DTA

Digital Theses Archive

 

Tesi etd-09062022-092828

Type of thesis
Dottorato
Author
NOCENTINI, OLIVIA
URN
etd-09062022-092828
Title
Study and development of AI-based approaches for cloth manipulation capabilities
Scientific disciplinary sector
ING-IND/34
Course
Istituto di Biorobotica - PHD IN BIOROBOTICA
Committee
relatore Prof. CAVALLO, FILIPPO
Keywords
  • assistive manipulator
  • cloth manipulation
  • neural networks
Exam session start date
26/01/2023;
Availability
parziale
Abstract
According to the World Health Organization forecasts, there are already more than one<br>billion people over the age of 60, with that figure predicted to climb to 1.4 billion by<br>2030. As a result, there is a growth in the need for caregivers, which may become<br>unsustainable for the future society. In this scenario, as the world population ages, the<br>demand for automated help grows. Service robotics is one area of robotics where robots<br>have shown significant promise in working closely with people. Hospitals, household<br>settings, and elderly homes will need intelligent robotic agents in use that perform daily<br>activities. Cloth manipulation is one such daily activity and represents a challenging<br>area for a robot.<br>This thesis details two main aspects: the cloth image classification and the identification<br>of grasping points for clothing manipulation. Taking into account the cloth classification<br>task, we first enhance state-of-the-art convolutions neural networks, e.g., LeNET, by<br>adding additional image processing features to the network structure. Finally, to further<br>improve the accuracy, we investigated the implementation of multiple convolutional<br>neural networks (MCNN). The proposed networks are trained and tested on the Fashion-<br>MNIST dataset.<br>The second research goal of this thesis focused on finding the grasping points of the<br>highest wrinkle (from a later point of view) of a folded hospital gown. The wrinkle is<br>detected using the Generative Grasping Convolutional Neural Network (GGCNN), while<br>the approach to the cloth by a manipulator is obtained by designing a visual servoing<br>algorithm that considers the input of the GGCNN.<br>In conclusion, the results described in this thesis tend to study by deep some AI-based<br>approaches for cloth manipulation capabilities; in particular, we concentrated on study-<br>ing the cloth image classification with neural networks and the Fashion-MNIST dataset.<br>Moreover, we analysed how to identify the first wrinkle of a cloth by combining the<br>visual servoing approach with a neural network.
Files