強化学習を活用した月着陸軌道制御のレジリエンス向上

抄録

Space missions require “resilience” to flexibly complete the mission in response to changes in the environment and system characteristics. The present study proposes a method for autonomously planning a corrective control law for lunar landing trajectory control to cope with off-nominal conditions and reflecting it in resilience improvement measures by utilizing reinforcement learning. The proposed method employs a reinforcement learning problem in which an agent is additionally placed in the control loop and the corrective control input as an action output by the agent is added to the original closed-loop control input. The results and insights are summarized for the resultant agent's characteristics which autonomously detect off-nominal conditions and proactively implement recovery measures, while verifying the capability and effectiveness of the proposed design framework enabled by a reinforcement learning architecture in a realistic and specific lunar landing sequence.