ACRONYM: A Large-Scale Grasp Dataset Based on Simulation

Clemens Eppner, Arsalan Mousavian, Dieter Fox

Seminal November 2020

TL;DR A dataset of 17.7M parallel-jaw grasps on 8872 ShapeNetSem objects from 262 categories, each labeled by closing and shaking the gripper in the FleX physics simulator. Retraining 6-DOF GraspNet and GPD on it, with unchanged models, raises their grasp performance 2-3 fold over training on a smaller dataset.

What they do

ACRONYM is a dataset for learning 6-DOF parallel-jaw grasping. It labels 2000 grasps on each of 8872 ShapeNetSem objects with the outcome of a physics simulation instead of an analytical quality measure, and includes a generator for cluttered multi-object scenes. To show its value, the authors retrain two existing grasp planners on it without changing the models.

How it works

Objects. Single-component ShapeNetSem meshes, made watertight, with uniform density and friction 1.0, scaled so the longest bounding-box side is uniformly between 6 and 35 cm.
Grasp sampling. Antipodal sampling with a Franka Panda gripper model at its 8 cm maximum aperture. Rejection sampling keeps 2000 proposals per object that are collision-free and have the object between the fingers.
Labeling. In FleX without gravity, the fingers close until a force threshold, then the hand shakes along its approach direction and rotates. A grasp succeeds if both fingers still touch the object. Compared with labels obtained under gravity, shaking has 91% precision and 93% recall; 59.21% of the 17.744M grasps succeed.
Clutter. Objects are placed on support surfaces of furniture meshes; single-object labels are reused, with grasps that collide with the scene counted as failures. Images are not shipped but rendered with provided pyrender code.

Results

On ACRONYM’s test objects, area under the success-coverage curve rises from 0.08 to 0.21 (6-DOF GraspNet, c-VAE), 0.14 to 0.26 (c-IMLE) and 0.02 to 0.04 (GPD) when training on ACRONYM instead of the smaller 6-DOF GraspNet dataset.
On unseen instances of the smaller dataset’s own categories, ACRONYM-trained models still score higher: 0.29 vs 0.11 (c-VAE) and 0.41 vs 0.31 (c-IMLE), although ACRONYM training contained no boxes or cylinders.
Every method scores lower on ACRONYM than on the smaller dataset, which the authors read as headroom for a long-term benchmark.

Limitations

Labels come purely from simulation. The paper argues for real-world transfer by citing prior work (FleX reached 86% average precision over 2625 real grasps) but runs no robot experiments of its own. Only the Panda parallel-jaw gripper is modeled, and all objects share one density and friction. Clutter labels reuse single-object results and only remove colliding grasps, which the authors note can produce false positives and negatives.

Figures

From Figure 1: ACRONYM contains 2000 parallel-jaw grasps for each of 8872 objects from 262 categories, 17.7M grasps i... — From Figure 1: ACRONYM contains 2000 parallel-jaw grasps for each of 8872 objects from 262 categories, 17.7M grasps in total.

From Figure 4: grasps on a mug that succeeded in simulation (green markers), showing 15% of the 2000 labels for the o... — From Figure 4: grasps on a mug that succeeded in simulation (green markers), showing 15% of the 2000 labels for the object.

From Figure 5: a randomly generated scene with structured clutter, objects clustered on the support surface of a furn... — From Figure 5: a randomly generated scene with structured clutter, objects clustered on the support surface of a furniture mesh. Colors are for visualization only.

From Figure 6: success-coverage curves on the ACRONYM test objects, with area under the curve in the legend. Every mo... — From Figure 6: success-coverage curves on the ACRONYM test objects, with area under the curve in the legend. Every model scores higher when trained on ACRONYM than when trained on the smaller 6-DOF GraspNet dataset (labeled [23] in the legend).

Abstract

We introduce ACRONYM, a dataset for robot grasp planning based on physics simulation. The dataset contains 17.7M parallel-jaw grasps, spanning 8872 objects from 262 different categories, each labeled with the grasp result obtained from a physics simulator. We show the value of this large and diverse dataset by using it to train two state-of-the-art learning-based grasp planning algorithms. Grasp performance improves significantly when compared to the original smaller dataset. Data and tools can be accessed at https://sites.google.com/nvidia.com/graspdataset.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong