Tactile Convolutional Networks for Online Slip and Rotation Detection Martin Meier, Florian Patzelt, Robert Haschke and Helge J. Ritter Neuroinformatics Group, Center of Excellence Cognitive Interaction Technology (CITEC), Bielefeld University, Germany {mmeier,fpatzelt,rhaschke,helge}@techfak.uni-bielefeld.de Abstract. We present a deep convolutional neural network which is capable to distinguish between different contact states in robotic ma- nipulation tasks. By integrating spatial and temporal tactile sensor data from a piezo-resistive sensor array through deep learning techniques, the network is not only able to classify the contact state into stable ver- sus slipping, but also to distinguish between rotational and translation slippage. We evaluated different network layouts and reached a final clas- sification rate of more than 97%. Using consumer class GPUs, slippage and rotation events can be detected within 10 ms, which is still feasible for adaptive grasp control. 1 Introduction In autonomous robotic manipulation tasks, for example grasping and placing objects, estimating the stability of the object in hand plays a major role. Objects may slip out of the manipulator. This can lead to a state in the desired action sequence from which the system cannot recover easily. Due to occlusions, vision- based systems can hardly keep track of the state of objects hold in manipulators and are therefore of limited usefulness when it comes to detecting loss of grasp stability. For that reason, the loss of an object can only be detected after such events already occurred. Humans perceive the onset of slippage by sensing high- frequency micro-vibrations through specialized nerves (Pacinian corpuscle) in the skin [4]. One possibility for early detection of slippage events in robotic systems is the integration of tactile sensing capabilities directly into robotic manipulators. By having human like sensing skills, the system should be able to directly evaluate the contact state during interactions. Compared to imaging technologies where standards are established for data acquisition and representation, current tactile sensors posses a large variety of data acquisition techniques, which can be either based on electric [12], optic [15] or acoustic [6] effects. For example the authors in [2] discuss eight different technologies which are based on these three effects and are used in current state of the art tactile sensors. For a detailed technical overview the interested reader is referred to [2]. The work presented in [13] used support vector machines and random forests to detect object slippage with a BioTac [6] sensor. The BioTac sensor offers mul- tiple modalities such as 19 electrodes to measure local contacts with a sampling