To contribute to the building of an automated textile steam sorting system, the main purpose of this master thesis research at the beginning was to automatically estimate, using computer vision, the 3D pick-up coordinates of an object. Once those goals were reached, in view to design a complete proof-of-concept, the system was extended to include an additional part : the transmission of the coordinates to an industrial ABB robot arm designed to pick-up the objects.

In summary, this thesis work which started from scratch, included 6 major steps:

1. The collection and labeling of the data using a zero-shot model (GroundingDINO)

2. The training of an object detection model (MaskRCNN Resnet)

3. The implementation of an instance segmentation model (SAM2)

4. The usage of a tracker over each instances (SORTTracker)

5. A distance estimation algorithm built on te usage ArUco markers, objectswhose dimensions are precisely known

6. The transmission of the coordinates via to Robot Web Service Rest API toRobot Studio, a software directly connected to the ABB Robot.

Eventually, a first automated robot prototype solution was created, successfully detecting, picking-up and removing clothes from a conveyor. While the aims of the thesis were met, showing the feasibility of such an undertaking, the research work will have to continue to pursue its final objective : Assisting co-workers by removing undesirable objects from conveyors. This implies labeling much more data in view to train new machine learning models with the capacity to detect and identify more categories of objects, keeping the desired ones identified with high enough confidence while removing the rest.