基于神经网络的机器人学习与控制:回顾与展望  

Learning and Control of Robots Based on Neural Networks: Review and Outlook

作　　者：谢正泰 樊佳亮 刘梅 金龙 XIE Zhengtai;FAN Jialiang;LIU Mei;JIN Long(School of Information Science and Engineering,Lanzhou University,Lanzhou 730000,China;Chongqing Key Laboratory of Big Data and Intelligent Computing,Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences,Chongqing 400714,China)

机构地区：[1]兰州大学信息科学与工程学院,甘肃兰州730000 [2]中国科学院重庆绿色智能技术研究院大数据与智能计算重庆市重点实验室,重庆400714

出　　处：《信息与控制》2023年第1期37-58,共22页Information and Control

基　　金：国家自然科学基金(62176109);腾讯RoboticsX犀牛鸟专项研究计划(2021-01);甘肃省自然科学基金项目(21JR7RA531,22JR5RA427,22JR5RA487);中央高校基本科研业务费(lzujbky-2021-65,lzujbky-2022-it02,lzujbky-2022-23);兰州大学超算中心。

摘　　要：机器人因其高效的感知、决策和执行能力,在人工智能、信息技术和智能制造等领域中具有巨大的应用价值。目前,机器人学习与控制已成为机器人研究领域的重要前沿技术之一。各种基于神经网络的智能算法被设计,从而为机器人系统提供同步学习与控制的规划框架。首先从神经动力学(ND)算法、前馈神经网络(FNNs)、递归神经网络(RNNs)和强化学习(RL)四个方面介绍了基于神经网络的机器人学习与控制的研究现状,回顾了近30年来面向机器人学习与控制的智能算法和相关应用技术。最后展望了该领域存在的问题和发展趋势,以期促进机器人学习与控制理论的推广及应用场景的拓展。Robots have great application value in the fields of artificial intelligence, information technology, and intelligent manufacturing due to their efficient perception, decision-making, and execution capabilities. At present, robot learning and control have become one of the most critical frontier technologies in the field of robotics. Meanwhile, different intelligent algorithms based on neural networks have been designed to provide a planning framework for synchronous learning and control of robot systems. Specifically, the research status of neural-network-based robot learning and control is reviewed from four aspects: neural dynamics(ND) algorithms, feedforward neural networks(FNNs), recurrent neural networks(RNNs), and reinforcement learning(RL). The intelligent algorithms and related application technologies utilized for robot learning and control in the past three decades are reviewed in detail. Finally, the remaining challenges and development trends in this field are provided to promote the development of robot learning and control theory and the extension of application scenarios.

关 键 词：机器人学习与控制 神经动力学方法 前馈神经网络 递归神经网络 强化学习 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]