Hybrid deep learning for lifelong 3D recognition and grasping in service robots

In recent years, service robots have become increasingly prevalent in both daily life and industrial environments. In these open and dynamic settings, robots are typically required to continuously and efficiently learn, accurately recognize novel objects, and flexibly adapt to diverse tasks. Achieving these capabilities relies heavily on advanced visual perception and manipulation technologies. However, current methods still face substantial limitations in addressing challenges such as learning from a few samples, fine-grained recognition, multi-view information fusion, and object grasping in complex scenes. To overcome these issues, this thesis proposes a series of innovative methods that integrate hybrid deep learning with multimodal representations, aiming to enhance the accuracy and robustness of robotic three-dimensional (3D) object recognition and grasping in real-world environments. Experimental evaluations demonstrate that these proposed methods surpass existing approaches in tasks including lifelong few-shot learning, fine-grained object recognition, multi-view 3D recognition, and robotic grasping. These contributions offer promising theoretical insights and practical applications for robotic systems.