Contact-based blind grasping of unknown objects

This paper explores to which extent contact information is enough to perform grasps on unknown objects. The specific scenario addressed consists of emptying a box which contains an undefined number of unknown objects using only force-torque and tactile data. The proposed approach to solve the task, follows the manipulation primitives paradigm, which is based on the use of atomic controllers specifically designed to perform robustly elemental manipulation actions. A set of sensor based primitives that implement a reactive controller that adapts to the uncertain real environment is implemented. Manipulation primitives are assembled within a Finite State Machine, which sequentially searches, grasps and transports all the objects in the box to a predefined location. Two different blind exploration approaches are used to deal with the task. Performance results showing grasp attempts, corrections performed, error corrected and elapsed time are presented and discussed. Moreover a simple vision system is added to compare its results to the blind approaches.