基于深度强化学习的机器人运动控制研究进展

复杂未知环境下智能感知与自动控制是目前机器人在控制领域的研究热点之一,而新一代人工智能为其实现智能自动化赋予了可能.近年来,在高维连续状态-动作空间中,尝试运用深度强化学习进行机器人运动控制的新兴方法受到了相关研究人员的关注.首先,回顾了深度强化学习的兴起与发展,将用于机器人运动控制的深度强化学习算法分为基于值函数和策略梯度2类,并对各自典型算法及其特点进行了详细介绍;其次,针对仿真至现实之前的学习过程,简要介绍5种常用于深度强化学习的机器人运动控制仿真平台;然后,根据研究类型的不同,综述了目前基于深度强化学习的机器人运动控制方法在自主导航、物体抓取、步态控制、人机协作以及群体协同等5个方面的研究进展;最后,对其未来所面临的挑战以及发展趋势进行了总结与展望.

Research progress of robot motion control based on deep reinforcement learning

Intelligent perception and automatic control in a complex unknown environment is one of the current research hotspots of robots in the field of control, and a new generation of artificial intelligence makes it possible to realize intelligent automation. In recent years, the new method of robot control using deep reinforcement learning in high-dimensional continuous state-action space has attracted the attention of relevant researchers. Firstly, the rise and development of deep reinforcement learning are first reviewed. The deep reinforcement learning algorithms for robot motion control are classified into two categories: value-based functions and policy gradients, and their typical algorithms and their related features are detailly described. Then, for the learning process before simulation to reality, five kinds of simulation platforms for robot motion control are briefly introduced, which are often used for deep reinforcement learning. Moreover, according to different types of research, the research progress of the deep reinforcement learning approach of robot motion control is expounded in five aspects, including autonomous navigation, object grasping, gait control, human-robot collaborative and multi-robot cooperation. Finally, the future challenges and development trends are summarized and anticipated.