ACRONYM: A Large-Scale Grasp Dataset Based on Simulation

An NVIDIA grasp dataset labeled in bulk through physics simulation, containing about 17.7 million parallel-jaw grasps.

ACRONYM was released in November 2020 by NVIDIA's Clemens Eppner, Arsalan Mousavian, and Dieter Fox. It selected 8,872 objects across 262 categories from the ShapeNetSem 3D model library, generated 2,000 candidate grasps per object using antipodal sampling (finding two opposing contact points on the surface), and then simulated each one in NVIDIA's FleX physics engine — closing a Franka Panda gripper and shaking the object to see whether it stayed held — yielding about 17.7 million labeled grasps, roughly 59% of them successful. Testing grasps one by one on a real robot is far too costly, so this kind of bulk simulated labeling gave learned grasping networks enough training data. The release also includes a tool for generating cluttered scenes with multiple objects randomly placed on support surfaces, capable of rendering depth images and point clouds.

ExampleNVIDIA's Contact-GraspNet, trained on about 17 million of these simulated grasps, exceeded a 90% success rate grasping previously unseen objects in real cluttered scenes.

Related

Grasping · Contact-GraspNet · Simulation Data · Antipodal Grasp · Grasp Pose Detection · ShapeNet

Sources

ACRONYM: A Large-Scale Grasp Dataset Based on Simulation (arXiv 2011.09584)
NVlabs/acronym (GitHub)
Contact-GraspNet (arXiv 2103.14127)

As of

2020-11