Bionic soft hands have attracted extensive attention due to their high flexibility, safety, and adaptability, enabling them to mimic human hand movements such as grasping. This paper presents the design a pneumatic-driven sorting modular soft hand based on visual object detection. Additionally, a comprehensive sorting system control platform was constructed, integrating vision sensor, the modular soft hand, object detection, drive units, and mobile sorting strategies. To verify the system's performance, we established datasets for three types of objects: Lightweight building blocks, fragile eggs, and fresh oranges, and utilized the YOLOv8 algorithm to train the object detection model, thus accurately identifying the type of objects to be sorted and the three-dimensional coordinates of their grasping points. The experimental results demonstrate that this control platform is efficient, stable, and non-destructive in both single-target and multi-target mixed sorting, fully verifying the system's practicality and reliability.