⚡ electrical engineering

LeLaR: The First In-Orbit Demonstration of an AI-Based Satellite Attitude Controller

This paper reports the first successful in-orbit demonstration of an AI-based satellite attitude controller, where a Deep Reinforcement Learning agent trained entirely in simulation was deployed on the InnoCube nanosatellite to achieve robust inertial pointing maneuvers that outperform classical PD controllers.

Original authors: Kirill Djebko, Tom Baumann, Erik Dilger, Frank Puppe, Sergio Montenegro

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kirill Djebko, Tom Baumann, Erik Dilger, Frank Puppe, Sergio Montenegro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very delicate, expensive toy spaceship floating in the dark, silent void of space. Your job is to spin it around and point it in a specific direction so its solar panels catch the sun or its camera can take a picture.

Traditionally, engineers build a "pilot" for this ship using strict math rules (like a recipe). If the wind blows, or the ship gets heavier, or a wheel gets sticky, the recipe might fail because it doesn't know how to improvise.

This paper is about LeLaR, a new kind of pilot. Instead of following a recipe, LeLaR is an AI that learned how to fly by playing a video game. And the best part? It didn't just play the game; it took the skills it learned in the game and successfully flew a real spaceship in orbit, without ever needing a human to teach it the real-world rules.

Here is the story of how they did it, explained simply:

1. The Problem: The "Video Game vs. Real Life" Gap

Think of training an AI like teaching a dog to fetch. You can teach the dog in your living room (the simulation), but when you take it to the park (the real world), the wind is stronger, the ground is muddy, and the ball bounces differently. Usually, the dog gets confused and fails.

In space, this is called the Sim2Real Gap.

The Simulation: A perfect, computer-generated world where physics works exactly as the engineers expect.
The Real World: A messy place where reaction wheels (the spinning wheels that turn the satellite) sometimes get stuck, sensors glitch, and magnetic fields behave strangely.

Most AI pilots fail when they leave the "living room" and enter the "park." They are too rigid.

2. The Solution: The "Gymnast" Approach

The team at the University of Würzburg didn't just train their AI on one perfect version of the game. They made the training environment chaotic on purpose.

Imagine training a gymnast not just on a perfect floor, but on a floor that sometimes tilts, sometimes has slippery spots, and sometimes has random gusts of wind.

They taught the AI to handle random variations in the satellite's weight.
They taught it to deal with noisy sensors (like trying to balance while wearing foggy glasses).
They taught it to cope with magnetic interference (like trying to walk a tightrope while holding a giant magnet).

By training in this "chaotic gym," the AI learned to be adaptable. It didn't memorize a single solution; it learned a strategy to handle anything that might go wrong.

3. The Mission: The InnoCube Satellite

The "spaceship" in this story is InnoCube, a tiny 3-unit CubeSat (about the size of a loaf of bread) launched in January 2025.

The Challenge: The satellite's reaction wheels had a weird quirk. If they spun too slowly, they would get "stuck" or give false readings. It was like a car engine that sputters if you drive too slowly.
The Fix: The AI was designed to avoid these "danger zones" automatically, learning to keep the wheels spinning at the right speed to stay in control.

4. The Safety Net: The "Bungee Cord"

Since this was a real satellite, the engineers couldn't let the AI go crazy. If the AI made a mistake and started spinning the satellite too fast, it could break the antenna or drain the battery.

They installed a Safety Cage. Think of this as a bungee cord attached to the AI pilot.

The AI is free to fly and make decisions.
But if it tries to spin the ship faster than a safe speed (20 degrees per second), the Safety Cage instantly cuts the power and takes over, putting the ship into a safe "idle" mode.
The Result: The AI flew the ship perfectly, and the Safety Cage never had to intervene. It was like a student driver who drove so well the instructor never had to touch the brake pedal.

5. The Big Test: The "Zero-Shot" Debut

This is the most impressive part. Usually, when you send a robot to a new planet, you have to spend months tweaking its software once it arrives.

LeLaR did something called "Zero-Shot Transfer."

It was trained entirely on Earth.
It was uploaded to the satellite.
It was turned on in space.
It worked immediately.

It didn't need to "re-learn" how to fly in space. It looked at the real, messy data from the satellite, recognized the patterns it had seen in its chaotic training, and said, "I've got this."

6. The Results: AI vs. The Old Way

To prove it worked, they compared the AI pilot against the satellite's standard "human-written" pilot (a classic PD controller).

The Standard Pilot: It was smooth but a bit clumsy. It often missed the target mark by a wide margin (up to 2 degrees off).
The AI Pilot (LeLaR): It was precise. It consistently stopped the satellite within 1 degree of the target, even when the reaction wheels were glitching or the satellite was spinning unexpectedly.

Why This Matters

This paper is a milestone because it's the first time an AI has successfully flown a real satellite in orbit without human help.

Think of it like this: For decades, we've been teaching computers to play chess. Now, for the first time, we've taught a computer to drive a real car in a real storm, and it didn't crash.

This opens the door for future space missions where satellites can fix themselves, adapt to broken parts, and make decisions on their own, far away from Earth where humans can't reach them in time. The "living room" training worked, and the "park" is ready for the AI.

1. Problem Statement

Attitude control is critical for satellite missions, yet classical controllers (e.g., Proportional-Derivative or PID) are often time-consuming to design, sensitive to model uncertainties, and struggle with nonlinear dynamics and varying operational boundary conditions. While Deep Reinforcement Learning (DRL) offers a promising alternative for learning adaptive control strategies, a significant barrier remains: the Sim2Real gap. Transferring an agent trained in a high-fidelity simulation to a real physical satellite without further in-flight training (zero-shot transfer) has historically been unachieved in orbit. Previous attempts, such as the HOPAS project on OPS-SAT, were restricted to simulation or ground testing.

2. Methodology

The authors developed LeLaR (In-Orbit-Demonstrator Lernende Lageregelung), an AI-based attitude controller for the InnoCube 3U nanosatellite. The methodology involved a multi-stage training and deployment process:

Satellite Platform: InnoCube is a 3U CubeSat launched in January 2025 into a 520 km sun-synchronous orbit. It utilizes three Reaction Wheels (RWs) and three Magnetorquers (MTs) for attitude control.
Simulation & Domain Randomization:
- A simulation environment was built using the Basilisk Astrodynamics Framework.
- Domain Randomization was applied during training to ensure robustness. Parameters varied included satellite moments of inertia (±15%), orbital elements, sensor noise (gyro bias and white noise derived from Thermal Vacuum Chamber testing), and residual magnetic dipole moments.
- Sensor Calibration: The team used automatic differentiation (PyTorch) to estimate and compensate for residual magnetic dipole moments and gyro biases using in-orbit telemetry, refining the simulation model.
Agent Architecture:
- Base-Agent: Initially designed with a split policy network (RW subnetwork for fast attitude control, MT subnetwork for slow momentum management) using the SkipPPO algorithm (a variant of Proximal Policy Optimization). This addressed the credit assignment problem caused by the vastly different time scales of RWs and MTs.
- Flight-Agent (LeLaR): Due to operational constraints (short ground passes preventing full momentum management cycles) and safety limits, the final flight agent utilized only the RW subnetwork of the base-agent. This subnetwork was post-trained (fine-tuned) to operate independently, adhering to new safety constraints (max body rate 20°/s, max torque 100 rpm/s) and handling active residual dipole compensation.
Safety Cage: A critical safety module was integrated into the onboard software. It monitors telemetry against hard limits (e.g., RW speeds, body rates). If limits are exceeded, it immediately disables the AI controller and enters a safe idle state, ensuring mission safety.
Observation Space: The agent receives a 39-dimensional vector including attitude error quaternions, rate errors, last actions, RW speeds, and magnetic field data. The input is normalized to $[-1, 1]$ .

3. Key Contributions

First In-Orbit Demonstration: This work presents the first successful in-orbit demonstration of an AI-based attitude controller for a spacecraft, achieving zero-shot transfer from simulation to flight hardware.
Overcoming Sim2Real Gap: The controller was trained entirely on the ground and deployed without in-flight retraining, successfully adapting to unmodeled dynamics such as:
- Random "dead times" (up to 20s) in RW activation.
- Sudden spikes in RW speed measurements (artifacts).
- Unmodeled residual magnetic dipole moments.
Hybrid Training Strategy: The use of a split policy network (RW/MT) followed by selective post-training of the RW subnetwork allowed the system to adapt to specific mission constraints while leveraging pre-learned dynamics.
Safety-First Integration: The implementation of a rule-based Safety Cage alongside the AI controller provides a robust framework for certifying AI in safety-critical aerospace applications.

4. Results

The LeLaR flight-agent was tested on InnoCube starting in October 2025. Key performance metrics include:

Zero-Shot Success: The agent successfully performed inertial pointing maneuvers (targeting a specific quaternion) in orbit without prior exposure to the real flight environment.
Steady-State Accuracy:
- The AI controller consistently achieved steady-state pointing accuracy within ±1° across all axes (Yaw, Pitch, Roll), meeting the training objective.
- In repeated maneuvers, the AI controller maintained a mean error close to 0° with low standard deviation (e.g., ~0.05°–0.15°).
Comparison with Classical PD Controller:
- The AI controller outperformed the satellite's default PD controller in precision. The PD controller often failed to reach the strict 1° threshold, requiring a relaxed threshold of 1.8°–2.0° to achieve steady-state in all maneuvers.
- While the PD controller showed smoother transient trajectories, the AI controller demonstrated superior robustness against unmodeled disturbances and sensor anomalies.
Transient Dynamics: Although the transient response (settling time) differed between simulation and reality due to unmodeled RW dead times and delays, the agent successfully converged to the target attitude in all tested scenarios.
Safety: No Safety Cage interventions were triggered during the experiments, confirming the agent operated within safe bounds.

5. Significance

This paper marks a pivotal milestone in aerospace engineering and AI:

Validation of DRL in Space: It proves that Deep Reinforcement Learning can be successfully deployed on flight hardware for critical control tasks, moving beyond theoretical simulations and ground testbeds.
Autonomy for Deep Space: The ability to achieve zero-shot transfer is crucial for future deep-space missions where high-fidelity modeling is impossible and iterative in-flight tuning is infeasible due to communication delays.
Scalability: The methodology is not limited to CubeSats; the approach can be adapted to larger satellites provided the observation space can be covered by existing sensors.
Future Outlook: The authors plan to extend the experiments to include full momentum management using Magnetorquers and to apply the methodology to other satellite platforms to further validate generalization.

In conclusion, LeLaR demonstrates that AI-based controllers can not only match but exceed the performance of classical control methods in real-world space environments, offering a viable path toward more autonomous and adaptive spacecraft.