During the execution of handling processes in manufacturing, it is difficult to measure the process forces with state-of-the-art gripper systems since they usually lack integrated sensors. Thus, the exact state of the gripped object and the actuating process forces during manipulation and handling are unknown. This paper proposes a deep learning regression model to construct a continuous stability metric to predict the maximum process forces on the gripped objects using high-resolution optical tactile sensors. A pull experiment was developed to obtain a valid dataset for training. Continuously force-based labeled pairs of tactile images for varying grip positions of industrial gearbox parts were acquired to train a novel neural network inspired by encoder-decoder architectures. A ResNet-18 model was used for comparison. Both models can predict the maximum process force for each object with a precision of less than 1 N. During validation, the generalization potential of the proposed methodology with respect to previously unknown objects was demonstrated with an accuracy of 0.4-2.1 N and precision of 1.7-3.4 N, respectively.

翻译：在制造过程中进行工件搬运作业时，由于现有夹爪系统通常缺乏集成传感器，过程力难以测量。因此，操作与搬运过程中被抓物体的确切状态及施加的过程力均处于未知状态。本文提出一种深度学习回归模型，利用高分辨率光学触觉传感器构建连续稳定性度量指标，以预测被抓物体上的最大过程力。通过拉力实验获取有效数据集进行训练，针对工业齿轮箱零件在不同抓取位置下，采集基于连续力标签的触觉图像对，训练受编码器-解码器架构启发的新型神经网络。使用ResNet-18模型进行对比实验。两种模型均能以小于1 N的精度预测各物体的最大过程力。在验证阶段，所提方法对于未知物体的泛化潜力得到验证，其准确度为0.4-2.1 N，精度分别为1.7-3.4 N。