Research On Recognition And Realization Of Human-robot Collaboration Intention Based On Human Motion Language

With the transformation of manufacturing towards personalization and flexibility,researchers are increasingly interested in high flexibility and conveniently operated humanrobot collaboration technology.However,limitations in the acquisition and transmission of interactive information,robot perception and decision-making abilities,and the intelligent control algorithm,mean that current human-robot collaboration technology cannot accurately identify and implement human motion intentions.With the development of the biomedical field,researchers have discovered that Electromyography(EMG)signals contain a wealth of human motion and force information.This provides a more natural way to acquire human instinctive decision-making abilities and collaborative strategies.This dissertation focuses on the key technical issues of human intention recognition and realization in human-robot collaboration,taking two typical collaborative scenarios,namely human-robot collaborative sawing and collaborative assembly,as examples.Imitative learning,reinforcement learning,and other intelligent algorithms are used to acquire human instinctive collaboration skills from the human motion language information,such as human EMG signals and joint rotation signals.The aim is to eliminate understanding differences in human-robot collaboration and establish a humanin-the-loop intelligent control model suitable for a wide range of task scenarios.The main research work is as follows:Specific solutions are proposed for filtering delays,joint rotation effects,and tremor effects in feature extraction of force-related information.Considering that equal weight Root Mean Square(RMS)filters can exacerbate the lag in feature extraction,a variable weight RMS filtering method with forgetting factors is proposed.By appropriately forgetting the past information of EMG signals and highlighting the current information,the real-time performance of feature extraction is improved.To balance the fluctuation and hysteresis of extracted EMG signal features and improve the quality of feature extraction,the Bayesian hyperparametric optimization method is used to optimize filter parameters.Aiming at the influence of joint rotation and output force-joint rotation coupling effects on force-related information extraction in motion space,a modified Fast Orthogonal Search(FOS)algorithm is proposed by using the correlation coefficient between extracted feature signals and arm output force as an optimization criterion.To solve the impact of involuntary arm tremors on the accuracy of the arm force/torque estimation model,a fast Fourier transform(FFT)is used to extract tremor features from raw EMG signals.An adaptive correction method based on neural networks is proposed to adaptively convert the tremor features into discount factors of low-pass discrete filters,further correcting the extracted information of EMG signals.Construct a human motion intention recognition model,and explore the feasibility of using human motion intention to guide robots to participate in the collaboration.Taking humanhuman collaborative sawing and lifting as examples,the regularity between the speed adjustment and the arm stiffness in human collaboration is analyzed,and the feasibility of sharing human motion intentions based on arm force/torque information is confirmed.To achieve the inference and sharing of human motion directions,the position adjustment intention inference model based on the naive Bayesian algorithm and the posture adjustment intention inference model based on multimodal information weight self-allocation network are constructed by exploring the characteristics of joint rotation angles and EMG signals during position/posture adjustment.Aiming at the anisotropy of the arm force/torque model,a parallel LSTM neural network based arm output force/torque estimation method with the help of the position/posture adjustment intention inference model is proposed,which improves the accuracy of the arm force/torque estimation model by fitting the arm force/torque in each direction separately.Three human-robot collaborative control methods are proposed,and human-robot cooperative sawing is taken as an example to verify and compare the proposed models,which provides guidance for expanding complex collaborative tasks.An adaptive variable impedance control method based on human body stiffness is proposed.Based on the tracking error of robots to human intentions,an adaptive variable impedance control controller is designed,and its stability and convergence are demonstrated.An imitation learning model of human skills based on multiple model Gaussian Process Regression is proposed,which uses probability estimation to learn human speed regulation skills in the face of external force changes from inaccurate human demonstrations.A data-driven reinforcement learning end-to-end collaborative control model is proposed.From the perspective of ergonomics,the collaborative control model is optimized online,enabling robots to gradually understand and match human motion intentions.Construct a fast transfer reinforcement learning method to rapidly expand and generalize control models from simple scenes to complex scenes.Using the ability of fuzzy rules to assign weights to multiple attribute input information,a fusion normalization method based on fuzzy rules is proposed to ensure that the control model has the same input and output scales before and after transfer learning.To reduce the input samples of the control model and simplify the complexity of the control model,a parallel collaborative control method based on the position/posture adjustment intention inference model is proposed for the multi-dimensional task space.Multiple sets of position/ posture adjustment control models are used to learn the collaborative skills of humans in different motion directions.To reduce the training cost of online learning human collaborative skills and bring superiority into play the commonalities among control models,a transfer learning network based on assembly time evaluation and a fast fine-tuning algorithm based on node importance are proposed.Through a more purposeful step-by-step transfer learning method,the compliance following ability and accurate execution ability of collaborative control models are gradually improved.Three experimental scenarios,namely,collaborative sawing tasks,collaborative assembly posture adjustment tasks,and collaborative assembly position adjustment tasks are used to verify the effectiveness and feasibility of the EMG signal feature extraction,motion intention recognition,and human-robot collaborative control and fast transfer generalization methods proposed in this dissertation.