Development of a learning system for robot control

Abstract

With the recent advancement of robotics manipulators and neural networks. As robot manipulator often requires managing a variety of tasks regarding grasping an object, it requires recognizing the object. Object detection has been a popular research topic, however, the existing method still proves to have challenges with occlusion handling. This project aims to solve the issue with occlusion by using the capability of the YOLOv5 model in fast and accurate object detection, and Transformer Network (TF) with trajectory predictions which proves to outperform current LSTM. This report will analyse the robustness of the YOLO model and TF network along with the capability of occlusion handling when combining both of them.