Modeling variation for learning-based and multi-task robots in fruit picking

Modern agriculture suffers from labor shortages for delicate tasks like fruit harvesting. This thesis worked on several challenges facing selective harvesting robotics. To reduce the manual annotation effort, 3D scene reconstruction was used for pose estimation, reducing the effort by 99.6%, while maintaining high pose estimation accuracy. To optimize the fruit detachment method, multiple algorithms were implemented to learn the fruit detachment motion from expert demonstrations. In a lab environment, a 67% success rate was achieved. Next, the challenge of dealing with wind disturbances was studied. A novel algorithm was developed, which could achieve a 99% success rate. Lastly, the multi-tasking capabilities were studied in a real orchard. Apple harvesting was performed best by training with 20 demonstrations and resulted in 90% of the fruits grasped and 65% picked. Pear harvesting was performed best by training with 40 demonstrations and resulted in 65% of the fruits grasped and 50% picked.