Diffusion Model-Based Directional Target Detection for Robotic Sorting Task

In the field of industrial automation, the efficiency and accuracy of robotic arm sorting tasks are crucial for increasing productivity. The core contribution of this study is the proposal of a new method, DiffDDet (Diffusion Modelbased Directional Target Detection), which not only achieves target detection but, more importantly, provides directional information of the targets, which is lacking in traditional target detection technologies. DiffDDet improves upon the DiffusionDet target detection method by outputting directional information of targets alongside bounding boxes and target categories, and by considering the cyclic nature of directions, it improves the computation method of the sigmoid focal loss function, making it better adapted to learning directions. Furthermore, to further enhance the efficiency of sorting path planning, we have improved the traditional genetic algorithm and developed a new path planning algorithm. This algorithm optimizes genetic operations by intervening in the initial generation of the population with the nearest neighbor algorithm, significantly improving the speed and adaptability of path planning, making it particularly suitable for robotic arm sorting tasks. The experimental results of this study demonstrate that the advantages of DiffDDet in providing target directional information, combined with the improved genetic algorithm path planning, can significantly enhance the overall performance and efficiency of robotic arm sorting operations.