심층 강화학습을 이용한 로봇의 자율 경로 탐색 알고리즘

Abstract

A deep reinforcement algorithm is an algorithm that successfully approximates components of a reinforcement learning algorithm with a deep neural network. And the most representative deep reinforcement learning algorithm is DQN (Deep Q-Network). DQN allows an agent to get a high score in various games with high-dimensional raw images as input. There have been many attempts to apply DQN to the high-level tasks of robots, but few have been successful. This is because most of the robot's tasks take place in a higher-order state-action space. Nevertheless, there are successful cases of a mobile robot and a robot manipulator using DQN and additional techniques to accomplish their intended tasks. In this research, we propose an autonomous path search method for a mobile robot to find a destination in an environment with numerous obstacles and a robot manipulator to avoid obstacles in real time and perform pick and place tasks. This could be achieved due to an improved DQN and simplified representation of the work environment. And we confirmed that autonomous path search of a mobile robot and a robot manipulator is possible through model training and simulation results.