SYN-PBOX: A large-scale benchmark dataset of box-shaped objects for scene understanding of bin picking

Robotic grasping and sorting tasks require a large amount of comprehensive and high-quality data, but collecting real-world data is too expensive and time-consuming in practice. To address this issue and bridge the domain gap when training with synthetic datasets for specific scenarios, this paper utilizes domain randomization (DR) and its enhanced techniques, and generates synthetic dataset of box-shaped objects with complex scenarios. Our dataset contains 15,000 realistic images (including: point clouds, RGB images, depth images) of 21 box-shaped objects and detailed annotation data indifferent poses and cluttered conditions (including: 6D pose, segmentation mask, instance segmentation, and instance segmentation of the reasonable grasping surface for each object). Experiments demonstrate the superior performance of our dataset in segmentation, pose estimation, grasping, and its ability to be effectively applied to real-world scenarios. Furthermore, we propose a deep neural network called YOLOv8-D, which can achieve a grasping success rate of over 95% on previously unseen real objects after training on our dataset.