Research On Key Technologies Of Lifelong Imitation Learning For Robot Motion Control

Imitation learning is one of the most common learning methods in robot motion control.The traditional robot imitation learning algorithms learn control models one by one for specific robot operational tasks,which results in poor robustness and weak portability.The incremental imitation learning of robot proposed in this paper is similar to the process of human lifelong learning based on experience.It has the ability to transfer the learned knowledge online,and this ability is the fundamental of robot's large-scale operation tasks through small sample data learning.In the future,this is an important development direction for imitation learning.This paper mainly studies the incremental imitation learning for the manipulator.It is the basis of the motion control and behavior decision of the robot manipulator.The imitation learning method proposed in this paper is divided into imitation learning based on behavioral cloning and imitation learning based on inverse reinforcement learning.Although some existing methods reflect the characteristics of imitation learning which are easy to learn robot motion control strategies,it is difficult to fully solve the problem of robot imitation learning when the number of learning tasks is large and the working environments are dynamic and complex.Therefore,this paper do the research on the incremental multi-task imitation learning problem based on lifelong learning framework,including incremental multi-task imitation learning based on behavioral clone methods such as DMP and incremental multi-task imitation learning based on inverse reinforcement learning.They are applied to the key issues in the field of robot imitation learning control,such as end-effector trajectory imitation learning and behavioral action imitation learning.The main contents of this article are as follows:Firstly,the lifelong imitation learning problem for robot operation control is proposed in this paper,which is discussed in detail from the necessity and rationality.A lifelong imitation learning framework is established,which enables the robot to acquire relevant empirical knowledge from the previously learned motion control strategies,and continuously expand and update the previously learned knowledge base to realize online multi-task imitation learning.Secondly,the behavior cloning based imitation learning method DMP and the inverse reinforcement learning based imitation learning method are taken as examples to study the application of the lifelong learning framework in the robot incremental multi-task imitation learning problem.Aiming at the robot imitation learning based on behavioral cloning,an incremental trajectory imitation learning framework based on sparse coding and DMP is proposed in this paper.For the robot imitation learning based on inverse reinforcement learning,an incremental motion control algorithm combined with sparse coding for multi-intention reinforcement learning is proposed.Finally,in order to verify the effect of the proposed methods,we tried the multitask imitation learning experiment based on three simulation platforms and the physical UR5 robot arm experimental platform.The ability of robot autonomous online learning,optimization and construction of multi-task imitation learning knowledge base were studied.These methods build an experience knowledge base which can performs a large number of multi-task robot imitation learning.