Human-robot Interaction Through Action Learning From Demonstration

As robotics technology developed the robots gradually walk into human life,for which Human-robot Interaction(HRI)became an important research topic in robotics.The tasks would be done better with HRI due to the combination of human and robot specialty.However,there,re still many challenges in HRI,which includes:1)Task recognition:Robot needs to recognize the executing task from the observations of human,which is the inevitable stage for HRI;2)Interactive motion generation:Robot is required to generate the corresponding motion to interact with human,which reflects the motion generalization for the interaction skill;3)Fast programming:Facing the complicate and varied task,the robotics programmers are eager for the friendly programming method,which reduces the programming skill and release the burden for robotics developer.Confronted with these challenges,this thesis introduces the imitation learning to make the robot learn the interaction skills with the human.Through the manually demonstrating the HRI task,we proposed a method to model it in a general way.With the proposed method,the robot can recognize the task from human observations.Then the robot generates the interactive motion to interact with the human.With this in mind,the paper mainly discusses the following three aspects:1.We built the system composed of Baxter robot and human motion tracking device and analyze the calibration between them.With such system,we can demonstrate every HRI task manually to collect the training samples.Every HRI human action is modelled in a general way and we can use it to recognize the human action.To aim at the redundant information in human motion tracking,we studied the feature selection.2.Based on modelling human action in a general way,we studied the relation modelling between the action of human and robot and we studied on how to improve the task recognition accuracy.From the observations from human,the robot can infer the corresponding robot motion trajectory.3.Several interaction tasks were designed in our experiments,we analyzed the proposed model performance from the dataset.We used the motion planning framework MoveIt! to compute the inverse kinematics solution of robot motion.Based on it,we run the interaction task in real Baxter robot to verify the proposed model.With imitation-learning,the robot is able to learn the interaction skill from the manually HRI task demonstrations,which makes the robot to learn the interaction skill with the human.The work makes the significant impact for the HRI application and development.