Research and Analysis of Robot Grasping Deep Learning: From End-to-End Model to Pre Training Foundation

Abstract

Robotic grasping, a cornerstone of autonomous manipulation, has been profoundly transformed by deep learning. This survey provides a comprehensive overview of the current state of deep learning-based methods for robotic grasping, highlighting the paradigm shift from traditional multi-stage pipelines to data-driven approaches (with end-to-end learning as a core branch). The paper systematically categorizes and analyzes key methodologies, including end-to-end grasp estimation (covering both planar and 6-DoF spatial grasping), multi-modal fusion (RGB-D, vision-language), reinforcement and imitation learning, and the emerging application of large-scale pre-trained models. The review synthesizes findings from prominent datasets (e.g., Cornell, GraspNet-1Billion) and evaluates performance against core metrics like grasp success rate, inference time, and generalization ability. Crucially, this survey emphasizes the practical applicability of these technologies, linking them to specific real-world scenarios such as industrial bin-picking and domestic service tasks. Despite significant progress, critical challenges persist, such as the sim-to-real gap, limited generalization, and the trade-off between real-time performance and computational cost. We discuss these open challenges and outline promising, executable future directions—such as domain adaptation for sim-to-real transfer and model lightweighting for edge deployment—to bridge the gap between academic research and industrial deployment.