Movement Primitives Based Robot Learning By Demonstration And Its Applications

For autonomous operation tasks and human-robot cooperation tasks of service robots,learning by demonstration is a strategy to learn operation skills by imitating the data presented by a given expert,which can make robots have the ability to reproduce tasks in unstructured environments.This paper focuses on the learning and generalization of dynamic movement primitive and interactive probabilistic movement primitive of trajectories.It deeply studies robot learning by demonstration and system design technology based on the hierarchical decomposition of tasks.In addition,a hierarchical learning by demonstration system is built based on Kinect RGB-D sensors and a UR5 robotic arm to verify ability of trajectory generalization and task reproduction for typical autonomous operation and human-robot cooperation tasks.Autonomous behavior reproduction mode and interactive cooperation reproduction mode are designed according to Lb D autonomous operation tasks and human-robot cooperation tasks respectively.Typical task sets are decomposed into action primitive sets,and action primitive sets are divided into for autonomous action primitives and for interactive action primitives according to whether robots interact with humans.Firstly,original teaching trajectories are collected for action primitives by using the direct teaching or indirect teaching,and trajectory noise is eliminated by Gaussian filtering,Gaussian mixture model and Gaussian mixture regression to obtain the optimal reference trajectories for robot imitation learning.Then,dynamic movement primitive can be used to learn a single-case reference trajectory to obtain trajectory generalization ability for autonomous action primitives.While for autonomous action primitives affected by task parameters or environmental factors,dynamic movement primitive parameterized model can be used to learn multi-demonstration action trajectories to enhance the generalization ability with respect to task parameters or environmental factors.For interactive action primitives,in order to allow robots to make real-time response to observed human partial motion trajectories,interactive probabilistic movement primitive is used to learn multi-demonstration human-robot interactive trajectories and predict the post-action humanrobot interaction trajectories by combining with Bayesian model-based action recognition.In addition,three phase estimation schemes are extended: phase estimation based on Gaussian distribution,phase estimation based on uniform distribution and phase estimation based on minimum distance.Environmental adaptive interaction probabilistic movement primitive is designed for interactive actions affected by the environment.The meta-model learns multidemonstration trajectories related to the environment to enhance its generalization ability with respect to the environment,and uniaxial equidistant interpolation phase estimation strategy is proposed.Finally,two types of action primitives perform simulation experiments to verify the generalization ability of the model.A hierarchical learning by demonstration system is built based on the above research content.Several functional modules such as trajectory acquisition and preprocessing,action primitive learning and generalization,and actual task reproduction are designed,and corresponding functional software and application interfaces are developed as well.For the multiple tasks under two modes,reproduction experiments on the UR5 robot are carried out in actual scenarios.The feasibility of imitation learning technology for action primitives and the practicality of application software development in this paper are verified through analysis of a series of task execution results.