Hang Yu

Generating reliable grasp configurations in cluttered scenes is an important guarantee for robots to successfully grasp objects. Existing learning-based methods encode the voxelized scene into a latent space to predict graspable points, which overlook both the reliability of grasp selection and the explicit geometric constraint imposed by the object’s surface occupancy. In this article, we propose a 3-D sparse grasp network (3-D SPGNet), which maps voxel features of object surfaces to grasp candidates to enforce explicit constraints, while leveraging the dense truncated signed distance function (TSDF) reconstruction to enhance grasp reliability through implicit synergy. Specifically, we first use a 3-D convolutional neural network to extract multiscale features at three different resolutions from the voxelized input. Then, we apply the proposed 3-D surface constraint block to aggregate features and perform explicit geometric mapping on the encoded nonempty voxels. Finally, we reconstruct the scene from sparse to dense and generate the grasp configuration accordingly. Moreover, we propose a novel loss function for estimating reliable grasp scores and poses. Simulation experiments demonstrate that our method improves the grasp success rates and declutter rates by approximately 9% compared to state-of-the-art baselines. In addition, we deploy the model on a real robot arm, and the real-world experimental results show that our method achieves over a 10% improvement in performance, which further verifies the effectiveness of the proposed method.

Although deep learning models may achieve successful grasps in some instances, they often struggle to accurately reflect the true likelihood of success for a given grasp. In this article, we introduce the trustworthy robotic grasping (TRG) problem, aiming to bridge the gap between predicted grasp probabilities and actual grasp success rates. We propose a novel credibility alignment framework through a two-branch network architecture. This architecture generates an adjusting tensor for nonprobabilistic outputs prior to the activation function of the backbone model, which is able to scale the output proportionally to improve the reliability of the predicted probability. To learn the adjusting tensor, a novel self-regulation encoder has been designed, which can extract 3-D local features of the scene for the local associative regulation of nonprobabilistic outputs. To facilitate research in this area, a new TRG dataset has been created. Experimental results reveal that our method not only significantly reduces the expected grasp error (EGE), maximum grasp error (MGE), and latter half expected grasp error (LH-EGE) by up to 50% compared to the precredibility alignment state, but also enhances the grasp success and declutter rates. Real-world experiments further validate the efficacy of our method.