| Tactile skins can enhance a robot's capabilities in many application domains, including whole-body manipulation, human-robot communication, and safe interaction. To support these enhancements, the expectation is that tactile skins will generate megabits of data every millisecond that need to be transferred to a central processing system. Due to the sheer volume of data and the distributed nature of tactile skins, acquiring the data at the processing system at the desired rate is a challenging problem. In this dissertation, we present our work to help address this problem.;First, we introduce the open-source tactile skin simulator BubbleTouch for generating tactile data since current tactile skin systems are not widely available to produce data for research. BubbleTouch uses kinematic and quasistatic mechanics to simulate interactions to avoid instabilities caused by dynamic simulators and produce clean signals. Tactile signals generated by BubbleTouch were used for evaluating the methods presented throughout this document.;We then disclose a new tactile data acquisition system for normal force tactile sensors based on compressed sensing. Compressed sensing simultaneously performs data sampling and compression. Given that the signals are sparse in some dictionary, compressed sensing provides guarantees on the accuracy of the recovered signals. We discuss the compressed sensing techniques used for tactile data acquisition to reduce hardware complexity and data transmission, while allowing fast, accurate reconstruction of the full-resolution signal at the processing center. We show that for a simulated test array of 4096 taxels (tactile sensing elements), the system achieves reconstruction quality equivalent to measuring all taxel values independently (the full signal) from just 1024 measurements (the compressed signal) at a rate over 100 Hz.;We note that we found there was no single dictionary that provided the sparsest representation for every signal, and thus, we describe a process that takes the compressed signal and decides which dictionary to use during reconstruction. The method trains the selection process with a set of labeled compressed signals, for which we discuss multiple labeling techniques to account for the availability of different types of signals, e.g., compressed or not. We demonstrate that overall the uncompressed signal labeling methods produce reconstructions almost as good as if it knew which dictionary actually provided the best reconstruction. All of the labeling methods perform near optimal on selecting the dictionary so that the reconstructed signals' accuracies are at least as good as measuring the taxel values individually or, when that is not possible given the available dictionaries, the reconstruction accuracy is as good as possible. We also compare the method to other dictionary selection approaches and show on average the presented approach produces higher quality reconstructions.;The volume of tactile data also makes it challenging for the processing system to analyze and utilize the tactile data for various applications. We discuss the particular processing application of tactile object classification and present an approach that uses the compressed signals as input for the classifier to reduce the amount of data the system processes. Specifically, the approach performs object classification on the compressed tactile data, based on a method called compressed learning. We show the approach obtains over 80% classification accuracy, even with a compression ratio of 16:1. |
