Colloquium: Lars Dijkstra (C&O)

Active Debris Removal missions are critical for mitigating the risks posed by space debris. Conceptual missions or design studies often consider a chaser with a space manipulator to capture targets and remove them from their orbits. However, in the post-capture phase of such missions, attitude control is not straightforward because many uncertainties exist in the mass and inertia of the combined spacecraft. In this research, model-free reinforcement learning algorithms were applied to directly deal with these uncertainties in a spacecraft attitude control setting. The trained agents displayed robust and highly performant results against several types of perturbations. Furthermore, domain randomization techniques were employed to enhance the robustness of the agents. The agents also showed potential to perform well when flexibilities in the combined spacecraft were introduced, underwriting their potential to learn highly performant and robust policies for control.

Supervisor: Jian Guo