The Impact of Data Augmentation on Tactile-Based Object Classification Using Deep Learning Approach

摘要

A safe and versatile interaction between humans and objects is based on tactile and visual information. In literature, visual sensing is widely explored compared to tactile sensing, despite showing significant limitations in environments with an obstructed view. Tactile perception does not depend on those factors. In this paper, a Machine Learning-based tactile object classification approach is presented. The hardware setup is composed of a 3-finger-gripper of a robotic manipulator mounted on the Doro robot of the Robot-Era project. This paper’s main contribution is the augmentation of the custom 20 class 2000 sample tactile time-series dataset using random jitter noise, scaling, magnitude, time warping, and cropping. The effect on the object recognition performance of the dataset expansion is investigated for the neural network architectures MLP, LSTM, CNN, CNNLSTM, ConvLSTM, and deep CNN (D-CNN). The data augmentation methods brought a statistically significant object classification accuracy increase compared to models trained on the original dataset. The best tactile object classification success rate of 72.58% is achieved for the D-CNN trained on an augmented dataset derived from scaling and time warping augmentation.