A Research On Robotic-arm Learning Method By Demonstration Based On Trajectory Imitation Learning And Task Planning

Robot learning by demonstration has mitigated the programming pressure of engineers to a certain extent and has a broad application prospect in the domain of service robot and industrial robot.The excellent learning method should enable the robot to have the flexible task representing and planning ability to cope with the mutative working environment and the needs of different users.At present,the common research routes of robot-arm learning by demonstration are admittedly divided into two derections: Trajectory-based learning method and task-based learning method.Trajectory-based demonstration learning applies machine learning methods to teaching robot trajectories,which has the advantages of being intuitive and easy to reproduce,but for more complex and changeable tasks,it will increase the complexity of learning to some extent,and its scene generalization ability is poor.However,learning based on task planning usually uses machine vision technology to represent task-related information,so as to carry out reappearance planning.Its advantage is that the scene generalization ability is excellent.However in the reproduction phase of the task,execution trajectory is mostly predefined generated,relatively inflexible.Based on the comprehensive analysis of the advantages and disadvantages of the two above-mentioned methods,this paper combines the two methods to proposes a robot learning framework based on machine vision in order to achieve better effect,and.First,compared with the traditional artificial drag robot trajectory model learning in the form of teaching,this article adopted kinectv2 camera to perceive the movement of the joints human hand,and based on DTW(Dynamic Time Warping),GMM(gaussian mixture model)/GMR(gaussian mixture regression),DMP(dynamic movement primitives),proposed an unmarked hands demonstration methods to learn trajectory of the movement primitive,and designed a trajectory imitation learning experiment,verify the feasibility of the method,to investigate the generalization ability of this method.Then,a task representation and reproduction planning method based on image processing and image analysis is designed.This method extracts and processes key frames by combining teaching context to represent task-related configuration information.According to the represented task model,the position of the target object and final state of corresponding in the new scene for repetition planning,and combine with the trajectory imitation learning of the action primitive to control the recurrence of the robotic arm to perform the task.Finally,the hand-eye experiment platform is set,and suitable hand-eye calibration method is proposed in view of the experimental platform configuration.Based on this,the demonstration data is mapped to the robot-arm coordinate system.Simple peg-in-hole experiment and assembly experiment are designed,in order to verifies the ability of reproduction by the overall learning framework.a large number of experiments proved the feasibility of the learning task and reproduction planning,as well as the reliability of the overall learning framework of in the phase of task reproduction.