A carregar...
Publicação
Vision Based Bin-Picking of Inserts for Injection Moulding
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 7.62 MB | Adobe PDF |
Autores
Resumo(s)
Este projeto consiste na identificação de metodologias de visão computacional e hardware para o desenvolvimento de um sistema de bin-picking. O protótipo desenvolvido identifica insertos, produzidos num tecido preto e rígido, dentro de uma caixa e posiciona-os de uma forma organizada. O posicionamento dos insertos tem de ser preciso e permitir que um robô cartesiano integrado numa máquina de injeção os apanhe corretamente para posicionamento dentro do molde de injeção.
Numa fase inicial deste projeto, foi realizado um estudo do estado da arte para identificar as metodologias de visão computacional e de hardware que melhor se adaptam a situações em que objetos se encontram de forma desorganizada e com oclusões. Com base nas conclusões deste estudo, um protótipo de um sistema de bin-picking foi desenvolvido incluindo um robô com 6 graus de liberdade, um sensor RGBD, métodos clássicos de visão computacional e, também metodologias de aprendizagem automática.
Os primeiros testes experimentais tiveram como objetivo determinar oportunidades de melhoria ao protótipo, e estas melhorias foram implementadas. O desempenho do sistema foi avaliado com base num teste de esvaziamento da caixa e ensaios com diferentes quantidades de peças dentro da caixa. Nesta avaliação foram utilizadas métricas como “Tempo de ciclo médio”, “Número máximo de tentativas até apanhar uma peça corretamente” e “Análise de rejeições”.
O protótipo atingiu o desempenho desejado, cumprindo o tempo de ciclo. Os resultados obtidos confirmam a viabilidade do sistema; ainda assim, são propostas melhorias para o desenvolvimento do sistema industrial nas suas condições de funcionamento finais.
This project involves identifying computer vision technologies and hardware for the development of a bin-picking system. The developed prototype identifies parts made from a dark, stiff fabric material within a bin and places them in an organized way. The placement of the parts has to be precise and allow a cartesian robot tending an injection machine to pick them in the correct orientation for placement inside the mould. In the initial stage of this project, research was conducted to identify computer vision technologies and hardware that better adapt to unknown cluttered environments. According to the conclusions taken from the previous stage, a bin-picking prototype was developed, with the aid of a 6 degree of freedom robotic manipulator, an RGBD sensor, classic computer vision technologies, as well as deep learning models. Initial tests were run to determine areas for optimization of the prototype, and improvements were implemented. The prototype’s performance was assessed through a single-fill test and trials with varying part densities in the bin. Multiple metrics were analysed under different conditions to assess the prototype’s performance, such as “Average cycle time,” “Number of attempts until successful pick” and “Rejection analysis”. Overall, the prototype achieved the desired performance, meeting the required cycle time. The results confirm the feasibility of the developed system; however, further refinements are proposed for the development of the final industrial system under its intended operating conditions.
This project involves identifying computer vision technologies and hardware for the development of a bin-picking system. The developed prototype identifies parts made from a dark, stiff fabric material within a bin and places them in an organized way. The placement of the parts has to be precise and allow a cartesian robot tending an injection machine to pick them in the correct orientation for placement inside the mould. In the initial stage of this project, research was conducted to identify computer vision technologies and hardware that better adapt to unknown cluttered environments. According to the conclusions taken from the previous stage, a bin-picking prototype was developed, with the aid of a 6 degree of freedom robotic manipulator, an RGBD sensor, classic computer vision technologies, as well as deep learning models. Initial tests were run to determine areas for optimization of the prototype, and improvements were implemented. The prototype’s performance was assessed through a single-fill test and trials with varying part densities in the bin. Multiple metrics were analysed under different conditions to assess the prototype’s performance, such as “Average cycle time,” “Number of attempts until successful pick” and “Rejection analysis”. Overall, the prototype achieved the desired performance, meeting the required cycle time. The results confirm the feasibility of the developed system; however, further refinements are proposed for the development of the final industrial system under its intended operating conditions.
Descrição
Palavras-chave
Bin-picking Manipulação robótica Robótica guiada por visão Deteção de objetos Sensores RGBD