Currently, there is a lot of pressure on agriculture to become more efficient, sustainable and to reduce the amount of menial labor. A partial solution can be provided by technology through advancing automation. The next leap in automation in agriculture will likely be the introduction of intelligent robotics. However, such systems have not reached the market yet, primarily because they fail to cope with all the natural variation they can encounter in the crops. For this thesis, it was researched which novel computer vision principles can help to cope with this crop variation. One approach was to create a flexible software framework to allow for fruit scanning and grasping with visual feedback during the motion. Furthermore, a machine learning based computer vision approach was researched to detect where plant parts in the image are located. Together, a new basis is provided for vision in agrobotics.