Unsupervised Learning of Spatio-Temporal Features for Tactile Data

Tactile feedback obtained during grasping of an object is crucial for a number of tasks in robotics. It allows to identify object properties and recognize its class, or to localize an object in a robot hand and assess quality and stability of the applied grasp [1]–[3]. The accomplishment of these tasks greatly depends on the choice of data representation, i.e. features that are used to capture signal properties, as it was was shown for visual data [4]–[7]. The representation of tactile data should not only provide robustness to real-world conditions, but also serve a wide variety of applications and be flexible to adapt to their specific requirements. In previous works tactile signal have been typically represented using a manually crafted set of features based on prior knowledge about the properties of the inputs. Early works aimed at identifying simple primitive shapes in object imprints in the tactile matrices (points and lines [8], [9]). Recent works use higher level geometric properties of pressure patterns, such as their position, area or higherorder moments [1], [3], [10], [11]. Other approaches build on features developed at Computer Vision, such as SIFT [12]. A very limited amount of work have aimed at unsupervised extraction of features by applying to tactile data the K-means algorithm or covariance analysis [13]–[15].