모바일 로봇의 자율 네비게이션을 위한 심층 강화 학습 기반자연행동 학습 (Natural Behavior Learning Based on Deep Reinforcement Learning for Autonomous Navigation of Mobile Robots)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 24권 / 3호 / 256 ~ 262페이지
· 저자명 : 김기서, 문정환, 김동규, 조승범, 이장명

초록

This paper proposes a new autonomous navigation method for two-wheeled mobile robots in an unknown environment using LiDAR sensors. Recently, a deep Q-network (DQN), which is a combination of deep learning and Q-learning theory, has attracted attention as a reinforcement learning algorithm. It is used to train the robot to recognize obstacles and avoid collisions while it is moving toward a designated destination. However, the existing DQN method can handle only discrete tasks and low-dimensional spaces, meaning it is not suitable for continuous, especially the control of mobile robots. In addition, the existing LiDAR sensor method uses distance as the input for learning, with the system determining the next action based only on the distance from the mobile robot to the obstacle. In this process, due to the frequent fluctuation in the action values, unnatural acceleration/deceleration actions are required, which leads to physical shocks to the robot and low driving power efficiency. In this paper, this problem is solved by applying a replay buffer to store the output of the network. That is, the action values are stored in the memory and fed back as input again following the action order of the network. Experiments are carried out on an actual robot after reinforcement learning in ROS-GAZEBO simulations and the validity of the algorithm is verified through an analysis of the experimental data.

영어초록

This paper proposes a new autonomous navigation method for two-wheeled mobile robots in an unknown environment using LiDAR sensors. Recently, a deep Q-network (DQN), which is a combination of deep learning and Q-learning theory, has attracted attention as a reinforcement learning algorithm. It is used to train the robot to recognize obstacles and avoid collisions while it is moving toward a designated destination. However, the existing DQN method can handle only discrete tasks and low-dimensional spaces, meaning it is not suitable for continuous, especially the control of mobile robots. In addition, the existing LiDAR sensor method uses distance as the input for learning, with the system determining the next action based only on the distance from the mobile robot to the obstacle. In this process, due to the frequent fluctuation in the action values, unnatural acceleration/deceleration actions are required, which leads to physical shocks to the robot and low driving power efficiency. In this paper, this problem is solved by applying a replay buffer to store the output of the network. That is, the action values are stored in the memory and fed back as input again following the action order of the network. Experiments are carried out on an actual robot after reinforcement learning in ROS-GAZEBO simulations and the validity of the algorithm is verified through an analysis of the experimental data.

참고자료

· 없음