摘要
基于视觉伺服控制的机器人手眼标定和逆运动学求解一直是视觉伺服领域的核心问题。随着应用场景的逐渐复杂,传统手眼标定方法已无法满足需求;随着网络深度的增加,单一BP神经网络逆运动学求解算法的误差趋于饱和甚至变大,无法进一步提升网络性能。针对以上问题,本文将手眼标定和逆运动学求解融合为拟合目标图像坐标到机器人各关节角之间的映射关系问题,提出了一种残差BP神经网络算法。使用多个残差网络模块的方式加深BP神经网络的深度,残差模块的输入信息可以在网络内跨层传输,较好地解决了因深度增加网络模型容易产生梯度消失而无法提升网络性能的问题;通过6自由度机器人雅可比方程对逆运动学解的空间进行划分,确定了8个独立的区域,基于独立区域方法对训练数据进行处理,从而避免了多自由度机器人逆运动学多解对网络学习的影响,网络训练精度提升了2个数量级,训练速度提高了2倍。在REBot-V-6R型6自由度机器人输送线分拣系统中进行二维平面抓取和三维实物抓取实验,实验结果验证了该方法的准确性。结果表明,该方法比1层BP神经网络、3层BP神经网络、5层BP神经网络的训练精度分别提高了4个数量级、2个数量级、5个数量级,测试精度提高2个数量级;与传统标定方法相比,本文方法节约了逆运动学求解过程的计算成本,抓取位姿精度提高了1个数量级。
The hand-eye calibration and inverse kinematics solution of the 6-degree-of-freedom robot based on visual servo control has always been the core problem in this field. With the application scene becoming more complex,the traditional hand-eye calibration method cannot meet requirements. At the same time,based on the single BP neural network inverse kinematics algorithm,the error tends to be saturated or even larger,which cannot further improve the network performance with the increase of network depth to a certain extent. In order to solve the above problems,the problem of hand-eye calibration and inverse kinematics was integrated into the problem of fitting the mapping relationship between the coordinates of the target image and the joint angles of the 6-degree-of-freedom series robot,and a residual BP neural network algorithm was proposed,the multiple residual network modules were used to deepen the depth of the BP neural network,and the input information of the residual module can be transmitted across layers in the network. It solved the problem that the gradient of the network model was easy to disappear and cannot improve the network performance with the increase of depth. In addition,the space of the inverse kinematics solutions was divided into eight unique regions by the sixdegree-of-freedom robot Jacobi equation,and the training data were processed based on this way,the influence of multi-solution of inverse kinematics of multi-degree-of-freedom robot on network learning was avoided,and the accuracy of network training results was improved by two orders of magnitude,and the training speed was increased by two times. Finally, two-dimensional plane grasping and threedimensional physical grasping experiments were carried out in the REBot-V-6 R 6 degree of freedom robot conveyor line sorting system,and the experimental results verified the accuracy of the method. Compared with single-layer BP neural network,three-layer BP neural network and five-layer BP neural network,the training accuracy was improved by four orders of magnitude,two orders of magnitude and five orders of magnitude,respectively,and the testing accuracy was improved by two orders of magnitude,and the computational cost of inverse kinematics was saved and the accuracy was improved by one order of magnitude.
作者
李光
章晓峰
杨加超
马祺杰
LI Guang;ZHANG Xiaofeng;YANG Jiachao;MA Qijie(College of Mechanical Engineering,Hunan University of Technology,Zhuzhou 412007,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2021年第4期366-374,共9页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家自然科学基金项目(11602082)
湖南省自然科学基金项目(2018JJ4079)。
关键词
6自由度机器人
BP神经网络
残差网络
手眼标定
视觉伺服
逆运动学
six degree of freedom robot
BP neural network
residual network
hand-eye calibration
visual servo
inverse kinematic
