Anthropomorphic Grasping of Complex-Shaped Objects Using Imitation Learning

This paper presents an autonomous grasping approach for complex-shaped objects using anthropomorphic robotic hand. Although human-like hands have a number of distinctive advantages, most the current pickup systems still use relatively simple gripper setups such as two-finger or even suction gripper. The main difficulty utilizing lies in sheer complexity system; it is inherently tough to plan and control motions high degree freedom (DOF) system. data-driven approaches been successfully used motion planning various recently, hard directly apply them high-DOF due acquiring training data. In this paper, we propose novel manipulation system consisting seven-DOF manipulator four-fingered hand with 16 DOFs. Human demonstration data are first acquired virtual reality controller 6D pose tracking individual capacitive finger sensors. Then, 3D shape target object reconstructed from multiple depth images recorded wrist-mounted RGBD camera. estimated residual neural network (ResNet), K-means clustering (KNN), point-set registration algorithm. moves following trajectory created by dynamic movement primitives (DMPs). Finally, robot performs one object-specific learned human demonstration. suggested evaluated official tester five promising results.