深層強化学習を用いたジャイロスタットの角速度制御

抄録

In recent years, deep reinforcement learning has been used in many applications including spacecraft attitude control. While attitude control of rigid bodies by deep reinforcement learning has been studied in many works, angular velocity and attitude control of gyrostats by deep reinforcement learning remain to be done. The purpose of this study is to control angular velocity and attitude of gyrostats by deep reinforcement learning and to explore more efficient control of gyrostats. The gyrostat of this study has three reaction wheels each of which is installed in each direction of principal axes of inertia. The reaction wheels are controlled by a method of deep reinforcement learning called Proximal Policy Optimization (PPO). Results show that robust control of angular velocity and attitude of the gyrostat is possible via PPO. In addition, value function mapped on the angular momentum sphere of the gyrostat tends to be higher along invariant manifolds of the dynamical system. These results indicate that deep reinforcement learning will be effective for angular velocity and attitude control of gyrostats in future space missions. Moreover, the flow in phase space may make it easier to control the attitudes of gyrostats.