Research Of Motion Planning And Compliant Control For Robot Manipulators Based On Imitation Learning

Robots have been widely used in human's daily life,which can liberate human from monotonous and repetitive work,and also improve production efficiency and quality.However,the capabilities for dexterous manipulation and intelligent decision-making are still far lower than humans.Robots can only be used for repetitive tasks in a structured environment.It is non-trivial to design a motion planning algorithm with high adaptability and stability for robots in a dynamic unstructured environment.For tasks in a complex contact-based environment,the planning and control of the force and position profiles still need to be designed carefully for specific tasks,and it is difficult to extend application to different scenarios.Therefore,this paper proposes a motion planning and compliant control method for robot manipulators based on imitation learning to improve the ease of use and intelligence of the robot,and to make robots accomplish tasks like a human in complex scenarios.In order to improve the ease of use and adaptability of the robot to the environment,a promising way is to use a dynamical system for motion planning.However,traditional dynamical system methods often use time-based planning or are complicated to design.This paper proposes a dynamical system design method based on a diffeomorphism,which can realize single-step training.Based on the constraints of the dynamical system,the constraints of the diffeomorphic mapping are analyzed.Then a unified framework of the diffeomorphic mapping that couples the position and orientation is proposed.Based on this framework,two mapping algorithms are proposed: an iterative algorithm based on radial basis functions,and a mapping algorithm based on invertible neural networks.Subsequently,comparison and analysis were carried out in simulation and experiment to verify the effectiveness of the mapping algorithm framework.Secondly,current time-invariant dynamical systems only consider the position planning,and the orientation planning still adopts time-domain interpolation,which cannot realize the synchronization between positon and orientation planning.This paper proposes a dynamical system design method that integrates position and orientation.This method establishes a diffeomorphic mapping between the teaching space and the latent space,and constructs a globally asymptotically stable system in the latent space,thereby obtaining a globally asymptotically stable system that can accurately represent the demonstration data.The stability of the dynamical system is proved by Lyapunov function.It is verified by two-dimensional space simulation and real robot experiments in three-dimensional space.The results show that the online motion planning method has good adaptability to time and space disturbances and can replan the motion on-the-fly in dynamicalally changing environment,which proves the effectiveness of the algorithm.Finally,for contact-based tasks,it is necessary to plan force and position profiles and design a compliant controller based on force and position feedback.This paper proposes a skill representation model for contact-based task manipulation.By collecting demonstration data in the process of human performing contact-based tasks,the relationship between position,speed,interaction force and interaction stiffness in contact tasks is constructed,which is used for the planning and control of the robot manipulator during autonomous execution.Then a skill learning framework for compliant manipulation is proposed based on imitation learning for contact-based tasks,and an adaptive hybrid force/position controller and an adaptive impedance controller are designed based on system state for typical contact tasks.In order to verify the framework,three types of robot experiments are designed: robot polishing experiments,peg-in-hole assembly experiments,and variable stiffness interactive experiments.In these experiments,the robot manipulators can realize compliant manipulation like a human to accomplish contact-based tasks,which verifies the effectiveness of the skill learning framework for compliant manipulation.This paper systematically studies the motion planning and compliant control methods of robot manipulators,and proposes a motion planning and control algorithm based on imitation learning,which can quickly adapt to different operating scenarios,and also adapt to dynamic environment and unknown disturbances.The researches in this paper are of great value and significance to the research on motion planning and compliant control,and also the application of robots.