提出了一种基于深度神经网络的机械臂最优抓取位置检测方法.相比传统手工设定的特征,基于深度神经网络的方法学习得到的特征具有较强的鲁棒性和稳定性,能够适应训练集中未曾出现的新物体.本方法首先使用基于深度学习的目标检测算法对图像中的目标物体进行检测,记录目标的类别和位置.然后根据分类检测结果,使用基于深度学习的机械臂抓取方法进行抓取位置学习.仿真实验表明所提方法能对图像中的目标物体进行较为准确的分类,在Universal Robot 5机械臂上得到的抓取实验结果证明了所提方法的有效性.
A method is proposed to detect the optimal position of robotic grasping based on deep neural network. Com-pared with conventional manually-set features, the features learned by deep neural network methods are more robust and stabler, and can be applied to objects outside of the training set. In this method, the object detection algorithm based on deep learning is first used to detect the objects in the image with the classes and locations of the objects recorded. Then, the robotic grasping method based on deep learning is used to learn the grasping positions according to the object classification and detection results. Simulation experiments indicate that the proposed method can classify the objects in the images accurately, and the grasping experimental results on Universal Robot 5 verify the effectiveness of the proposed method.