Kinesthetic Human-Robot Interaction Based Learning From Demonstration Technology For Humanoid Robot

The robot is a collection of mechanics,electronics,control and artificial intelligence which represents the highest achievement of mechatronics.With the continuous development of science and technology,the humanoid robot has become one of the hottest research topics.Researches on humanoid robot have made a great breakthrough in many aspects,such as the key mechanical components,walking balance and coordinated motion control,speech recognition and interaction,etc.But it is far from our expectation,the main reason lies in the lack of environmental awareness and learning ability of humanoid robots.Compared to the industrial robots with 6 degrees of freedom,which are repeated in a structured environment,the control of humanoid robots is full of challenges.Humanoid robots work together with human,which contains a wide variety of uncertain objects.So,humanoid robots need to complete the complex operation and navigation tasks.To accomplish these tasks,it is necessary to make a good coordination of the sensing device,decision system and limb movement and it is difficult to program in the traditional teaching method.At present,the adaptability and learning ability of humanoid robot is insufficient,the generalization ability is weak.New learning algorithm and control theory still need to be researched and developed constantly.This thesis developed a humanoid robot which uses the advantage of 3D printing.Based on the humanoid robot platform,we made a research of the key technology of learning from Demonstration including the robot kinematics model,multimodal teaching,kinesthetic teaching,visual perception,hand eye calibration,kinesthetic human-robot interaction and enhanced kinesthetic human-robot interaction.First of all,the hardware platform of humanoid robot is established.A two wheeled differential motion platform and a full 3D print humanoid body are combined to develop a humanoid robot imNEU.With the idea of modularization,imNEU is divided into 6 modules including two arms modules,hands module,two head modules and body module.The embedded drive control module is designed with the idea of layered driving and the corresponding software system is designed on the ROS platform.Second,imNEU is equipped with multimodal human-computer interaction interface and teaching means.The control system,communication protocol,communication mode is designed.Under the ROS(robot operating system)software framework,Kinect sensors,binocular vision and other functional modules are developed and integrated.Based on the control system above,a variety of humanoid robot control methods are designed including the Kinect Kinesthetic Demonstration control mode,the GUI contorl mode,the remote control mode with F710,the data glove control mode and the script file control mode.The control experiments were made based on the 5 kinds of control methods.Third,the forward and inverse kinematics of the humanoid robot is established and the motion control problem of humanoid robot is studied.Based on the D-H joint coordinate parameter model,the forward kinematics model of the robot is established and the numerical optimization algorithm is used to solve the inverse kinematics.The simplex method and the L-BFGS algorithm are used respectively.Speed and accuracy of the two algorithms are analyzed and the L-BFGS optimization algorithm is selected to solve the inverse kinematics problem finally.Based on the kinematics,the precision of the robot's repeated positioning and the working range of the robot are determined.Fourth,the visual perception system of humanoid robot imNEU is developed.It includes the left and right camera calibration and the calibration of binocular camera.The objects was identify through the HSV color space.Based on Opencv,we obtain the coordinates of the object.On the basis of this,multiple eye coordinates datas and base coordinate datas were obtained,then the least squares method were use to obtain the optimal transform matrix H*.Fifth,the study of kinesthetic human-robot interaction and enhanced kinesthetic human-robot interaction is carried out.Due to the high dimension and uncertain characteristics of the teaching datas,the GMM(Gaussian Mixture Models)and GMR(Gaussian Mixture Regression)methodes are put forward.The GMM is used to model the raw training datas and the GMR is a regression algorithm which is used to get the robot trajectory.In view of the internal coupling problem of environmental constraints and teaching data under unstructured environment,the teaching data reconstruction under multi coordinate system and multiple GMM hybrid methodes are proposed to improve the generalization performance of algorithm under unstructured environment.Finally,based on the imNEU humanoid robot experiment platform,taking the assembly process as an example,the experiment research of kinesthetic Human-Robot interaction and enhanced kinesthetic Human-Robot interaction are carried out.The experimental results show that the proposed method is reasonable and effective and it can solve the problem of human-robot interaction in unstructured environment of humanoid robot.