Researchers led by Jiarui Guo have built RFW-S, a dataset of 83,365 deformable objects for robot simulation, each with physical properties that have been checked, according to a paper on arXiv. They cover 485 everyday categories in five families: soft solids, cloth sheets, shells, garments and bags.
An automated pipeline generates the 3D shapes from reference images, models their structure, and estimates and calibrates their physical parameters with help from a large language model. A separate check then tests each object in interactions that differ from those used for calibration.
The team evaluated robot grasping on the objects, for both success and whether the object stayed intact, and replayed the handling of fragile objects on a real robot to show that the simulated physics is realistic.



