Active haptic exploration with a manipulator for a grasping task

Perception is a basic human ability that provides awareness of the environment and of the person itself in space. Many sensors and strategies have been developed to replicate this feature in autonomous robotic systems, in particular, vision sensors are the ones most used. However, in some scenarios, they are not enough to estimate an unknown object's pose. Therefore this project aims to implement an active haptic exploration based on a force-torque sensor to localize an object for a grasping task. The exploration is accompanied by a particle filter implemented by Johannes Pankert. In order to define the optimal trajectory that the manipulator has to follow, a rough initial visual estimation of the object's position and shape is first generated, based on the point cloud collected from an RGB-D camera. The trajectory is then implemented considering the particles' distribution and only the feedback information from the force-torque sensor; it is an optimal solution since it maximizes the information gain and minimizes the exploration time. This work was tested in a simulation environment for a handle grasping task, which was successfully achieved with a calibration error in the initial visual estimation under 2 cm, with the exception of some corner cases.