부산대학교 도서관

Deep reinforcement learning (DRL) exhibits a promising approach for controlling humanoid robot locomotion. However, only values relating sensors such as IMU, gyroscope, and GPS are not sufficient robots to learn their locomotion skills. In this article, we aim to show the success of vision based DRL. We propose a new vision based deep reinforcement learning algorithm for the locomotion of the Robotis-op2 humanoid robot for the first time. In experimental setup, we construct the locomotion of humanoid robot in a specific environment in the Webots software. We use Double Dueling Q Networks (D3QN) and Deep Q Networks (DQN) that are a kind of reinforcement learning algorithm. We present the performance of vision based DRL algorithm on a locomotion experiment. The experimental results show that D3QN is better than DQN in that stable locomotion and fast training and the vision based DRL algorithms will be successfully able to use at the other complex environments and applications.