GraspFactory: A Large Object-Centric Grasping Dataset

Large datasets have been a major contributor to the success of AI models. The fields of Computer Vision and Natural Language Processing have seen tremendous progress due to the presence of internet-scale datasets like ImageNet [1] and Laion-5b [2]. Models such as Chat-GPT [3] and Dall-E[4] demonstrate strong generalization capabilities for tasks that were not explicitly represented in their training data, thanks to the use of diverse training datasets and large-scale transformer-based architectures. Similar efforts have been undertaken in robotics to collect large datasets, such as Open X-Embodiment [5] and DROID [6]. These datasets focus on end-to-end training of robots but there is still a need for task-specific datasets. Robot grasping is one such task, and a generalized grasping model remains elusive, in part due to the lack of geometrically diverse objects in existing datasets. In this work, we present an object-centric grasping dataset that offers greater geometric diversity compared to existing datasets. Currently, object-centric grasping datasets [7, 8, 9] and scene-based grasping datasets [10, 11, 12] are mostly geared toward domestic robotics applications. These datasets have been used to train robot grasping models such as [13, 14, 15, 16].