harness) wireless satellite bus, which replaces conventional wiring with wireless data transmission. This not only saves mass, but also reduces potential sources of failure.

Perspectives: the next stage of space autonomy

This successful in-orbit test makes the University of Würzburg a pioneer in AI-driven space systems. The demonstrated AI-based controller represents an important building block for future deep space exploration. The results of the LeLaR project could enable the faster and more cost-effective development of new complex AI-based controllers for a wide range of satellite platforms.

“The next goal is to capitalize on this head start,” says Djebko.

“This is a major step towards full autonomy in space,” adds Montenegro. “We are at the beginning of a new class of satellite control systems: intelligent, adaptive and self-learning.”