Hook-picking robots are increasingly widely used in automated production lines, especially in logistics and manufacturing. Hook-picking tasks require robot grippers to have high-precision spatial position control capabilities to cope with complex working environments, while traditional gripper control methods are often difficult to cope with changeable environments and diverse target objects. This paper proposes a deep learning-based claw position control model for hook-picking robots. Using Deep Convolutional Neural Network (DCNN) and reinforcement learning algorithms, a hand claw position control system with strong adaptability and high accuracy is constructed.In addition, the system also uses an improved layer descent algorithm to optimize the model parameters to improve the control accuracy and response speed. Experimental results show that the control system has excellent positioning accuracy and response speed in complex and changeable scenes, which significantly improves the operation performance and robustness of the hook-picking robot.