Rocket Recycling via Reinforcement Learning

Rocket recycling offers significant potential to reduce launch costs, increase launch frequency, and reduce space debris, as evidenced by SpaceXs success with rocket relandings. Traditional model-based control methods face challenges in handling the complex, non-linear dynamics inherent in the rocket recycling process. Reinforcement learning (RL), by contrast, provides a flexible, data-driven approach which is better suited to managing these complicate dynamic conditions. In this paper, we investigate two research problems in RL-based rocket recycling control: 1) how the granularity of control variables (e.g. thrust force levels) impacts the control performance; 2) how the environmental interference during training affects the robustness of RL controller. Our evaluation results demonstrate that increasing the granularity of control variables significantly slows the convergence of RL-based controller to a stable state. Moreover, introducing environmental interference during trainingsuch as wind disturbancesimproves both the efficiency and robustness of the RL-based controller. Our findings provide insights into how the RL-based controller can be more effectively used for rocket recycling tasks and highlight considerations that can enhance its applicability and robustness in dynamic environments.