Daily objects Dataset
Description
For the experiments, we select a set of five different ‘daily life’ objects shownin Figure 2b. These objects have three types of surfaces: flat (e.g. one side ofthe eyeglasses case and the cubic box), circular (e.g. the cola can and the tennisball), and curved (e.g. the shampoo bottle and the other side of the eyeglassescase) surfaces. The objects are different in shape, materials, and size. But, theyall have a good impact resistance (i.e. high stiffness) and are big enough to bein contact with many sensors when grasped. Data was collected by grasping each of the five objects 400 times, at differentpositions, using the Baxter gripper as shown in Figure 2c. Our experiments donot involve any post–grasp manipulation; therefore, the arm position is fixed.Before the experiments, we determine the closure position of the gripper foreach object, which corresponds to the gripper position when the touch occurs.We set the gripper velocity and force to v = 5cm/s and F = 0.03 ∗ F max(F max = 35N ). The chosen objects are hard enough to avoid any deformation 8 Youssef Amin et al.when grasped. During each trial, the gripper closes –at a constant velocity andforce– to the predefined position, and remains closed for 1.2 seconds; then, itopens for two seconds. The trials are repeated with no sensors and/or gripperfeedback (open–loop control). Grasping control is done by the robotic operatingsystem (ROS). The IE continuously acquires the tactile data from all sensorssimultaneously at a high sampling frequency (2KSps) and transmits them to thePC. According to Section 3.2, the signals are pre–processed before extractingthe features. Using Matlab, we automatically extract 150 samples from the grasppeak by taking 25 samples before the sample that has the minimum value and125 samples after. This process is done for all the channels simultaneously, andfor all the trials.Eventually, we build six datasets based on the combination of the numberof sensors (i.e., excluding four sensors or using all of them) and the extractedfeatures. Each dataset contains 2000 data, i.e. 400 samples for each class. Table3 gives the details of the six datasets. Each column corresponds to a dataset,the first row reports the name of the datasets, while the second, third, fourth,and the fifth rows report the number of sensors used to build the datasets, thenumber of features extracted from each sensor, the type of feature (i.e. Mean,standard deviation=STD, kurtosis=Kur, and skewness=Skew), and the totalnumber of features, respectively. The six datasets correspond to six differentfeature extraction stages. The loss function will be evaluated to find the bestcombination of features extraction stage and predictor.