UniGraspAll: Efficient Grasp Detection with Object Detection for Generalized Objects

Embodied intelligence provides a novel paradigm for robotic grasping, enabling adaptive strategy refinement through environmental interaction, yet efficient grasp detection in cluttered multi-object scenes remains a pivotal unsolved challenge in industrial automation and service robotics, plagued by the trade-off between detection accuracy, real-time performance and computational efficiency in existing methods, and further constrained by poor adaptability to fast-decision multi-grasp scenarios. This paper addresses the core issues of task synergy and real-time inference in multi-object grasp detection by proposing UniGraspAll, an end-to-end network for effective multi-object grasp detection. Unlike computationally expensive 3D-based grasp detection methods that are confined to single-object scenarios, UniGraspAll leverages only 2D visual information to achieve efficient multi-object grasp detection, and it realizes joint learning of grasp detection and object recognition/localization for global scene perception, which simplifies the integration of object recognition and grasping into a single system and reduces the reliance on high-cost sensing equipment. We enrich the GraspNet-1Billion dataset by adding contour labels (annotated at 20-pixel intervals) and corresponding category IDs for each RGB image. Extensive experiments demonstrate that UniGraspAll maintains high detection accuracy while achieving efficient multi-object grasp prediction, outperforming state-of-the-art baseline methods, with real-time inference performance of 38 milliseconds per target and 242 milliseconds per frame, and it exhibits strong generalization for a wide range of target objects in practical multi-grasp scenarios requiring rapid decision-making. The code and dataset methodology are available on GitHub: https://github.com/SuperLuu7/GraspNet-1-Billion-dataset-enrichment