Material Recognition with a Tactile Skin on the Fingertip of a Humanoid Robot by Learning Temporal Features

Kurzfassung

Robot manipulation reaches a new level with the help of tactile sensing. Many robotic tactile sensors have been developed in the past decades. However, due to the different advantages and disadvantages of the existing robotic tactile sensors, none of them are perfectly applicable for a wide range of humanoid robot applications. In the first part of this thesis, a commercial tactile sensor fromTekscan, which was originally designed for human hands, is tested with the humanoid robot Agile Justin from DLR. Some potential robotic applications are outlined, such as slippage detection and surface exploration. The second part of the thesis focuses on the application of texture discrimination. Two feature vectors, L-Feature and H-Feature, are proposed to represent the properties of different textures. Several environmental variations are identified, which negatively affect the overall performance. Gaussian, kNN and SVM classifiers are applied to these two feature vectors and the performance is evaluated with random and leave-one-position-out cross validation respectively. The classifier trained from the H-Feature with SVM obtains the best performance with an accuracy of 0.96 without and 0.86 with environmental variations.