Research on Robot Grasping Based on Deep Learning for Real-Life Scenarios

Abstract

The application of robots, especially robotic arms, has been primarily focused on the industrial sector due to their relatively low level of intelligence. However, the rapid development of deep learning has provided a powerful tool for conducting research on highly intelligent robots, thereby offering tremendous potential for the application of robotic arms in daily life scenarios. This paper investigates multi-object grasping in real-life scenarios. We first analyzed and improved the structural advantages and disadvantages of convolutional neural networks and residual networks from a theoretical perspective. We then constructed a hybrid grasping strategy prediction model, combining both networks for predicting multi-object grasping strategies. Finally, we deployed the trained model in the robot control system to validate its performance. The results demonstrate that both the model prediction accuracy and the success rate of robot grasping achieved by this study are leading in terms of performance.