摘要
针对移动机器人在未知环境下的无图导航问题,本文提出了一种基于深度强化学习的端到端的控制方法。机器人需要在没有地图的情况下,仅仅依靠视觉传感器的RGB图像以及与目标之间的相对位置作为输入,来完成导航任务并避开沿途的障碍物。在任意构建的仿真环境中,基于学习策略的机器人可以快速适应陌生场景最终到达目标位置,并且不需要任何人为标记。实验表明,这种端到端的控制策略可以实现仿真环境中的导航任务,且与普通离散控制的深度强化学习的方法相比,机器人学习导航策略的平均收敛时间降低了75%。
Collision-free exploration in an unknown environment is a core function of mobile robots. This paper proposes an end-to-end control method based on deep reinforcement learning. The robot needs to rely solely on the RGB image of the vision sensor and the relative position to the target as input to complete the navigation task and avoid obstacles. In an arbitrarily constructed simulation environment,a learning strategy- based robot can quickly adapt to unfamiliar scenes and eventually reach the target location without any artifacts. Experiments show that this end- to- end control strategy can achieve navigation tasks in the simulation environment,and the average convergence time of the robot learning navigation strategy is reduced by 75% compared with the classical deep reinforcement learning method.
作者
陈杰
程胜
石林
CHEN Jie;CHENG Sheng;SHI Lin(The China Manned Space Engineering of Ice,Beijing 100083,China;China Aerospace Science and Technology Corporation,Beijing 100094,China;School of Computer Science,Northwestern Polytechnical University,Xi’an 710072,China)
出处
《电子设计工程》
2019年第15期61-65,共5页
Electronic Design Engineering
关键词
自主导航
深度强化学习
自动控制
移动机器人
autonomous navigation
deep reinforcement learning
automatic control
mobile robot
