{"id":{"repo_id":"alicante","oai_identifier":"oai:rua.ua.es:10045/164116"},"canonical_url":"https://search.dev.ndltd.org/etd/alicante/oai:rua.ua.es:10045/164116","repository":{"repo_id":"alicante","name":"University of Alicante","base_url":"https://rua.ua.es/server/oai/request"},"display":{"title":"Multisensorial perception for grasping objects with multifingered grippers","abstract":"The ability to perceive and manipulate the environment is key to the survival of the human species. This underscores the significance of perception and grasping for humans, which is essential for robots. According to some of the world’s finest roboticists on perception and manipulation, there are still open challenges. This thesis seeks to enhance our profound yet limited understanding of these domains. Firstly, it is imperative to comprehend the environment through perception and sensing. The present research has explored the utilisation and analysis of a variety of available perception technologies. The research was initiated with RGB cameras and was subsequently expanded to RGBD cameras and less commonly used sensors, such as Light Detection And Ranging (LiDAR) and multispectral cameras. RGB cameras were utilised to train several State-Of-The-Art (SOTA) Neural Networks (NNs) to detect and segment objects in outdoor environments. The development of several applications followed the training of these models. One such application, using RGBD cameras, was designed to calculate the depth of objects. While the obtained results were satisfactory, the analysis of LiDAR-camera fusion technology aimed to enhance the work range. Consequently, a method for enhancing the accuracy and density of low-layered sensors was developed. Furthermore, other applications utilising LiDAR sensors have been inquired, incorporating high-layered sensors that yield images analogous to those captured by low-resolution RGB cameras. In this case, the objective was to determine whether these images permit the direct detection and segmentation of objects in LiDAR images. Finally, multispectral cameras offer the ability to discern the invisible, but multispectral multilens cameras are faced with the problem of labelling. Another work has enabled the automation of the labelling process, thereby reducing the time required. Specifically, the application employs lens transformation calculation to label multiple images with minimal effort. Secondly, it is vital to be able to interact with the perceived environment. It is fortunate that humans are born with this innate capacity, which is developed even after reaching a conscious state. Nevertheless, the transition to robotics is challenging, and it has been a subject of study by the scientific community for many years. The two main approaches are analytical and data-driven methods. Analytic methods utilise the properties of the object surface to mathematically determine the optimal candidate points for grasping. Conversely, data-driven methods employ datasets to identify hidden data patterns to achieve the same goal, but using Deep Learning (DL) techniques. The former technique was employed to develop a fully functional pipeline for a robotic system to collect litter in outdoor environments, utilising the GeoGrasp algorithm. It is important to note that GeoGrasp is only able to handle parallel grippers; creating GeoGraspEvo was therefore necessary. This algorithm provides grasping points for multifinger grippers, with experiments in the real world demonstrating its effectiveness. The latter technique was utilised in the development of QDGset. It consists of a large-scale dataset of approximately 60M parallel jaw gripper grasps on 40k simulated objects. The generation of this dataset involved the extension of QDG-6DoF (a system based on an evolutionary algorithm called Quality-Diversity (QD), which generates object-centric grasps) to produce grasping datasets. The field of robotics shows great potential for development over the next few years. The objective of this thesis is to facilitate the advancement of scientific knowledge through a range of works, with the hope that they will serve the community to advance the frontiers of knowledge.","abstract_html":"The ability to perceive and manipulate the environment is key to the survival of the human species. This underscores the significance of perception and grasping for humans, which is essential for robots. According to some of the world’s finest roboticists on perception and manipulation, there are still open challenges. This thesis seeks to enhance our profound yet limited understanding of these domains. Firstly, it is imperative to comprehend the environment through perception and sensing. The present research has explored the utilisation and analysis of a variety of available perception technologies. The research was initiated with RGB cameras and was subsequently expanded to RGBD cameras and less commonly used sensors, such as Light Detection And Ranging (LiDAR) and multispectral cameras. RGB cameras were utilised to train several State-Of-The-Art (SOTA) Neural Networks (NNs) to detect and segment objects in outdoor environments. The development of several applications followed the training of these models. One such application, using RGBD cameras, was designed to calculate the depth of objects. While the obtained results were satisfactory, the analysis of LiDAR-camera fusion technology aimed to enhance the work range. Consequently, a method for enhancing the accuracy and density of low-layered sensors was developed. Furthermore, other applications utilising LiDAR sensors have been inquired, incorporating high-layered sensors that yield images analogous to those captured by low-resolution RGB cameras. In this case, the objective was to determine whether these images permit the direct detection and segmentation of objects in LiDAR images. Finally, multispectral cameras offer the ability to discern the invisible, but multispectral multilens cameras are faced with the problem of labelling. Another work has enabled the automation of the labelling process, thereby reducing the time required. Specifically, the application employs lens transformation calculation to label multiple images with minimal effort. Secondly, it is vital to be able to interact with the perceived environment. It is fortunate that humans are born with this innate capacity, which is developed even after reaching a conscious state. Nevertheless, the transition to robotics is challenging, and it has been a subject of study by the scientific community for many years. The two main approaches are analytical and data-driven methods. Analytic methods utilise the properties of the object surface to mathematically determine the optimal candidate points for grasping. Conversely, data-driven methods employ datasets to identify hidden data patterns to achieve the same goal, but using Deep Learning (DL) techniques. The former technique was employed to develop a fully functional pipeline for a robotic system to collect litter in outdoor environments, utilising the GeoGrasp algorithm. It is important to note that GeoGrasp is only able to handle parallel grippers; creating GeoGraspEvo was therefore necessary. This algorithm provides grasping points for multifinger grippers, with experiments in the real world demonstrating its effectiveness. The latter technique was utilised in the development of QDGset. It consists of a large-scale dataset of approximately 60M parallel jaw gripper grasps on 40k simulated objects. The generation of this dataset involved the extension of QDG-6DoF (a system based on an evolutionary algorithm called Quality-Diversity (QD), which generates object-centric grasps) to produce grasping datasets. The field of robotics shows great potential for development over the next few years. The objective of this thesis is to facilitate the advancement of scientific knowledge through a range of works, with the hope that they will serve the community to advance the frontiers of knowledge.","abstract_has_math":false,"creators":["Páez Ubieta, Ignacio de Loyola"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025","date_published":"2025","updated_at":"2026-07-24T00:52:09Z","subjects":["LiDAR","LiDAR-camera Fusion","Multispectral Imagery","Grasping Analytic Methods","Grasping Data-driven Methods"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:10045/164116"],"render_values":[{"text":"hdl:10045/164116","href":null,"code":true}]}]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null},"metadata_groups":[{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2025"]},{"key":"dc:type","label":"Dc Type","values":["info:eu-repo/semantics/doctoralThesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["LiDAR","LiDAR-camera Fusion","Multispectral Imagery","Grasping Analytic Methods","Grasping Data-driven Methods"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:10045/164116"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.other","label":"Dc Description Other","values":["The ability to perceive and manipulate the environment is key to the survival of the human species. This underscores the significance of perception and grasping for humans, which is essential for robots. According to some of the world’s finest roboticists on perception and manipulation, there are still open challenges. This thesis seeks to enhance our profound yet limited understanding of these domains. Firstly, it is imperative to comprehend the environment through perception and sensing. The present research has explored the utilisation and analysis of a variety of available perception technologies. The research was initiated with RGB cameras and was subsequently expanded to RGBD cameras and less commonly used sensors, such as Light Detection And Ranging (LiDAR) and multispectral cameras. RGB cameras were utilised to train several State-Of-The-Art (SOTA) Neural Networks (NNs) to detect and segment objects in outdoor environments. The development of several applications followed the training of these models. One such application, using RGBD cameras, was designed to calculate the depth of objects. While the obtained results were satisfactory, the analysis of LiDAR-camera fusion technology aimed to enhance the work range. Consequently, a method for enhancing the accuracy and density of low-layered sensors was developed. Furthermore, other applications utilising LiDAR sensors have been inquired, incorporating high-layered sensors that yield images analogous to those captured by low-resolution RGB cameras. In this case, the objective was to determine whether these images permit the direct detection and segmentation of objects in LiDAR images. Finally, multispectral cameras offer the ability to discern the invisible, but multispectral multilens cameras are faced with the problem of labelling. Another work has enabled the automation of the labelling process, thereby reducing the time required. Specifically, the application employs lens transformation calculation to label multiple images with minimal effort. Secondly, it is vital to be able to interact with the perceived environment. It is fortunate that humans are born with this innate capacity, which is developed even after reaching a conscious state. Nevertheless, the transition to robotics is challenging, and it has been a subject of study by the scientific community for many years. The two main approaches are analytical and data-driven methods. Analytic methods utilise the properties of the object surface to mathematically determine the optimal candidate points for grasping. Conversely, data-driven methods employ datasets to identify hidden data patterns to achieve the same goal, but using Deep Learning (DL) techniques. The former technique was employed to develop a fully functional pipeline for a robotic system to collect litter in outdoor environments, utilising the GeoGrasp algorithm. It is important to note that GeoGrasp is only able to handle parallel grippers; creating GeoGraspEvo was therefore necessary. This algorithm provides grasping points for multifinger grippers, with experiments in the real world demonstrating its effectiveness. The latter technique was utilised in the development of QDGset. It consists of a large-scale dataset of approximately 60M parallel jaw gripper grasps on 40k simulated objects. The generation of this dataset involved the extension of QDG-6DoF (a system based on an evolutionary algorithm called Quality-Diversity (QD), which generates object-centric grasps) to produce grasping datasets. The field of robotics shows great potential for development over the next few years. The objective of this thesis is to facilitate the advancement of scientific knowledge through a range of works, with the hope that they will serve the community to advance the frontiers of knowledge."]},{"key":"dc:title","label":"Title","values":["Multisensorial perception for grasping objects with multifingered grippers"]}]}],"canonical_facts":{"dc:date.issued":["2025"],"dc:description.other":["The ability to perceive and manipulate the environment is key to the survival of the human species. This underscores the significance of perception and grasping for humans, which is essential for robots. According to some of the world’s finest roboticists on perception and manipulation, there are still open challenges. This thesis seeks to enhance our profound yet limited understanding of these domains. Firstly, it is imperative to comprehend the environment through perception and sensing. The present research has explored the utilisation and analysis of a variety of available perception technologies. The research was initiated with RGB cameras and was subsequently expanded to RGBD cameras and less commonly used sensors, such as Light Detection And Ranging (LiDAR) and multispectral cameras. RGB cameras were utilised to train several State-Of-The-Art (SOTA) Neural Networks (NNs) to detect and segment objects in outdoor environments. The development of several applications followed the training of these models. One such application, using RGBD cameras, was designed to calculate the depth of objects. While the obtained results were satisfactory, the analysis of LiDAR-camera fusion technology aimed to enhance the work range. Consequently, a method for enhancing the accuracy and density of low-layered sensors was developed. Furthermore, other applications utilising LiDAR sensors have been inquired, incorporating high-layered sensors that yield images analogous to those captured by low-resolution RGB cameras. In this case, the objective was to determine whether these images permit the direct detection and segmentation of objects in LiDAR images. Finally, multispectral cameras offer the ability to discern the invisible, but multispectral multilens cameras are faced with the problem of labelling. Another work has enabled the automation of the labelling process, thereby reducing the time required. Specifically, the application employs lens transformation calculation to label multiple images with minimal effort. Secondly, it is vital to be able to interact with the perceived environment. It is fortunate that humans are born with this innate capacity, which is developed even after reaching a conscious state. Nevertheless, the transition to robotics is challenging, and it has been a subject of study by the scientific community for many years. The two main approaches are analytical and data-driven methods. Analytic methods utilise the properties of the object surface to mathematically determine the optimal candidate points for grasping. Conversely, data-driven methods employ datasets to identify hidden data patterns to achieve the same goal, but using Deep Learning (DL) techniques. The former technique was employed to develop a fully functional pipeline for a robotic system to collect litter in outdoor environments, utilising the GeoGrasp algorithm. It is important to note that GeoGrasp is only able to handle parallel grippers; creating GeoGraspEvo was therefore necessary. This algorithm provides grasping points for multifinger grippers, with experiments in the real world demonstrating its effectiveness. The latter technique was utilised in the development of QDGset. It consists of a large-scale dataset of approximately 60M parallel jaw gripper grasps on 40k simulated objects. The generation of this dataset involved the extension of QDG-6DoF (a system based on an evolutionary algorithm called Quality-Diversity (QD), which generates object-centric grasps) to produce grasping datasets. The field of robotics shows great potential for development over the next few years. The objective of this thesis is to facilitate the advancement of scientific knowledge through a range of works, with the hope that they will serve the community to advance the frontiers of knowledge."],"dc:identifier":["hdl:10045/164116"],"dc:subject":["LiDAR","LiDAR-camera Fusion","Multispectral Imagery","Grasping Analytic Methods","Grasping Data-driven Methods"],"dc:title":["Multisensorial perception for grasping objects with multifingered grippers"],"dc:type":["info:eu-repo/semantics/doctoralThesis"]},"updated_at":"2026-07-24T00:52:09Z"}