Dexterous Grasping For Three-Finger Robotic Hand In Multi-Object Environments

Object grasping is a fundamental function of robotics,especially for robots with multi-finger dexterous hands.With the growing demand for robotics in various applications,this dissertation aims to help robots with multi-finger dexterous hands perform collision-free and high-quality grasping in complex scenarios involving multiple objects.The main contributions include:· Novel Grasp Representation:We propose a novel grasp representation that improves upon traditional methods.Our approach takes into account the fact that a dexterous hand will make contact with an object during the grasping process,allowing us to limit the sampling range in space based on these contact points.This also results in a significant reduction in sampling the grasp space and leads to faster and more effective generation of multi-finger dexterous grasps.· Multi-finger Dexterous Hand Grasping Dataset: We design two distinct procedures for constructing grasp datasets for a multi-finger dexterous hand in a simulation environment using Grasp It! and Py Bullet.One procedure is tailored for generating a multi-finger dexterous grasp dataset for single objects,while the other is focused on constructing a grasp dataset for multi-object scenes.By implementing these procedures,we are able to create two separate and specialized datasets that provided diverse data for training our neural network model.· Dexterous Hand Grasping Algorithm:We develop CMG-Net,a novel grasp network model that leverages a new grasp representation.By inputting point cloud data from a single viewpoint,CMG-Net can accurately predict the grasp position and robot’s joint angles for a given scene.This enables the dexterous hand to perform high-quality grasping in even challenging multi-object scenarios.· Dexterous Hand Grasping System:We develop a grasping system for multi-finger dexterous hands,operating both in simulation environments and in real-world environments.In terms of hardware,the system consists of a Franka Emika Panda robotic arm,a Realsense camera and a three-finger gripper,all of which are integrated in ROS.