Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics

To reduce data collection time for deep learning of robust robotic grasp plans, we explore training from a synthetic dataset of 6.7 million point clouds, grasps, and robust analytic grasp metrics generated from thousands of 3D models from DexNet 1.0 in randomized poses on a table. We use the resulting dataset, Dex-Net 2.0, to train a Grasp Quality Convolutional Neural Network (GQ-CNN) model that rapidly classifies grasps as robust from depth images and the position, angle, and height of the gripper above a table. Experiments with over 1,000 trials on an ABB YuMi comparing grasp planning methods on singulated objects suggest that a GQ-CNN trained with only synthetic data from Dex-Net 2.0 can be used to plan grasps in 0.8s with a success rate of 93% on eight known objects with adversarial geometry and is 3× faster than registering point clouds to a precomputed dataset of objects and indexing grasps. The GQCNN is also the highest performing method on a dataset of ten novel household objects, with zero false positives out of 29 grasps classified as robust and a 1.5× higher success rate than a point cloud registration method.