Visualization of Tactile Features for Object Recognition with a Multi-Fingered Hand

摘要: Multi-fingered robot hands can be extremely effective to physically explore and recognize objects, especially if they are extensively covered with distributed tactile sensors. Convolutional Neural Networks (CNNs) have been proved successful in processing high dimensional tactile information and we introduced a Morphology-Specific CNN (MS-CNN) in which hierarchical convolutional layers which were formed following the physical configuration of the tactile sensors on the robot. However, why the network achieved high recognition rates of objects was not revealed. In this study, Grad-CAM++ as one of visualization methods which is suitable for CNN architectures is utilized. From the visualization result, we investigated which MS-CNN architecture enables the robot hand to successfully recognize 9 types of physical properties of objects by a single touch. A recognition rate of over 95% was achieved.