Rotation and translation invariant object recognition with a tactile sensor

In this paper a novel approach is proposed to recognise different objects invariant to their translation and rotation by utilising a tactile sensor attached to a robotic arm. As the sensor is small compared to the tested objects, the robot needs to access those objects multiple times at different positions and is prone to move or rotate them. This inevitably increases difficulty in object recognition during manipulations. To solve this problem, it is proposed to extract tactile translation and rotation invariant local features to represent objects; a dictionary of k words is therefore learned by κ-means unsupervised learning and a histogram codebook is then used to identify objects. The proposed system has been validated by classifying real objects with data from an off-the-shelf tactile sensor. The average overall accuracy of 91.2% has been achieved with only 10 touches and a dictionary size of 50 clusters.