Geometric Transformation: Tactile Data Augmentation for Robotic Learning

In recent years, the popularity of robotic tactile learning has grown with data-driven approaches being widely utilized for various tasks. It is generally acknowledged that a large and high quality dataset is critical for a data-driven approach. However, collecting such a dataset is challenging and time consuming with a real robot and tactile sensors. Instead, data augmentation techniques are widely used in visual-based learning because it could increase the size/quality of the dataset and improve the learning performance. In this letter, by considering properties of tactile data, we propose geometry transformation tactile data augmentation methods by adopting conventional image data augmentation methods. Our proposal is tested on two different tactile learning tasks, slip prediction and object recognition on a real robot. Two datasets are collected and our datasets are publicly available. The results show that our proposal could improve the learning accuracy additionally by up to 8.43%. Furthermore, our proposal outperforms conventional image data augmentation methods on tactile learning task overall.