基于分层学习的四足机器人运动自适应控制模型  

A Quadruped Robot Motion Adaptive Model Based on Hierarchical Learning

作　　者：崔俊文 刘自红[1] 石磊[1] 刘福强 乐玉 Cui Junwen;Liu Zihong;Shi Lei;Liu Fuqiang;Yue Yu(Southwest Unversity of Technology and Science,Mianyang 621010,China)

机构地区：[1]西南科技大学制造科学与工程学院

出　　处：《计算机测量与控制》2020年第1期105-110,共6页Computer Measurement &Control

基　　金：四川省大学生创新创业训练项目基金（S201910619035）

摘　　要：针对四足机器人面对腿部损伤无法继续有效自主运作的问题,提出一种基于分层学习的自适应控制模型;该模型结构由上层状态策略控制器(SDC)和下层基础运动控制器(BDC)组成;SDC对机器人腿部及姿态进行决策并选择运动子策略,BDC子运动策略表达该状态下机器人的运动行为;在Unity3D中构建反关节多自由度的四足机器人,训练多种腿部受损状况的BDC子运动策略,BDC成熟后20s周期随机腿部受损并训练SDC;该模型控制流程为SDC监测机器人状态,激活BDC策略,BDC输出期望关节角度,最后由PD控制器进行速度控制;其实现机器人在腿部受损后自我适应继续保持运作;仿真与实验结果表明,该控制模型能在机器人损伤后能自我快速、稳定调整运动策略,并保证运动的连贯性及柔和性。Aiming at the problem that the quadruped robot can not continue to operate effectively and independently after the leg accidently be damaged,the adaptive control model based on hierarchical learning is proposed.The model structure consists of an upper state policy controller(SDC)and a lower base motion controller(BDC).The SDC estimates the expected motion sub-strategy of the robot’s legs and posture,and the BDC sub-motion strategy is activated to control the robot to express the athletic behavior.Damage to the robot is manifested in the complete loss of athletic ability in any leg.The adaptive control of the model is reflected in the robot’s self-adjusting strategy after the leg fails.In Unity 3D,a four-legged robot with anti-joint multi-degree of freedom is built.The SDC monitors the state of the robot and adjusts the strategy.The BDC output gives the joint PD controller speed control.Simulation and experimental results show that the model shows a fast and stable effect on the robot’s self-adjusting motion strategy after the legfails.

关 键 词：分层学习 深度强化学习 四足机器人 部分马尔可夫决策 步态控制 机构失效 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]