Learn to Grasp Objects with Dexterous Robot Manipulator from Human Demonstration

摘要

This study proposed a method to transfer human grasping skills to dexterous robot manipulator. Dynamic motion primitives (DMPs) were used to model the trajectory for the robot manipulator. Users demonstrated a motion task by moving the robot manipulator and thus realizing 3D trajectory. Each joint of the robot manipulator set their own independent DMPs model. This was to enhance the stability of the generated motion towards the target. Furthermore, the generalization of DMPs that the trained model can be applied to similar trajectories, was found and verified in the experiment. A robotic hand (ZealGiant hand) was installed at a 7-DOF robot arm (Kuka iiwa 14), with force sensitive resistance units at the end of the five fingers. A simple data glove was designed to record the human grip posture. A synergistic control was developed for grasping according to the real-time force. Learning from human demonstration for the robotic manipulator developed in study may facilitate programming and mimic a human-like grasping control.