深層強化学習と形状空間上のゲージ場を用いた宇宙ロボットの姿勢制御と形態進化

抄録

With the recent development of novel spacecraft such as solar sails and microsatellites, active use of shape maneuvers has gained increasing interest for the purpose of attitude control. However, to use shape maneuvers effectively for attitude control, it is critically important and challenging to design the morphologies of spacecraft in an appropriate manner. In this study, we apply deep reinforcement learning to find shape maneuvers of an origami-inspired space robot that achieves attitude control under conditions of vanishing total angular momentum. The space robot consists of four triangular panels connected to the edges of a central square panel. We evolve the morphology of the space robot via particle swarm optimization, employing an objective function approximated by the Q-value function obtained from the deep reinforcement learning for attitude control. As a result of morphology evolution, the triangular panels of the space robot became narrower and longer. Moreover, the evolved space robot successfully reduced the control time for attitude changes.