Robot Skill Learning Based On Episodic Memory And Meta Learning

From birth to old age,human beings can always learn and accumulate knowledge,and can store and memorize the learned skills,so as to realize the effective reuse of knowledge to quickly solve new tasks.This paper studies the learning method of robot skills,so that it can master the skills needed to complete the task in the learning process,remember the multiple segments of the training task,and can be integrated and applied to new tasks.In this paper,the acquisition of robot arm motor skills based on episodic memory and meta learning is studied in chapters:First,we design the object recognition and grasping system to realize the recognition and collision-free operation of objects in the working space through vision.After the camera calibration,the point cloud information of the object is obtained to realize the positioning of the object;The vision system based on RGB-D image is built to realize the object recognition in simple environment and complex environment respectively.Based on Move It! motion planning library carries out the path planning of the robotic arm,and designs the basic actions commonly used in the task.Then,we carry out the method research of robot learning skills based on meta-learning.In this paper adopt meta-learning network built with LSTM as the core.The manipulator conducts cross task learning through the training on multiple tasks.In the model training stage,the meta-learner sets up multiple sets of tasks for the skill acquisition goals.The basic learner learns the current task for each task,and masters the current task.The meta-learner built by LSTM receives the gradient information transmitted by the meta-classifier.After the meta-learning network is trained,the network parameters are stored as experience modules through the established contextual memory model to guide the meta-classifier in learning new tasks.Third,by studying the biological basis and cognitive characteristics of episodic memory,after we analyze the episodic memory function,and construct episodic memory including robot experience and skill models.The episodic memory of each task is modeled as a collection of multiple events and the network weight of the meta-learner for the task.At the low level,the robot's state perception and behavior actions of the event are encapsulated for each event of the task.At the high level,the network weight of the task meta-learner is encapsulated for each task and stored as contextual experience.Finally,this paper builds a robot experimental platform composed of Kinect V2 camera,ur3 e manipulator,mobile operation platform and robotiq gripper,designs tasks such as cleaning mixed objects on the desktop and stacking blocks,and learns simple skills such as grasping and stacking by teaching the robot to interact with objects;Then,the generalization ability of the model on complex tasks is verified by block stacking task,configuration change task and mixed item classification task.