Robuste Lageregelung unter Einsatz von Reinforcement Learning für einen Raumfahrt-Landedemonstrator

Kurzfassung

The applicability of so-called green propellants and propulsion systems for spacecraft landers is to be evaluated through the development of a lander demonstrator at the German Aerospace Center in the context of the project “The LÄNDer”. Propellant technology researched and developed onsite is tested on a modular testing platform, enabling testing of varying degrees of freedom with the intention of safely studying components and control systems, gradually stepping towards the final milestone of the project, i.e., the conduction of a free-flight demonstration. This work contributes to the development of an attitude control system for the lander demonstrator. A conventional control approach is elaborated as a baseline, and a reinforcement learning-based control system is built as a representative of promising data-driven methods. A consumer-grade inertial measurement unit is used to measure the kinematics of the system. For estimation of states, a complementary filter is implemented as a state-observer. With tremendous requirements for the application of any system in the space industry, safety is of major concern. Therefore, the robustness of the attitude control system is assessed extensively and the effect of deviations from the model are studied by means of typical control engineering performance metrics. For this purpose, the most influential properties on control performance are identified through a sensitivity analysis through evaluation of the control performance deterioration in the scope of Monte Carlo experiments. This work focuses on studying the closed-loop performance of the control system in the context of state estimation uncertainties as well as stochastic and deterministic environmental disturbances. Development and evaluation of controllers is conducted within a simulation environment in a first step. The controller trained through interaction with the simulation is then deployed on the hardware without further tuning. To achieve this, the trained neural network is extracted and run on an embedded real-time system. Tests with the real system are conducted using a dedicated test bench built in the context of the project. In conclusion, insights gained from the sensitivity analysis and observations of the performance on the real system are used to improve the controller’s robustness, followed by a final evaluation of the tuned controller.