Estimation of Image-Based End-Effector Approach Angles for Tomato Harvesting Robots

We propose a method to estimate a suitable approach angle for the end-effector of a tomato harvesting robot based on image data.Agricultural harvesting robots often face obstacles such as other fruits or stems around the target crop.Additionally, it is important to approach the target from a direction appropriate for harvesting, considering the shape of the end-effector.The proposed method uses a deep learning-based instance segmentation model to extract regions of fruits and stems, and estimates the suitable approach angle based on their positional relationships.We demonstrated the usefulness of the proposed method using an image dataset acquired in an actual tomato greenhouse.