Grasp It like a Pro: Grasp of Unknown Objects with Robotic Hands Based on Skilled Human Expertise

This work proposes a method to grasp unknown objects with robotic hands based on demonstrations by a skilled human operator. Not only are humans efficacious at grasping with their own hands but are also capable of grasping objects using robotic hands. Therefore, we consider how the grasping skills of a human trained in robotic hand use can be transferred to a robot equipped with the same hand. We propose that a skilled human user manually operates the robotic hand to grasp a number of elementary objects, consisting of different boxes. A Decision Tree Regressor is trained on the data acquired from the human operator to generate hand poses able to grasp a general box. This is extended to grasp objects of unknown and general shape leveraging upon the state-of-the-art Minimum Volume Bounding Box decomposition algorithm that approximates with a number of boxes the point cloud of the object. We report on extensive tests of the proposed approach on a Panda manipulator equipped with a Pisa/IIT SoftHand, achieving a success rate of 86.7% over 105 grasps of 21 different objects.