Deep learning techniques for modelling human manipulation and its translation for autonomous robotic grasping with soft end-effectors

One of the key enablers for the extraordinary dexterity of human hands is their compliance and capability to purposefully adapt with the environment and to multiply their manipulation possibilities. This observation has also produced a significant paradigm shift for the design of robotic hands, leading to the avenue of soft endeffectors that embed elastic and deformable elements directly in their mechanical architecture. This shift has also determined a perspective change for the control and planning of the grasping phases, with respect to (w.r.t.) the classical approach used with rigid grippers. Indeed, instead of targeting an accurate analysis of the contact points on the object, an approximated estimation of the relative hand-object pose is sufficient to generate successful grasps, exploiting the intrinsic adaptability of the robotic systems to overcome local uncertainties. This chapter reports on deep learning (DL) techniques used to model human manipulation and to successfully translate these modelling outcomes for enabling soft artificial hands to autonomous grasp objects with the environment. Chapter Contents: • 1.1 Introduction • 1.2 Investigation of the human example • 1.2.1 Methods • 1.2.2 Experiments • 1.2.2.1 Evaluation on ECE data set • 1.3 Autonomous grasping with anthropomorphic soft hands • 1.3.1 High level: deep classifier • 1.3.1.1 Object detection • 1.3.1.2 Primitive classification • 1.3.2 Transferring grasping primitives to robots • 1.3.3 Experimental setup • 1.3.3.1 Approach phase • 1.3.3.2 Grasp phase • 1.3.3.3 Control strategy • 1.3.4 Results • 1.4 Discussion and conclusions • Acknowledgement • References.