Not stated in source 2021 Needs review

Real-Time Object Detection and Grasping for a Robotic Arm

Abstract

Research and development with robotic arms have been explored for their potential to perform various physical tasks such as sorting items by colors and shapes.Prior work and research made advancements with robotic arm systems by combining them with deep neural networks (DNN) and computer-based vision to detect and interact with objects.These systems take a massive amount of time to deploy and develop due to time-consuming DNN training costs.This training cost staggers the development of robotic arms since computer vision needs a DNN to operate.Many of these systems rely on object detection to pick up items from a fixed camera overhead within their field of vision.Fixed overhead cameras compensate for the limitation of computer vision to properly locate and estimate the distances of the object in 3D space.Such that an overhead camera will assume detected objects will be on the surfaces directly in front.Research to estimate distance with computer vision relies on well-trained DNN to estimate accurately.Computer vision distance estimation has had yet to be successful with robotic arm systems.This project aims to reduce the training cost on DNN by using Transfer-Learning, to promptly deploy a well-trained DNN to a robotic arm system to detect an object and estimate its distance from a non-fixed field of vision to grasp objects in real-time.

Keywords

Not available

Specialisations

Similar theses