Research On Human-Robot Skill Transfer System And Related Methods

In recent years,with a new round of global industrial revolution and the wave of scientific and technological development,the robot industry at home and abroad has developed rapidly,and its scale and market are constantly expanding.By improving the cognitive and learning ability of the robot,the robot can plan the motion path independently according to the task environment under the condition of non-customized programming.However,the traditional industrial robot mainly uses the programming-by-teaching method to realize the motion planning.This method only makes the robot repeat the teaching trajectories,which has many problems such as low programming efficiency and poor generalization ability.The human-robot skill transfer is a key technology to realize the efficient programming of robot.It transfers the human skills to robot after the universal representation,which improves the efficiency of robot programming and the ability of skill generalization in different task scenarios.In this paper,the human-robot skill transfer system and related methods will be studied.By improving the existing skill model and combining with robot intelligent control algorithm,the human-robot skill transfer system based on skill modeling has been developed.Firstly,this paper discusses and analyzes the classical methods of robot skill model,including dynamic motion primitive(DMP)and Gaussian mixture regression(GMR).Meanwhile,we design a skill transfer framework based on DMP model.The framework consists of robot skill modeling and trajectory tracking controller.In the part of robot skill modeling,we propose an improved skill modeling method based on GMR and DMP,which makes the skill model not only have the ability of space expansion,but also can encode the spatial probability distribution of the same set of demonstrations.Furthermore,we use fuzzy Gaussian mixture regression(FGMR)instead of GMR to improve its fitting ability to DMP nonlinear terms.In the part of trajectory tracking controller,we design a controller based on function approximation,which uses radial basis function(RBF)neural network to estimate the unknown dynamic part of the manipulator,compensates the dynamic disturbance caused by the external environment,and reduces the trajectory tracking error of the controller.Considering that the skill modeling based on multi-source information plays an important role in the successful execution of robot tasks,this paper also designs a human-machine skill transfer system based on force and position modeling,which will simultaneously model the contact force information and position information contained in the demonstrations.For the teaching stage,we realized the traction teaching function of the end of the mechanical arm based on the mobility model.Considering the lack of force sensors or the number of force sensors,we design a force observer based on momentum to estimate the contact force between the end of the manipulator and the environment.For the skill learning stage,we further improve the dynamic system part of the DMP model,using the first-order exponential decay system to replace the fixed target point parameters in the model,to solve the problem of abrupt acceleration of the model output trajectory.For the recurrence stage,we use active stiffness control to control the stiffness between the end of the manipulator and the environment,so as to indirectly control the contact force.