Download Link¶
Dex-Net HDF5 databases and datasets of 3D object models are available from our data repository. For GQ-CNN training datasets, see the gqcnn data repository New datasets and 3D models will be uploaded to this location as they become available.
Dex-Net Object Mesh Dataset v1.1¶
The Dex-Net Object Mesh Dataset v1.1 contains 1,500 3D object models in .OBJ format. The source of the models are: 1,371 synthetic models from the 50 category subset of 3DNet and 129 laser-scanned models from the KIT Object Database. The models were rescaled, rotated to align with the principal axes of the object, and centered on the center of hte object bounding box.
More details can be found in the Dex-Net 2.0 paper.
Dex-Net 2.0 HDF5 Database¶
The Dex-Net 2.0 HDF5 Database contains the Dex-Net Object Mesh Dataset v1.1 stored in HDF5 format. Each 3D model is labeled with up to 100 parallel-jaw grasps specified as an end-effector pose for an ABB YuMi custom gripper. Each grasp is labeled with force closure and the robust epsilon metric (also known as robust Ferrari-Canny) under uncertainty in object pose, gripper pose, and friction.
License¶
The datasets include 3D object mesh models from 3DNet and the KIT Object Database that may be subject to copyright. Please see the original datasets for more information.
The grasps and grasp metrics are available for unrestricted use.