基于卷积神经网络的改进机械臂抓取方法

An Improved Manipulator Grasping Method Based on Convolutional Neural Network

为了提高机械臂抓取的精度,提出一种基于Mask R-CNN的机械臂抓取最佳位置检测框架。基于RGB-D图像,所提框架通过精确的实例分割确定抓取对象的类别、位置和掩码信息,由反距离加权法在去噪后的深度图上获取中心点的加权深度坐标,构成目标对象的三维目标位置,经坐标系转换得到最终的最优抓取位置。建议的框架考虑到目标对象的姿态与边缘信息,可以有效地提高抓取性能。最后,基于UR3机械臂上的抓取实验结果验证了该框架的有效性。

In order to improve the accuracy of manipulator grasping,an efficient framework is proposed for detect the optimal position of robotic grasping based on Mask R-CNN.The architecture,w hich uses RGB-D images as input,m akes an accurate in stance segmentation to determine the category,location and mask information of grabbing objects.Then the weighted depth coordi nates of the center point are obtained from the denoised depth map by the inverse distance weighted method to form the three-dimen sional target position.And the final optimal grasping position is obtained by coordinate transformation.As such,the performance can be effectively improved by considering the posture and edge information of target object.Finally,some experiments for grasping network on Universal Robot 3 are utilized to demonstrate the effectiveness of the proposed framework.