Embedded Artificial Intelligence for Tactile Sensing

Electronic tactile sensing becomes an active research field whether for prosthetic applications, robotics, virtual reality or post stroke patients rehabilitation. To achieve such sensing, an array of sensors is used to retrieve human-skin like information, which is called Electronic skin (E-skin). Humans through their skins, are able to collect different types of information e.g. pressure, temperature, texture, etc. which are then passed to the nervous system, and finally to the brain in order to extract high level information from these sensory data. In order to make E-skin capable of such task, data acquired from E-skin should be filtered, processed, and then conveyed to the user (or robot). Processing these sensory information, should occur in real-time, taking in consideration the power limitation in such applications, especially prosthetic applications. The power consumption itself is related to different factors, one factor is the complexity of the algorithm e.g. number of FLOPs, and another is the memory consumption. In this thesis, I will focus on the processing of real tactile information, by 1)exploring different algorithms and methods for tactile data classification, 2)data organization and preprocessing of such tactile data and 3)hardware implementation. More precisely the focus will be on deep learning algorithms for tactile data processing mainly CNNs and RNNs, with energy-efficient embedded implementations. The proposed solution has proved less memory, FLOPs, and latency compared to the state of art (including tensorial SVM), applied to real tactile sensors data. Keywords: E-skin, tactile data processing, deep learning, CNN, RNN, LSTM, GRU, embedded, energy-efficient algorithms, edge computing, artificial intelligence.