Tactile Contact Patterns for Robotic Grasping: A Dataset of Real and Simulated Data

Advancing tactile sensing in robotics and machine learning necessitates high-quality datasets encompassing realworld and simulated interactions. In this paper, we present a comprehensive dataset containing 46,200 samples collected from a deformable, capacitive-based tactile sensor. The dataset is equally divided into three main groups: 15,400 real samples, 15,400 synthetic samples generated using Abaqus, and 15,400 synthetic samples generated using Isaac Gym through finite element analysis (FEA). Data acquisition was performed under two experimental scenarios. In the first scenario, 49 unique indenters were pressed onto the sensor at various force levels, producing various contact patterns. In the second scenario, the sensor was integrated into a 2F-85 Robotiq parallel gripper to grasp 12 different objects. We provide a detailed account of the dataset construction process, elaborate on its composition, and introduce a graphical user interface that enables the creation of customized datasets tailored to specific application needs. Ultimately, we present a case study employing Transfer Learning to exemplify the dataset's potential by utilizing real and synthetic data to recognize surface types (flat or curved), showcasing how synthetic data can be effectively leveraged alongside real data to enhance performance. To access the code and resources used in this research, all files are available in our GitHub repository at [TactileDataset](https://github.com/Lab-CORO/TactileDataset).