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

AI summary

Overview

Research area: Spacecraft guidance, navigation and control (specifically satellite attitude control), deep reinforcement learning (DRL), and simulation-to-reality (Sim2Real) transfer for embedded systems.

Technical level: Advanced. The paper assumes familiarity with quaternion attitude representations, proportional-derivative control, actor-critic reinforcement learning, and CubeSat subsystem architecture.

Scope in one sentence: The paper reports the design, training, and first in-orbit flight test of a deep-reinforcement-learning attitude controller on the InnoCube 3U nanosatellite, benchmarked against the satellite's own proportional-derivative controller.

What This Paper Is About

Classical satellite attitude controllers such as proportional-derivative (PD) controllers take a long time to design and are sensitive to model uncertainty and changing operating conditions. Deep reinforcement learning is a promising alternative, but until now AI-based attitude controllers have only ever been demonstrated in simulation or on ground testbeds. This paper claims the first successful in-orbit demonstration of an AI-based attitude controller for inertial pointing maneuvers, trained entirely in simulation and deployed to a real satellite, thereby overcoming the simulation-to-reality gap.

Key Contributions

1. First claimed in-orbit demonstration of an AI-based attitude controller. The authors state that to date AI-based attitude control has been restricted to simulation environments or laboratory testbeds, with no successful in-orbit demonstrations, and that this work represents the first successful test in orbit. The paper also notes that the earlier HOPAS project on ESA's OPS-SAT CubeSat attempted an early in-orbit demonstration using an AI agent as a residual on top of a PID controller, but its full AI closed-loop control was ultimately restricted to simulation and ground testing.

2. A complete training methodology combining a base-agent and post-trained flight-agent. The controller was trained only on the ground, uploaded to InnoCube, and tested in orbit. Training used a base-agent with a split policy network (a reaction-wheel subnetwork and a magnetorquer subnetwork) trained with a modified PPO variant called SkipPPO, then post-trained into the LeLaR flight-agent.

3. A quantified characterization of simulation-to-reality discrepancies. The paper reports a sensor noise model derived from Thermal Vacuum Chamber (TVAC) testing, a refined inertia tensor, and a machine-learning-based calibration of residual magnetic dipole moments and gyroscope bias from in-orbit telemetry.

4. A safety architecture for AI-controlled flight maneuvers. The "Safety Cage" monitors telemetry against predefined safety limits and triggers a fallback to a safe idle state, immediately turning off the AI-based controller if anomalous behavior occurs during AI-controlled maneuvers.

Main Findings

• Reaction-wheel subnetwork alone was sufficient. While the controller was originally designed with both a reaction-wheel (RW) and magnetorquer (MT) subnetwork, initial in-orbit tests showed that the RW subnetwork's performance on its own was sufficient, and it also fit the later safety and power constraint of a maximum experiment duration of 15 minutes. This was the reason the flight-agent was derived by fine-tuning the RW subnetwork only.

• Momentum management was not viable in the early experiment phase. Command list execution was not available early on, limiting tests to live teleoperation during passes, with a maximum pass duration of approximately 12 minutes. Because the maneuver length is significantly longer than the pass duration (reported as greater than 2000 s), the momentum-management logic was not viable during these tests.

• Residual magnetic dipole moments were unexpectedly high. InnoCube exhibited unexpectedly high residual magnetic dipole moments during early in-orbit operations. The authors' automatic-differentiation-based calibration yielded a residual dipole of μ = [-0.459, -0.024, 0.069] A·m² and a gyroscope bias of b = [-0.028, 0.761, -0.032] °/s.

• Magnetorquer validation matched theory. Activating the magnetorquers sequentially and re-running the PyTorch calibration script produced magnetic-field changes corresponding to dipole moments of 0.31–0.39 A·m², in accordance with expected theoretical values, confirming both the MT characteristics and the learned residual dipole.

• Partial residual dipole compensation reduced momentum buildup. Following application of the dipole moment compensation from one set of magnetorquers, in-orbit checks showed a significant reduction of the residual dipole moment, reducing the buildup of angular momentum during maneuvers. The authors note that fully eliminating the residual on the X-axis would require using the second set of magnetorquers, which is currently not implemented for safety reasons because it would require running both ADCS nodes simultaneously.

• Steady-state metrics confirm robust performance. The abstract states that steady-state metrics confirm the robust performance of the AI-based controller during repeated in-orbit maneuvers, and that the in-orbit performance was evaluated against both simulation results and the in-orbit performance of InnoCube's default PD controller under identical mission conditions. The specific numerical in-orbit performance values and comparison tables are not included in the truncated content provided here.

• Inertia model was accurate but still randomized for training. The final inertia tensor I = diag(0.0428, 0.0422, 0.00985) kg·m² deviated by at most 2.5% from initial CAD estimates, yet inertia was still varied by ±15% per axis during domain randomization for all agents.

• Safety limits were tightened after deployment. Shortly after deployment and prior to the first test, the agent's maximum rotation rates were limited to 20 °/s per axis, and the effective applied torque was constrained to 50 rpm/s per axis initially and shortly after relaxed to 100 rpm/s per axis (the text is truncated at this point).

Methodology in Plain English

The team built a simulation model of the InnoCube satellite from component specifications, CAD-based inertia estimates, and experimental subsystem data, then trained a neural-network controller inside that simulation using deep reinforcement learning. The controller is given a 39-dimensional observation vector, normalized to [-1, 1], that includes the attitude error quaternion, the rate error, the last executed reaction-wheel action, reaction-wheel speeds, the magnetic field vector, the reaction-wheel speed error, a scaled cross product of wheel speed and magnetic field, and the magnetic field norm. Because several of these are supplied at two consecutive time steps, the agent can infer the system dynamics from the state deltas.

The base-agent used a split policy network: one subnetwork for the reaction wheels and one for the magnetorquers, each of which could be DISABLED, FROZEN, or ACTIVE. Training used SkipPPO, a modification of Proximal Policy Optimization that allows action freezing and time-step skipping, which mitigates the credit-assignment problem caused by the fact that reaction wheels and magnetorquers operate on time horizons two orders of magnitude apart. The reaction-wheel network was trained first with its own reward function, the magnetorquer network was then trained with the wheel network frozen to bring wheel speeds to ±500 rpm, and finally both subnetworks were post-trained together using a combined reward.

To improve simulation fidelity, the authors recorded gyroscope and magnetometer data from the InnoCube engineering qualification model in a Thermal Vacuum Chamber evacuated to 10⁻⁴ mBar, cropped the data to the in-orbit temperature range of -3.55 °C to 15.6 °C, split it into 60-second segments, and built a bias-plus-white-noise model whose standard deviations were scaled by a factor of 1.2 for domain randomization. They also refined the inertia tensor from telemetry and used a differentiable PyTorch-based optimization (AdamW with automatic differentiation) to estimate residual magnetic dipole moments and gyro bias by minimizing the residual of Euler's rotational dynamics.

The resulting flight-agent was the reaction-wheel subnetwork of the base-agent, fine-tuned through post-training for the tightened post-launch safety constraints. It was uploaded to the satellite and tested in orbit, with the Safety Cage running in parallel to turn the AI controller off and fall back to a safe idle state if telemetry exceeded predefined limits.

Why This Matters

Impact on research: AI-based control for spacecraft has largely been a simulation-only pursuit. This paper provides the first reported flight evidence that a DRL agent trained entirely in simulation can control a real satellite through inertial pointing maneuvers, which directly addresses the field's central open problem: the simulation-to-reality gap. It also contributes reusable artifacts — a sensor noise model from TVAC testing, a machine-learning calibration method for residual dipole and gyro bias, and a public telemetry dataset (available at https://github.com/kdjebko/lelar-in-orbit-data).

Real-world applications:

• Earth-observation and remote-sensing satellites, which need precise inertial pointing and repeated slew maneuvers, and can benefit from controllers that generalize across uncertain inertia and disturbance conditions.
• Small-satellite and CubeSat constellations, where the cost of manually designing and tuning a controller for each mission is high, and where an autonomously learned controller could reduce development effort.
• GNSS and radio-occultation payloads, since InnoCube's PD controller was originally designed for GNSS pointing with a relaxed requirement of around ±90° anti-nadir pointing accuracy.
• Laser-ranging tracking, where InnoCube's design requires pointing the retroreflector toward a ground station with ±17° accuracy during a pass.

Industry relevance: The paper matters for satellite manufacturers and operators because it frames AI control as a way to cut design time and to handle model uncertainty without maintaining high-fidelity models for every environmental contingency. The Safety Cage architecture and the documented safety-limit tightening (20 °/s maximum rotation rate per axis; torque rate constraint initially 50 rpm/s per axis, relaxed to 100 rpm/s per axis) show how the authors bridged the gap between laboratory AI and flight-safety certification. The claim of zero-shot transfer — the agent operating successfully upon deployment without further training — is the property most relevant to commercial adoption, since iterative post-launch fine-tuning is often infeasible.

Future Directions

1. Full momentum management in orbit. The magnetorquer subnetwork was never exercised in a closed-loop flight test because command list execution was unavailable and pass durations (approximately 12 minutes) were shorter than the maneuvers (greater than 2000 s). Once unsupervised command execution is available within the 15-minute experiment limit, the combined RW-and-MT agent could be tested.

2. Complete residual dipole elimination. Fully eliminating the residual dipole on the X-axis requires using the second set of magnetorquers, which is not currently implemented because it would require running both ADCS nodes simultaneously. Resolving this would reduce angular momentum buildup during maneuvers.

3. Absolute attitude determination. The evaluation relied on gyroscope integration only, because InnoCube lacks highly accurate absolute attitude sensors such as star trackers and the current sensor-fusion accuracy from magnetometer and sun-sensor measurements does not produce accurate continuous determination solutions. Improving on-board determination would allow the controller to be evaluated against absolute rather than purely relative pointing accuracy.

4. Better extrapolation guarantees for Sim2Real transfer. The paper notes that neural networks tend to perform well on interpolation but fail on extrapolation, and that domain randomization is used to avoid out-of-distribution errors. How far the trained policy generalizes beyond the randomized training distribution remains an open question.

Target Audience

This paper is most valuable to spacecraft guidance, navigation and control engineers; researchers working on reinforcement learning for control of physical systems; CubeSat and nanosatellite developers; and teams interested in flight qualification and safety certification of AI-based controllers. Readers focused on the numerical benchmark comparisons should note that the detailed in-orbit performance tables referenced in the abstract and Section III are not contained in the truncated content provided here, so specific in-orbit steady-state figures, settling times, and the direct numerical comparison against the default PD controller under identical mission conditions are not reported in this summary.

Authors’ abstract

Attitude control is essential for many satellite missions. Classical controllers, however, are time-consuming to design and sensitive to model uncertainties and variations in operational boundary conditions. Deep Reinforcement Learning (DRL) offers a promising alternative by learning adaptive control strategies through autonomous interaction with a simulation environment. Overcoming the Sim2Real gap, which involves deploying an agent trained in simulation onto the real physical satellite, remains a significant challenge. In this work, we present the first successful in-orbit demonstration of an AI-based attitude controller for inertial pointing maneuvers. The controller was trained entirely in simulation and deployed to the InnoCube 3U nanosatellite, which was developed by the Julius-Maximilians-Universität Würzburg in cooperation with the Technische Universität Berlin, and launched in January 2025. We present the AI agent design, the methodology of the training procedure, the discrepancies between the simulation and the observed behavior of the real satellite, and a comparison of the AI-based attitude controller with the classical PD controller of InnoCube. Steady-state metrics confirm the robust performance of the AI-based controller during repeated in-orbit maneuvers.