Towards Multidimensional Textural Perception and Classification Through Whisker
Texture-based studies and designs have been in focus recently. Whisker-based multidimensional surface texture data is missing in the literature. This data is critical for robotics and machine perception algorithms in the classification and regression of textural surfaces. In this study, we present a novel sensor design to acquire multidimensional texture information. The surface texture's roughness and hardness were measured experimentally using sweeping and dabbing. Three machine learning models (SVM, RF, and MLP) showed excellent classification accuracy for the roughness and hardness of surface textures. We show that the combination of pressure and accelerometer data, collected from a standard machined specimen using the whisker sensor, improves classification accuracy. Further, we experimentally validate that the sensor can classify texture with roughness depths as low as $2.5μm$ at an accuracy of $90\%$ or more and segregate materials based on their roughness and hardness. We present a novel metric to consider while designing a whisker sensor to guarantee the quality of texture data acquisition beforehand. The machine learning model performance was validated against the data collected from the laser sensor from the same set of surface textures. As part of our work, we are releasing two-dimensional texture data: roughness and hardness to the research community.