Learning-Based Control Scheme Of A Robotic Exoskeleton With Physical Human-Robot Interaction

As a class of wearable auxiliary devices,robotic exoskeletons are applied in numerous fields,calling for the higher requirements in safety and efficiency.Therefore,research work on the safety and efficiency of physical human-robot interaction(pHRI)control strategies for robotic exoskeletons becomes more important.Focusing on the robotic exoskeleton,a hierarchical control strategy based on Learning from Demonstration(LfD)for pHRI is proposed in the thesis.The proposed control strategy consists of two layers,the low-layer is to obtain the motion control enabling the robotic exoskeleton to conform to the human movement and achieve the safe pHRI,and the high-layer is to get the LfD-based strategy so the robot can assist the wearer to complete the interactive task and provide power only when needed.The main research contents of the thesis are summarized as follows:(1)Aiming at the back-drivability problem of robotic exoskeleton,an admittance control scheme with an internal position controller is designed.Based on the asymmetric barrier Lyapunov function(ABLF)adaptive neural network(NN)controller,an internal position admittance controller is designed to solve the difficulties in the asymmetric physical constraints of the robot and the unknowns in the dynamic model,and the inverse kinematics solution process is avoided as well.(2)Research on a hierarchical control strategy of robotic exoskeleton based on LfD.In the proposed control method,the admittance control scheme is based on ABLF adaptive NN controller as the low-layer motion control method.For the high-level interaction,a strategy for learning human skills from demonstration is proposed in the thesis,which consists of three phases,demonstration,learning and reproduction.In the demonstration phase,the robot observes the way of the demonstrator performing a specific interactive task successfully;in the learning phase,the robot uses the Gaussian Mixture Model(GMM)to learn the data being recorded during demonstration;in the reproduction phase,the robot can predict the assistive force by utilizing the learned GMM which fuses with the Gaussian mixture regression(GMR)algorithm.To improve the generalization ability of the strategy,the task parameterized Gaussian mixture model(TP-GMM)is extended to combine with the proposed method so that the interactive motion trajectory can converge to a new target point.In this thesis,the proposed methods are validated on the upper extremity exoskeleton robot platform.The experimental results show the feasibility of the methods.The proposed methods provide a new training pattern for the medical rehabilitation to reduce the work intensity of physical therapists and realize the active training of patients.