Texture classification on uneven surfaces using deep learning techniques

Abstract

Robots are increasingly essential in various fields, excelling in tasks from routine operations to hazardous situations. Enhancing robots with human-like capabilities, such as tactile sensing, broadens their potential applications. Tactile sensors enable robots to perceive and interact with their environment similarly to humans. This research focuses on leveraging tactile sensors to classify textures on uneven surfaces, an area previously unexplored in the literature. By collecting data points along predefined paths on object surfaces, we minimized assumptions about the object’s geometry, making the system more flexible and adaptable. These data points guided the robot’s trajectory, during which tactile data were systematically gathered on the surface of uneven objects, marking a pioneering effort in this area. To improve texture classification and reduce processing time, we employed a sliding window approach, segmenting the dataset into smaller overlapping windows for multi-scale analysis. In addition to data from uneven surfaces, we supplemented our dataset with tactile data from even surfaces from another study. We applied advanced deep learning models, including convolutional neural networks (1D CNN), recurrent neural networks (bidirectional LSTM), and hybrid architectures, to classify tactile textures using time-series data. The models achieved average accuracy, precision, and recall rates of 92.3%, 92.4%, and 92.3% for uneven surfaces, and 96.9%, 97.0%, and 97.0% for even surfaces. This study demonstrates the importance of tactile sensing in robotic systems, particularly for texture classification on uneven surfaces. By incorporating MARG and barometer sensors into the Open Manipulator X, this research advances tactile perception in robotics, equipping robots to interact more effectively with diverse environments. The findings set the stage for future applications where precise tactile perception is essential.