Research On SEMG-based Human Hand Motion Recognition And Its Applications

Robotic manipulators work as the end-effectors of robotic arms.With the increasing complexity of the grasping and manipulation tasks,the development of robotic manipulators is heading towards increasing the degree-of-freedoms,dexterity and adaptability.At present,the control method of robotic manipulators lags behind the development of the mechanical design,and faces the problems of few motion modes and poor grasping adaptability,which makes it difficult to handle tasks in unknown and changeable scenes.This work aims to transfer humans’ ability of fast motion planning and the adaptability of grasping force to robotic manipulators through ”human-in-the-loop” criterion so that they can realize multi-mode motions,including human-like gestures,and the ability to adjust the contact force in grasping tasks.To realize ”human-in-the-loop” strategy,this thesis studies s EMG-based human hand motion recognition method and its applications on robotic manipulators.The contributions are summarized as follows:This thesis proposes the human hand movement recognition method based on hidden Markov model(HMM).This method uses the observable s EMG signals to model the transitions between the unobservable action primitives(hidden states),thus building the generative model to represent a hand movement.The maximum mutual information(MMI)criterion is applied to optimize the hyperparameters of HMMs.The model built with hyperparameters optimized by MMI have the ability to distinguish itself from the other hand movement models.MMI criterion has increased the recognition accuracy of the generative classifier built with all hand movement models.This thesis proposes the hand movement recognition method based on nonparametric HMM.The hierarchical Dirichlet process(HDP)is introduced as the prior of the model parameters to extend the classical HMM to a model with infinite number of hidden states.In this model,the number of action primitives that belong to a hand movement is not limited to a pre-assigned value,and can adjust as more s EMG signals are observed.The training of HDP-HMM is based on the approximate posterior inference,i.e.,Markov Chain Monte Carlo(MCMC)to sample from the posterior distributions of model parameters.In addition,an real-time probability update procedure based on forward algorithm is proposed for online hand movement recognition.The result of the comparative experiments shows that the number of hidden states has learned from the observed s EMG signals,and competitive accuracy is achieved by nonparametric HMMs compared to the optimal classical HMMs via model selection.This thesis proposes an operation method for multi-finger robotic manipulators based on the pattern recognition of human hand movement.HMMs are built for predesigned hand movements including gestures and grasping movements.Then the probabilities of the processed s EMG signal on each hand movement model are updated in real-time and the recognition result is sent as the high-level pose command to robotic manipulators.The motion planning of the reference pose is completed in the finger joint space and realized by the position controller.Experimental result shows that the proposed method has operated a multi-finger robotic manipulator to realize eight gestures and to grasp objects with different shapes successfully with four grasping movements.This thesis proposes the estimation method of human hand grasping force and its guided grasping control for robotic grippers.The neural network model is used to establish the nonlinear relationship between the forearm s EMG signals and the contact force between the finger and the object,which is used to estimate the contact force during the grasping tasks.The noise from s EMG signals is handled through a combination of signal processing and the intrinsic properties of Neural Networks.In the framework of forceguided control,the grasping force estimated from s EMG signal is used as the force reference,which is tracked by an admittance controller via the inner position control loop.Our method is verified by the successful grasping of typical fragile and deformable objects in daily life.