DOFS: A Real-world 3D Deformable Object Dataset with Full Spatial Information for Dynamics Model Learning

Robot manipulation of 3D Deformable Objects is essential for many activities and applications in the real world, such as household [1, 2] and healthcare [3], and is still an open challenge despite extensive studies. Recently, data-driven solutions have shown impressive and promising results in 3D deformable object manipulation by learning-based approaches [4, 5], where sufficient data is essential to improve model training or policy learning. To obtain training data, some previous works collected synthetic data from simulators [6]. Still, there is an unavoidable gap between the real world and the simulator since the existing simulators cannot accurately simulate all real-world physical characteristics (e.g., friction, impact, and stiffness) [7]. To mitigate the gap, some researchers [8, 9, 10, 11] collect Real-World Data (RWD); for example, [8, 9] collects RGB-D images and point clouds, [10] collects 3D mesh models, [11] uses a professional system with 106 cameras to obtain the 3D reconstructions of deformed mesh.