Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Design Aspects of a System For Research of Integration of Computer Vision and Machine Learning in Robotics
Submission Author:
Carlos Rodrigues Rocha , RS , Brazil
Co-Authors:
Carlos Rodrigues Rocha, Amanda Jorge Mendes, Kauã Silveira
Presenter: Carlos Rodrigues Rocha
doi://10.26678/ABCM.COBEM2021.COB2021-1243
Abstract
This paper presents the first results in the design and prototyping of an open source manipulator robot system with an integrated intelligent vision module. The project objective is to build a robotic cell for learning and research purposes, with a particular focus on the integration of machine learning and computer vision in robotic manipulation systems. These computational technologies are increasingly adopted in industrial applications, providing extended workspace perception for robotic systems and, consequently, larger operational autonomy. The system was designed for pick-and-place tasks, consisting of a four degree of freedom (P3R) manipulator and a digital camera. Different types of end effector can be use, according to the type of objects to be handled. The relative low-cost, open source characteristics of the project lead to some component and constructive choices. The system controller is composed by a four core single board computer (SBC) and a dual core microcontroller. The computer vision subsystem uses a CCD camera connected to the SBC via a dedicated port. The manipulator has some of its structural/mechanical parts built by additive manufacturing, from its CAD design, and other parts were are common usual in numerical control machines, such as stepper motors, trapezoidal thread spindle and timing belt transmissions. The software for the controller is being developed mainly in Python, both for the SBC and the microcontroller. In the first are the high-level elements, such as trajectory generators, task execution managers, computer vision and machine learning. In the second are the actuator control for the manipulator and sensor data acquisition. OpenCV is used for computer vision, while Tensorflow is used for the machine learning implementation. The first results considered a screws and nuts classification task. Using different image capture/processing resources from the workspace and a convolutional neural network, an accuracy of approximately 99% was obtained. This is a work in progress. The embedded electronics has already been completed and tested, and the physical structure of the manipulator is still under construction. The next stage of the project will carry out the integration of the modules and finalize the implementation of the manipulator control software. It should also be noted that this is a final term work for a Industrial Automation Technical course, and the project should be made available as open source after its completion.
Keywords
Robotic Manipulation, Computer Vision, machine learning, Robotic Cell, Educational Platform
DOWNLOAD PDF VIEW PRESENTATION
