Search papers, labs, and topics across Lattice.
The paper introduces DGS-Net, an end-to-end grasp prediction network that learns dense grasp configurations from single-view point clouds in multi-object scenes. To train this network, the authors created a large-scale dataset of 307 objects, 240 multi-object scenes, and over 350k validated grasps, generated using a novel two-stage synthesis strategy that incorporates grasp offsets and pre-grasp configurations. Experiments demonstrate that DGS-Net achieves high grasp success rates (88.63% in simulation, 78.98% on a real robot) and low penetration compared to existing methods, highlighting its effectiveness and generalization.
Forget single-object grasping – this new dataset and network lets robots learn dexterous grasps directly from cluttered, real-world-like scenes.
Dexterous grasping in multi-object scene constitutes a fundamental challenge in robotic manipulation. Current mainstream grasping datasets predominantly focus on single-object scenarios and predefined grasp configurations, often neglecting environmental interference and the modeling of dexterous pre-grasp gesture, thereby limiting their generalizability in real-world applications. To address this, we propose DGS-Net, an end-to-end grasp prediction network capable of learning dense grasp configurations from single-view point clouds in multi-object scene. Furthermore, we propose a two-stage grasp data generation strategy that progresses from dense single-object grasp synthesis to dense scene-level grasp generation. Our dataset comprises 307 objects, 240 multi-object scenes, and over 350k validated grasps. By explicitly modeling grasp offsets and pre-grasp configurations, the dataset provides more robust and accurate supervision for dexterous grasp learning. Experimental results show that DGS-Net achieves grasp success rates of 88.63\% in simulation and 78.98\% on a real robotic platform, while exhibiting lower penetration with a mean penetration depth of 0.375 mm and penetration volume of 559.45 mm^3, outperforming existing methods and demonstrating strong effectiveness and generalization capability. Our dataset is available at https://github.com/4taotao8/DGS-Net.
