Deep learning-driven industrial robot visual grasping system modelling and optimization

Driven by artificial intelligence and intelligent algorithms, in order to raise the grasping efficiency and accuracy of industrial robots in complex industrial scenarios, a modeling and optimization method for the visual grasping system of industrial robots driven by deep learning is studied and proposed. The target recognition model selects the You Only Look Once vision 5 (YOLOv5) framework, and optimizes the model performance through transfer learning and model pruning techniques. By using the principle of depth cameras and integrating the convolutional neural network with the pose estimation algorithm, a visual positioning and pose estimation model was formed. The results show that under normal illumination, optimized YOLOv5 has a recognition accuracy rate of 95.2% and a three-dimensional positioning error of 1.8mm. The A* algorithm shortens the path length by 20.2% and reduces the obstacle avoidance time by 34.2%. The success rate of deep learning-guided grasping is 5.5% higher than that of traditional methods. The self-adaptive stress control reduces the damage rate of rubber parts from 5% to 0.8%. Studies show that this system can adapt to complex lighting scenarios. The optimized algorithm effectively improves the grasping efficiency and success rate, which is conducive to promoting the intelligent development of industrial robots.