Deep Differentiable Grasp Planner for High-DOF Grippers


Abstract

We present an end-to-end algorithm for training deep neural networks to grasp novel objects. Our algorithm builds all the essential components of a grasping system using a forward-backward automatic differentiation approach, including the forward kinematics of the gripper, the collision between the gripper and the target object, and the metric of grasp poses. In particular, we show that a generalized Q_1 grasp metric is defined and differentiable for inexact grasps generated by a neural network, and the derivatives of our generalized Q_1 metric can be computed from a sensitivity analysis of the induced optimization problem. We show that the derivatives of the (self-)collision terms can be efficiently computed from a watertight triangle mesh of low-quality. Put together, our algorithm allows the computation of grasp poses for high-DOF grippers in unsupervised mode with no ground truth data or improves the results in supervised mode using a small dataset. Our new learning algorithm significantly simplifies the data preparation for learning-based grasping systems and leads to higher qualities of learned grasps on common 3D shape datasets, achieving 22% higher success rate on physical hardware and 0.12 higher value of the Q_1 grasp quality metric.

Video

Paper

Deep Differentiable Grasp Planner for High-DOF Grippers, RSS 2020.
Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha.

Supplemental Material

Supplemental material can be found here, just after the regular paper.

Dataset

High-DOF grasping dataset can be downloaded here. To use this dataset, you should install GraspIt! and change the paths in .xml file to point to shapes path.

Erratum

We have updated the gradient equation of normal at a point when the closest feature to this point is an edge, which can be found in Subsection C of Section IV in our arXiv version.

@inproceeding{liu2020diffgrasping,
  title={Deep Differentiable Grasp Planner for High-DOF Grippers},
  author={Liu, Min and Pan, Zherong and Xu, Kai and Ganguly, Kanishka and Manocha, Dinesh},
  booktile={2020 Robotics: Science and Systems (RSS)},
  year={2020}
}