Research On Shared Control Method Of Multi-Degree-of-Freedom Manipulator Based On Learning From Demonstration In Unstructured Environments

Robotic manipulators have played irreplaceable roles in implementing rapid,accurate,and repetitive tasks in structural environments such as industrial applications.However,lots of tasks in unstructured environments are still executed by low-level automated machines and skilled operators.Due to environmental changes,operation fatigue,and other factors,it is easy to lead to safety accidents or operational efficiency decline.When both manual operation and automation have obvious defects,the humanmachine shared control integrating manual operation and automation has the advantages of human creativity,adaptability,and cooperation,as well as the advantages of automatic operation being fast,reliable,precise,and good at repetitive tasks,which can reduce the operator’s burden and improve work efficiency.To help the operator complete the task,the robot needs to first predict the target,and then provide assistance according to the prediction result.To make predictions while performing tasks,it is usually necessary to establish the motion model of the robot.However,since the task type and scene of the manipulator may change frequently,it is necessary to efficiently and quickly establish the motion model of the robot for a specific task.Learning from demonstration(Lf D)in machine learning technology can realize the programming of robots by simply demonstrating how to perform tasks,which is efficient,intuitive,easy to use,and flexible.However,due to the problems of dissimilar kinematics,operation noise,and trajectory inconsistency with large spatial-temporal variations,it is difficult to obtain high-quality demonstrated trajectories to reflect the operator’s true intentions through teleoperation-based Lf D for heavy-duty manipulators.The statistical modeling method can alleviate the problem caused by operation noise,but it requires a large number of demonstrated samples,so how to use a small number of imperfect demonstrated samples for trajectory learning of the hydraulic manipulator is a problem to be solved in this paper.Secondly,Lf D can not only generate motion trajectories,but the Lf D based shared controller can learn the demonstrated samples,predict the operator’s intention,and then help the operator to complete the task.The advantage of this framework is that there is no need to pre-set targets/paths or detect complex environments.The shared controller is divided into blended shared control and haptic shared control.How to design these two types of shared controllers based on Lf D to assist the operator in completing tasks is another problem to be solved in this paper.For the above problems,this paper aims to meet the requirements of auxiliary control of multi-degree-of-freedom manipulators in unstructured environments.Based on Lf D,research on blended shared control and haptic shared control is carried out to realize the human-robot cooperative operation of the manipulator in unstructured environments and take into account the safety of interaction.The difficulty of trajectory learning of heavyduty manipulators in unstructured environments and how to realize shared control through Lf D are solved.The research on Lf D based shared control method of the manipulator in unstructured environments is systematically carried out.This paper is divided into six chapters,which are described in detail as follows:In chapter 1,the background and research value of Lf D based shared control for multi-degree-of-freedom manipulator in unstructured environments are introduced,and the existing problems and technical challenges are pointed out.The current research status of Lf D and shared control is expounded and their shortcomings are pointed out.The significance and main contents of this research are summarized.In chapter 2,the teleoperation-based Lf D for the master-slave hydraulic manipulator system has some problems,such as size scaling,response delay,and oscillation tendency.As a result,it is difficult to obtain the ideal demonstrated sample,and thus it is difficult to generate a smooth trajectory that can accurately reflect the real intention of the operator.To overcome this problem,a locally weighted task-based trajectory learning is proposed to deal with the unsatisfactory demonstrated samples caused by velocity oscillations and operational noise.Firstly,A locally weighted chattering cancellation algorithm is proposed to reduce the influence of operational noise in the demonstrating process.Secondly,a sequentially hierarchical Dirichlet process with temporal encoding algorithm is proposed to extract the subgoals of demonstrated trajectory.The algorithm can select cluster penalty parameters adaptively to filter the operating noise,so as to reflect the real intention of the operator.Then,NURBS is used to fit subgoals to reconstruct the trajectory,and DMP is used to generalize the reconstructed trajectory.Finally,experimental tests are carried out in two scenarios to prove the effectiveness of the proposed method.In chapter 3,aiming at the problem of inaccurate intention prediction due to poor dynamic performance of heavy-duty hydraulic manipulators,such as oscillation tendency and response delay,a blended shared control method based on locally weighted intent prediction is proposed.First,the task learning is implemented on the basis of the research in the second chapter.Second,the robustness to operational noise is improved by calculating two possibilities: empirical probability and real-time probability.Then,Arbitration rules based on the predicted results are established to seamlessly blend the control signals of the operator and the robot.Finally,comparative tests in two typical scenarios are carried out on the self-developed teleoperated hydraulic manipulator system.The advantages of the proposed shared controller are proved from the aspects of trajectory tracking error,operation efficiency,collision times,and operator’s user experience.In chapter 4,virtual fixture is a powerful tool to improve safety and efficiency for teleoperation or collaborative tasks.However,traditional virtual fixtures with constant stiffness cannot handle such scenarios in which robots need to leave the constraints to perform tasks.Therefore,an adaptive virtual fixture based on the motion refinement tube is proposed,which can dynamically adjust the guiding force according to the distribution of trajectories.Firstly,trajectory distribution is learned by the hidden semi-Markov model and Gaussian mixture model.Secondly,an adaptive virtual fixture based on a motion refinement tube in Cartesian space is proposed,which can easily adjust the interaction force through nonlinear stiffness terms.To extend the method to Cartesian space and avoid tube deformation caused by ignoring non-diagonal elements of a covariance matrix,a local coordinate system was proposed based on the covariance matrix,and the tube radius and nonlinear stiffness terms were calculated in the local coordinate system.Then,a method based on a virtual energy tank is designed to ensure the stability of the system.Finally,the effectiveness of the proposed method is verified from the aspects of operation efficiency,safety,and interaction force through comparative experiments and user studies.In chapter 5,aiming at the fact that the traditional virtual fixture cannot be applied in complex multi-task situations,based on the research in Chapter 4,the adaptive virtual fixture is extended to the multi-task scenario.The multi-task function is realized by activating multiple adaptive virtual fixtures at the same time in a probabilistic way.The guiding force of each adaptive virtual fixture is superimposed in a probabilistic way.Then,a passive controller based on an energy tank is designed for the multi-task adaptive virtual fixture,and the system stability is analyzed by the Lyapunov function.Finally,a multitask scenario is built.Compared with the existing method,the effectiveness of the proposed method is verified from the aspects of efficiency,security,and interaction force.In chapter 6,the main research contents and the innovations are summarized.Finally,future research is given.