Vision-based approach for task reconstruction of a robot

Abstract

In this study, we developed a task reconstruction technique with a high success rate, which allows a robot to acquire information to create and reconstruct tasks by using a two-dimensional (2D) camera mounted on the ceiling in a closed space. The robot uses information related to each joint based on the motions learned via imitation learning when specific behaviors are performed. The robot then uses the 2D camera to reconstruct tasks based on the location data related to objects in images after they have been moved. The reconstructed tasks are executed in a closed space where a robot can move around. We introduce a technique that creates and reconstructs tasks, which can be implemented in sequential order using the extracted location data.