Study On Human-robot Skill Transfer

Production pattern needs to be improved to adapt to the new demands of industrial products by the market.It is getting more and more difficult for industrial robots with drawbacks of low programming efficiency,high cost for applications and the lack of flexibility to meet the requirements of the new generation of manufacturing.It has become a bottleneck for industry upgrading and needs to be addressed urgently.Human-robot skill transfer(HRST)can enable the transferring of human compliant skills to robots.HRST has two distinguishing characteristics:the first one is that robots can be efficiently programmed through HRST,therefore,it will be very convenient for robotic applications and lower labor and time cost;the second one is that HRST can allow robots to learn human-like compliant manipulation skills in order to improve the flexibility of the robot-based manufacturing system,and therefore,the application domain of robot-based manufacturing can be accordingly expanded.In this sense,HRST can be very promising for the addressing of the above problems,and becomes a very hot topic in this field.This thesis first elaborates the research background and significance of HRST,from the point view of the development status of industrial robots.The state-of-the-art is then summarized which is mainly focused on two aspects: i)the acquisition and representation of motion skills;and ii)the learning of compliant skills.For the first aspect,the general procedure of HRST is introduced,including three phases: demonstration,model learning,and task reproduction;the main human-robot interaction systems for HRST is presented,including vision-based;teleoperation-based,and human-robot physical interaction;the common used models that can encode motion skills are introduced,including dynamical movement primitives and probabilistic models.For the second aspect,two approaches of learning of compliant skills are given,including the EMG-based stiffness transfer from humans to robots;and human-like biomimetic controller based on human motor learning strategies.A teaching-by-demonstration system is established for transfer of human variable stiffness skills to robots.The human arm stiffness during demonstration is estimated based on EMG signals collected from human arm muscles,then the human arm stiffness is mapped into the robotic arm joint space,thus that the robot can learn the stiffness regulation strategies from a human user.In order to improve the generalization ability,a HRST framework based DMP is presented,in which the motion trajectories and stiffness profiles are encoded in parallel.Through this framework the motion trajectories and the stiffness profiles can be adjusted to enable the robot to adapt to new task situations.Furthermore,alignment and segmentation techniques are also integrated into the framework to make it more convenient for adaptation of these profiles.Several tasks are performed to verify the proposed framework.A multimodal learning approach is proposed to enable robots to learn more complete skill features from humans.Three types of signals including robotic endpoint movement states,human arm EMG signals and interaction force between the robot and its environment are collected during the demonstration phase.Hidden semi-Markov model(HSMM)is used to separately encode the variable concentration,i.e.,position-velocity,position-stiffness,and position-force.Then Gaussian mixture regression(GMR)is used to generate the desired control variables based on the estimated HSMM parameters and the position trajectory.Comparative experiments have validated the effectiveness of our approach.In order to improve the efficiency of HRST,a robotic learning approach based on humanlike biomimetic control strategies is proposed in this thesis.This approach includes two phases:kinematic demonstration and compliant reproduction.During the reproduction phase,the robot can online adapt its arm impedance and feedforward force to the task based on the motion errors.Through human-robot collaboration and interaction tasks the robot has shown that it can compliantly adapt to its environment or partner by using the proposed approaches.