Artificial data generation pipeline for visual grasping deep learning

• Ano de defesa: 2023
• Autor(a) principal: Casteluci, Larissa Cassador
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: aprendizagem profunda; computer vision; dados sintéticos; deep learning; detecção de preensão com visão; synthetic data; visão computacional; visual grasp detection
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/18/18162/tde-16062023-112737/

The rise of deep learning algorithms in academia has changed the area of robotic grasping. Before, methods involving analytical analysis and grasping modelling were the most common strategies. However, deep learning strategies have become recently more prevalent. They have presented incredible results in the last decade. However, they present disadvantages of their own. A major drawback is that they require large amounts of representative data to be trained on. For specific applications, a specific dataset with custom targets is required. But generating data for robotic grasping is not an easy task. It is more challenging than creating datasets for classification or object detection problems, since it requires lab experiments. Manual acquisition of this data can be time-consuming. In this context, the generation of synthetic data using rendering and simulation tools can be a viable solution. This strategy, on the other hand, also has its own set of problems. The most relevant is the reality gap, i.e. the intrinsic difference between reality and simulated data. There are a few techniques developed to mitigate this problem, such as domain randomization and photorealistic data. We provide a tool that allows the creation of datasets for robotic grasping for a configurable set of targets. We compare in a real life scenario a neural network trained on this custom dataset and compare results with the same network trained on a state-of-the-art dataset and show that our tool creates viable datasets that neural networks can be trained on and produce suitable results.