Research And Implementation Of Robot Arm Task Imitation Based On Learning From Demonstration

Making robots behave similar to human intelligence is a hot topic in the field of robotics,which requires robots to have the ability to learn.Robot arm is the most widely used mechanical device in robotics field.The research of its behavior acquisition is an important aspect of robot motion skill research.Learning from demonstration(LfD)is a fast and efficient way of learning.Through learning from demonstration,robot arm can quickly acquire motion skills,thus simplifying the complex motion planning and improving the learning efficiency.This paper focus on the research of robot arm's behavior acquisition through LfD.Main contents are as follows:Firstly,the construction of robot arm LfD system and kinematics analysis.Learning system of robot arm is set up based on LfD framework,and a hands-on demonstration system with LfD mechanism is designed.According to the requirements of this paper,robot arm is selected and kinematics analysis of robot arm is completed.In the physical environment,robot arm is dragged to conduct demonstration,and the joint angle information of robot arm in the process of movement is obtained.Through the positive kinematics calculation,the motion information of robot arm in the demonstration process is obtained,and the data for robot arm LfD is provided.Secondly,the design of state-action strategy based on RNN and the imitation of point to point task.In view of the difficulty in obtaining the behavior of robot arm,RNN is adopted to expression strategy,and RNN network is constructed according to the demonstration information,and strategy is obtained through training samples.In the physical environment,the imitation of the original task and the complex task are completed by robot arm,and imitation of generalization task is carried out on the basis of the imitation of demonstration task.The experimental study shows that robot arm can not only imitate demonstration task,but also can realize imitation of new task by generalization.Thirdly,the design of state-state strategy and imitation of target task based on LSTM.Aiming at the problem of target task with different initial points but the same target,the strategy of learning state to state through LSTM network is designed,and the trajectory is planned through iterative output by strategy,and the action information is obtained by the inverse kinematics calculation.In the physical environment,the research of the target task imitation under single demonstration and multiple demonstration is carried out.Imitation of target task by robot arm is realized through the physical experiment.This paper studies the acquisition of behavior based on teaching LfD,and the research results have positive significance for the acquisition of motor skills through learning from demonstration.