Irregular Depth Tiles: Automatically Generated Data Used for Network-based Robotic Grasping in 2D Dense Clutter

Recent advances in deep learning have enabled robots to grasp objects even in complex environments. However, a large amount of data is required to train the deep-learning network, which leads to a high cost in acquiring the learning data owing to the use of an actual robot or simulator. This paper presents a new form of grasp data that can be generated automatically to minimize the data-collection cost. The depth image is converted into simplified grasp data called an irregular depth tile that can be used to estimate the optimal grasp pose. Additionally, we propose a new grasping algorithm that employs different methods according to the amount of free space in the bounding box of the target object. This algorithm exhibited a significantly higher success rate than the existing grasping methods in grasping experiments in complex environments.