Learning From Demonstration For Humanoid Robot ARM Based On Kinect

The emergence of Learning from Demonstration(LfD)technology makes the robot behavior strategy shift from manual design to machine learning algorithm approximation,and thus greatly enhance the learning efficiency and effectiveness.As the input of environmental information,visual signal is intuitive,efficient and easy to interact with human.It is an ideal source of information for robot environment perception.Kinect camera has many advantages,such as low price,low cost,easy to use and rich data.Therefore,it is of great significance to apply Kinec t to robot demonstration learning.Therefore,in this paper,Lf D for humanoid robot arm based on Kinect is carried out.The research contents include the following aspects:In this paper,three kinds of learning tasks in human-robot interaction are studied,including the trajectory tracking of the human arm joint,the tracking of human arm end effector posture and the autonomous painting.This article mainly completes the following work:Firstly,the study of human-robot joint motion based on elbow constraint is carried out.In this paper,the 7-DOF joint mapping between human arm and humanoid robot arm is established,and a redundant 7-DOF inverse kinematics analysis method based on elbow constraint is proposed and deduced.The trembling problem of r obot arm is analyzed and the problem was solved by applying the shuffle filter method.Finally,the real-time joint motion control of the manipulator based on human-robot joint mapping was realized.Secondly,the estimation of hand posture based on convolution neural network has been studied.Aiming at the problem of hand image acquisition,this paper proposes a hand image acquisition method based on arm posture normalization,and completes the definition of hand TIM gesture and the solution of parameters.In this paper,a multilayer convolutional neural network is constructed based on Caffe platform in order to establish the mapping relation between depth impage of human hand and posture parameter,to present two methods of hand gesture acquisition,and to realize the real-time posture estimation of human hand.Thirdly,the research on autonomous painting of humanoid robot arm based on SAE-LSTM has been carried out.After analyzing the painting process,this paper defines the painting angle sequence.In this paper,a neural network training data set was made with geometric figure,and the SAE-LSTM neural network model has been constructed.The feature extraction of the geometric image is applied by SAE and the drawing angle sequence is stored by LSTM.And then the homemade training set is used to complete the training and testing of the combined network.Finally,the feasibility of the scheme is verified by simulation experiment.Finally,in this paper,several corresponding experimental platforms are set up for different demonstration learning tasks.Joint experiment is designed to verify the correctness of the inverse kinematics solution,and the real-time performance of the scheme is analyzed in the experiment.The hand gesture estimation experiment is designed to determine the accuracy of the posture estimation method using NDI data as a standard value to evaluate the method proposed in this paper.The experiment of manipulating the geometric figure of the manipulator is designed,and the precision of the geometric figure is improved with the advancement of the training gradually.