On explainability of tactile data representation for robots

• Main Author: Tian, Tian
• Other Authors: Lin Zhiping
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/150209
• Institution: Nanyang Technological University

The ability of texture classification is highly desired for robots. Current machine learning models working on texture classification tasks, despite having high accuracy, are inherently limited in interpretability. Therefore, modifications for better explainability are expected. In this work, we propose an explainable and efficient representation learning method. Texture property, stiffness and roughness characteristics specifically, as two crucial aspects when humans distinguish among different materials, are collected in a low-cost manner and infused to bootstrap the representation learning process in order to achieve better performance. In particular, two dedicated neurons in the representation vector are assigned to learn stiffness and roughness. By leveraging on these properties with physical notion in the latent space, the proposed method improves the classification accuracy. By testing on degrading the number of training samples and length of latent vector, we demonstrate that our method is able to retain performance under adverse condition. In general, the proposed representation learning method provides improvements in texture classification accuracy and efficiency as well as the interpretability of the latent representation.