*Equal contribution, 1University of California, Los Angeles, 2Toyota Research Institute, 3The University of Utah
Robotic grasping is fundamental to manipulation tasks, with soft grippers gaining increasing attention for their compliance and adaptability. However, existing grasp datasets predominantly focus on rigid grippers and manipulants, lacking data for soft-rigid interactions. This limitation hinders the development of generalizable grasp prediction models and constrains their applicability in real-world scenarios. To bridge this gap, we introduce a fully automated generative pipeline for grasp synthesis that supports both rigid and soft grippers as well as manipulants. At the core of this pipeline is a high-performance, parallelized multi-environment simulator designed to ensure accurate soft-rigid coupled dynamics and physically realistic grasping poses. Leveraging this pipeline, we construct a large-scale grasping dataset GRIP comprising 1,200 objects and 100K grasp poses, annotated with manipulant deformation and stress distributions. GRIP includes diverse grasping configurations, accommodating rigid and soft grippers as well as manipulants under both unimanual and bimanual grasp settings. Extensive experiments are conducted to validate our dataset, including neural grasp generation and stress field prediction, and the results show high consistency with real-world experiments, demonstrating the effectiveness and applicability of our dataset.