A Review of Deep Learning-Based Industrial Robot Grasping Systems

Traditional industrial robot grasping methods rely on predefined rules and structured environments, making them ill-suited for complex scenarios and dynamic changes. With breakthroughs in deep learning across visual perception and reinforcement learning, robotic grasping technology is progressively shifting toward a data-driven intelligent paradigm. This paper focuses on two key technologies: visual perception and grasping strategy optimization. In visual perception, grasp detection and 6D pose estimation based on convolutional neural networks significantly enhance system robustness in complex environments. In strategy optimization, imitation learning and deep reinforcement learning offer novel pathways for autonomous decision-making and skill acquisition in dynamic environments. However, current systems still face challenges including data scarcity, simulation-reality gaps, and insufficient multimodal perception fusion. Future research should focus on few-shot learning, transfer techniques from simulation to reality, multimodal information fusion, and non-rigid object grasping to advance the development of general-purpose intelligent grasping systems.