3D PERCEPTION OF DEFORMABLE OBJECTS IN ROBOTIC APPLICATION

The manipulation of deformable objects, particularly garments, remains a significant challenge in robotics due to their infinite-dimensional state space and unpredictable dynamics. This thesis addresses the critical need for robust perception systems by designing, implementing, and evaluating an end-to-end vision pipeline for autonomous robotic handling of clothing. The research is contextualized within the VolPix project from EUROBIN, which targets the automation of laundry tasks involving ten distinct garment categories in both wet and dry states. The proposed system employs a multi-stage approach to transform a single RGB image into actionable data for a robotic manipulator. The pipeline begins with instance segmentation to isolate individual garments from cluttered scenes, followed by object recognition to determine each item's category. Subsequently, a specialized keypoint detection module localizes semantic landmarks crucial for grasping and folding, and a final stage, reconstructs the garment's 3D mesh using a monocular depth estimation technique. To train and validate these components, a custom dataset was collected and annotated, supplementing pretraining on the large-scale DeepFashion2 dataset. This work establishes a comprehensive perception framework that integrates segmentation, recognition, keypoint detection, and 3D reconstruction, providing a strong foundation for advancing autonomous robotic manipulation of deformable objects.