Innovative Artificial Intelligence / Machine Learning (AI/ML) solutions for smart harvesting of high value crops

This thesis explores the integration of artificial intelligence (AI)-based perception and low-cost robotic manipulation for the autonomous harvesting of underexplored high-value crops. Two case studies, edible flowers and blackberries, were investigated to evaluate a modular pipeline combining detection, segmentation, pose estimation, and robotic manipulation. For edible flowers, the proposed framework, FloralAI, combined YOLOv5 detection with the Segment Anything Model (SAM) for zero-shot segmentation and semi-automatic annotation. Pose estimation was performed using principal component analysis (PCA), while a novel plucking-point estimation method was introduced based on inferred flower diameter. This strategy reduced reliance on large annotated datasets and enabled generalization across different flower species. For blackberries, 3D plucking-point estimation was achieved through ellipsoid fitting combined with PCA-based pose estimation, enabling accurate localization and orientation of berries within dense and irregular canopies. This perception pipeline was coupled with a custom soft robotic gripper and a low-cost 6-DOF arm, leading to the development of the first fully evaluated autonomous blackberry-harvesting prototype. Task-based experiments confirmed successful autonomous picking across berries with diverse orientations and positions, establishing replicable benchmarks for future research in agricultural robotics. The results highlight that while detection models exhibit moderate cross-crop transferability, pose estimation and manipulation strategies require crop-specific adaptations. The findings also emphasize the interdisciplinary nature of agricultural automation: advances in robotics and AI must be complemented by crop research, including plant architecture, growth patterns, and cultivation methods, to enable scalable robotic harvesting. Overall, this work demonstrates the feasibility of affordable and adaptable robotic harvesting systems. It outlines key future directions, including large-scale dataset generation, multi-modal sensing, real-world deployment, and crop–robot co-design, thereby contributing to the foundation for more sustainable and resilient food production systems.