Research On Five-Finger Grasp Planning Based On Machine Learning

With the continuous development of space exploration in various contries,spacecraft of various tasks have been sent into space.The increasing intensity of space experiment tasks and on-orbit services have become a severe test to astronauts.Due to the physical constraints and great risks astronauts face when carrying out space activities,autonomous and unmanned technique is an important development direction of space experiments and on-orbit services.Therefore,the autonomous manipulation of space robot gradually becomes one of the important ways to complete the space experiments and on-orbit services.Equipped with a variety of sensors,controllers and actuators,the robot can independently manipulate specific objects without external manual control.Autonomous grasping is a major challenge in the research of robots' autonomous manipulation.This paper combines Machine Vision with Machine Learning to solve the problem of grasping novel objects with the five-finger dexterous hand,especially the perception of grasping objects,the selection of five-finger grasping points and the strategy of how to execute the generated grasps.The main work is as follows:A five-finger grasping point selection method based on SVM-ranking is designed.The five-finger grasping point of the unknown object can be planned based on vision sensors.The planner obtains colors image and depth images through RGB-D sensor to generate point cloud.Based on the sensor data and the human hand grasping analysis,a three-step optimal selection strategy is proposed;on this basis,the corresponding search coordinate system is established for each step of selection,and the features of the corresponding selection object are extracted;the training set is obtained by manual marking,and the selection model of each step is learned by the SVM-ranking learning algorithm;tests have been conducted to verify the effectiveness of the proposed algorithm.A fast selection algorithm of optimal grasp region based on deep learning is proposed.A sparse auto-encoder is designed to learn features autonomously.For RGB-D multi-modal input,a learning algorithm with structural penalties is adopted to regularize the number of connections between the input layer and first hidden layer so that the over-fitting is restrained.A two-step learning strategy is proposed to learn the optimal grasp region.Based on the training set,the deep neural network model is trained and evaluated by the cross-validation.A five-finger dexterous hand grasp control strategy is designed.Firstly,the structure of the five-finger dexterous hand is analyzed and the simplified model of the hand is established.Then,the kinematic model and the dynamic model of the hand are derived.Based on the common robot arm-hand system,the strategy of how to execute the generated grasp and how to control the arm and the hand respectively is design.The simulation of autonomous grasping is conducted.Based on ROS,an autonomous grasp simulation platform was established.The effectiveness of the five-finger dexterous hand grasp planning system was verified by the autonomous grasping simulation experiment.In summary,for the problem of five-finger dexterity grasping,the grasping plan method based on RGB-D data input and machine learning theory has been proposed.This dissertation can provide meaningful theoretical and technological guide for the practical application of five-finger dexterous hand and the further study of autonomous manipulation of space robot.