Dex-Net 2.0

Berkeley's grasp-scoring network, trained on simulated data, that judges from a depth image which grasp will hold best.

Dex-Net is the grasping research project of Ken Goldberg's group (AUTOLAB) at UC Berkeley; its best-known version, Dex-Net 2.0, was published at RSS 2017. Instead of collecting real-robot data, it labels a huge collection of 3D object models using analytic grasp metrics (formulas from geometry and mechanics that score how stable a grasp is), synthesizing 6.7 million “point cloud + grasp + label” examples, then trains a convolutional network, GQ-CNN: given a depth image and a candidate grasp (planar position, angle, depth), it outputs a success probability, and the robot executes whichever candidate scores highest. It's an early, representative example of “train on synthetic simulated data, deploy directly to the real robot” for grasping; a later version, 3.0, extended this to suction grippers, and 4.0 (Science Robotics, 2019) uses both parallel-jaw grippers and suction together for bin picking.

ExampleDex-Net 2.0 used GQ-CNN to plan two-finger parallel-jaw grasps on an ABB YuMi robot: 93% success on 8 known objects, and about 99% precision for grasps judged “stable” on household objects it had never seen.

Also called

Dexterity Network, GQ-CNN, Grasp Quality CNN

Related

Grasping · Grasp Quality Metric · Bin Picking · Synthetic Data · Convolutional Neural Network · Grasp Pose Detection

Sources

Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics (arXiv:1703.09312)
Dex-Net 项目主页（Berkeley AUTOLAB） (Chinese)

As of

2019