6-DoF Dual-Arm Imitation Learning with Position Control based Self-Collision Avoidance for Service Robot

Abstract

This thesis proposes an imitation learning system which consists of an imitation system a self-collision avoidance system and a motion learning system The imitation system captures the human skeleton with an RBG-D camera Kinect 2 0 to map the human joint angles to the motor angles of the robot at here an integrating method of the forward kinematics and the space vector method is developed To improve safety the self-collision avoidance system is added so that the robot will not collide with itself while learning human motions This system is based on the position control method; with an impedance subsystem it calculates the possibility of collisions and generates a repulsive force to further calculate the corresponding displacement of the robot arm In order to maintain the similarity to the human motion while avoiding collision integrated adaptive constricted particle swarm optimization algorithm (PSO-IAC) is used to calculate the corresponding motor angles The learned motion is then recorded in a motion dataset as a reference motion To deal with different situations and objects the learned motion is then adjusted by another 2-stage PSO-IAC The first PSO-IAC generates a suitable motion trajectory based on the real situation and the second PSO-IAC calculates the motor angles according to the trajectory The real experiments demonstrate the efficiency of the whole imitation learning system The robot can learn a reference motion using its vision and generate a suitable motion based on the learned one for coping with different objects and situations

Date of Award	2018 Jul 6
Original language	English
Supervisor	Tzuu-Hseng S. Li (Supervisor)