Geometrical block assembly path planning using Faster RCNN and Q-learning

Abstract

Recently, assembling robots have been used, not only in general processing, but also in a eld where new assembling methods and procedures for various objects are applied, and the assembling process is frequently changed. Accordingly, situations have occurred in which it is dicult for non-experts to respond promptly to the situation. The reason for this is that, if the robot does the work, it is necessary to change its procedure according to the robot assembly process plan, such as Robot Grasping Planning, Robot Motion Planning, and Robot Assembly Planning. However, such robot control algorithms are not easy to change quickly, even for robot experts, and it is even more dicult for non-experts who are not accustomed to this. A simpler, countermeasure is needed to apply in the eld. As a method to solve this problem, this paper provides an implementation of an assembly process that can minimize human intervention by using an articial intelligence algorithm. The simulator for verication of the results of this paper is divided into two stages based on Gazebo, which is one of the 3D simulates of the Robot Operating System (ROS). Block assembly simulation is also implemented. First, as the center of the articial intelligence algorithm, Faster R-CNN is implemented as a step of recognizing the block and grasping the gripping position of the block to be assembled. The second step focuses on the articial intelligence algorithm of Reinforcement Learning, which is modeled by trajectory generation based on two points of pose and trajectory, generated after gripping the block and moving to the position to assemble the block position. This paper shows that it is possible to assemble by articial intelligence algorithms other than Robot Grasping Planning, Robot Motion Planning, and Robot Assembly Planning, which was used in assembly processes and showed the possibility of block assembly through assembly sequence and methods. This paper also proposes that non-experts can quickly respond to articial intelligence algorithms and geometric trajectories at the site where the assembly process is frequently changed.