부산대학교 도서관

The basic thing to consider when building a quadruped robot is the issue of balance. These factors greatly determine the success of the quadruped robot in carrying out movements such as stabilizing the body on an inclined plane, walking movements and others. Conventional feedback control methods by performing mathematical modeling can be used to balance the robot. However, this method still has weaknesses. The application of conventional feedback control methods often results in an inaccurate controller, so it must be manually tuned for its application. In this study, reinforcement learning methods were used using Q-Learning algorithms. The use of reinforcement learning methods was chosen because no mathematical calculations are needed to control the balance of quadruped robots. The process of learning the system to train the agent's abilities is carried out using a Gazebo simulator. The learning results show that the system could run well as evidenced by the higher value of sum rewards per episode.