強化学習を用いた慣性ロータ型倒立振子の安定化制御

抄録

The reaction wheel is a mechanical device for stabilizing the attitude of the controlled object by the reaction torque generated when the flywheel is rotated. In order to operate the reaction wheel as an attitude stabilizing device, it is necessary to construct a control system for this reaction wheel, and the attitude stabilizing performance depends on the performance of this control system. In the design of the control system, all parameters of the control object need to be known. However, there are many unknown parameters, which makes the control system design difficult. Therefore, we apply reinforcement learning to control system design. By applying reinforcement learning, it is possible to construct a control system that learns the optimal operation using only input / output data related to the state and action of the control object. Using a 3D dynamic simulator equipped with multiple physics engines as a reinforcement learning environment, a real machine model of reaction wheel type inverted pendulum is constructed in 3D virtual space. In addition, by using a robot operation system (ROS) equipped with an adjustment function with Gazebo, a reinforcement learning algorithm is implemented on ROS.