A virtual data generator system for shape recognition in haptic robotics

Abstract

In robotics, the current state of object recognition in haptic sensory mode falls significantly short of the results obtained in visual mode. One of the main reasons for this is the lack of haptic data sets for training recognition models. A major impediment is the time-consuming and difficult task for a real robot to capture large amounts of haptic information. This paper introduces a virtual haptic dataset generator system that captures haptic features based on the curvatures of an object. The main goal is to show that this capture system is a feasible approach that can eventually be implemented not only in virtual settings but in actual robots. The virtual haptic capture system described speeds up the learning process, where a real robot would learn through virtual simulation. The paper shows three important points that make the system feasible. The capture is independent of the angle of inclination of the end-effector as it approaches the explored object. The system recognition is performed on everyday objects. Since a real system is exposed to noise during data acquisition, the data of the virtual system must also contain noise. High performance is still achieved within the noise ranges of current sensor systems.