A Model for Robot Grasping: Integrating Transformer and CNN With RGB-D Fusion

In recent years, the continuous development of robotics and artificial intelligence technology, with the robot application is also more and more advanced, especially the robot grasping task, but at present it is difficult to take into account the grasping accuracy and runtime, so we propose a model through the following design can be used in robot grasping task has excellent performance. The RCrossFormer module, which can establish temporal relationships in multiple levels and dimensions through Multi-level Long Short Distance Attention, and the Swin Transformer are combined to form a backbone, and together with the CNN-based module, we form a model that can better capture global and local features, while achieving a lightweight model. Features while realising the advantages of model lightweight; Moreover, the RGB-D Fushion can be implemented in multi-scale and multi-stage by our proposed tiny Residual Feature Fushion module, which can improve the performance of grasping detection. Experiments show that the detection accuracies in the public Jacquard and Cornell datasets are 96.6% and 99.3%, respectively, with high detection accuracy, and the real-world grasping experiments also have good results in predicting the bit position.