摘要
智能抓取机器人能够代替人类完成高强度工作,为实现物体的准确定位,提升机器人抓取的成功率,对基于感兴趣区域的机器人抓取系统进行研究。对深度相机进行标定,对深度卷积神经网络损失函数进行改进,使用焦点函数代替传统的交叉熵函数,训练模型,得到目标的类别、二维包络框中目标的像素坐标值与深度值等信息。利用手眼标定方法将深度传感器坐标转换到机械臂基坐标系下,依据相机成像原理完成物体的定位。通过机器人逆运动学求解关节角度,驱动机器人实现抓取。对实验过程进行分析,在aubo_i5机械臂上进行实验验证。实验结果表明,目标的识别定位误差较小,平均精度值提升了2.36%,抓取的平均成功率达到93.4%,较改进前提升了13.4%,能够满足机器人抓取的需求。
Intelligent grasping robot can replace human to perform a high-intensity works. To realize the accurate positioning of objects and improve the success rate of robot grasping, a robot grasping system was designed based on region of interest. The system calibrated the depth camera, and then improved the loss function of the deep convolution neural network.The focus function was used instead of the traditional cross entropy function, by which the model was trained to get the category of the target, the pixel coordinate value, and depth value of the target in the two-dimensional envelope box. The coordinates of the depth sensor were transformed into the manipulator coordinate system by hand-eye calibration method, and the positioning of the object was completed according to the imaging principle of the camera. The joint angle was determined in inverse kinematics of the robot, which drove the robot to grasp. The experimental course was analyzed and the results were verified with the aubo_i5 manipulator. Results show that the target recognition and positioning error was small, the average accuracy was increased by 2.36%, the average success rate of grasping was 93.4%, which is 13.4% higher than that before the improvement, which can meet the needs of robot grasping.
作者
马世超
孙磊
何宏
郭延华
MA Shi-chao;SUN Lei;HE Hong;GUO Yan-hua(School of Electrical and Electronic Engineering College,Tianjin University of Technology,Tianjin 300384,China)
出处
《科学技术与工程》
北大核心
2020年第11期4395-4403,共9页
Science Technology and Engineering
基金
国家自然科学基金(61871173)
天津市科技计划项目(18YFZCSF00600,18ZXJMTG00160,18ZXZNSY00270)。
关键词
机器人
卷积神经网络
深度传感器
目标检测
目标定位
物体抓取
robot
convolution neural network
depth sensor
object detection
object location
object grasping
