ACRONYM is a dataset of 17.7M simulated parallel-jaw grasps of 8872 objects. It was generated using NVIDIA FleX.

This repository contains a sample of the grasping dataset and tools to visualize grasps, generate random scenes, and render observations.

For using the full ACRONYM dataset, see instructions below.

License

The source code is released under MIT License. The dataset is released under CC BY-NC 4.0.

Requirements

• Python3
• python -m pip install -r requirements.txt

Installation

• python -m pip install -e .

Use Cases

Visualize Grasps

usage: acronym_visualize_grasps.py [-h] [--num_grasps NUM_GRASPS] input [input ...]

Visualize grasps from the dataset.

positional arguments:
  input                 HDF5 or JSON Grasp file(s).

optional arguments:
  -h, --help            show this help message and exit
  --num_grasps NUM_GRASPS
                        Number of grasps to show. (default: 20)
  --mesh_root MESH_ROOT
                        Directory used for loading meshes. (default: .)

Examples

The following command shows 40 grasps for a mug from the dataset. Grasp markers are colored green/red based on whether the simulation result was a success/failure:

acronym_visualize_grasps.py --mesh_root data/examples/ data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5

Generate Random Scenes and Visualize Grasps

usage: generate_scene.py [-h] [--objects OBJECTS [OBJECTS ...]] --support
                         SUPPORT [--support_scale SUPPORT_SCALE]
                         [--show_grasps]
                         [--num_grasps_per_object NUM_GRASPS_PER_OBJECT]

Generate a random scene arrangement and filtering grasps that are in
collision.

optional arguments:
  -h, --help            show this help message and exit
  --objects OBJECTS [OBJECTS ...]
                        HDF5 or JSON Object file(s). (default: None)
  --support SUPPORT     HDF5 or JSON File for support object. (default: None)
  --support_scale SUPPORT_SCALE
                        Scale factor of support mesh. (default: 0.025)
  --mesh_root MESH_ROOT
                        Directory used for loading meshes. (default: .)
  --show_grasps         Show all grasps that are not in collision. (default:
                        False)
  --num_grasps_per_object NUM_GRASPS_PER_OBJECT
                        Maximum number of grasps to show per object. (default:
                        20)

Examples

This will show a randomly generated scene with a table as a support mesh and four mugs placed on top of it:

acronym_generate_scene.py --mesh_root data/examples/ --objects data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 --support data/examples/grasps/Table_99cf659ae2fe4b87b72437fd995483b_0.009700376721042367.h5

Same as above but also showing green grasp markers (maximum: 20 per object) for successful grasps (filtering those that are in collision):

acronym_generate_scene.py --mesh_root data/examples/ --objects data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 --support data/examples/grasps/Table_99cf659ae2fe4b87b72437fd995483b_0.009700376721042367.h5 --show_grasps

Render and Visualize Observations

usage: render_observations.py [-h] [--objects OBJECTS [OBJECTS ...]] --support
                              SUPPORT [--support_scale SUPPORT_SCALE]
                              [--show_scene]

Render observations of a randomly generated scene.

optional arguments:
  -h, --help            show this help message and exit
  --objects OBJECTS [OBJECTS ...]
                        HDF5 or JSON Object file(s). (default: None)
  --support SUPPORT     HDF5 or JSON File for support object. (default: None)
  --support_scale SUPPORT_SCALE
                        Scale factor of support mesh. (default: 0.025)
  --mesh_root MESH_ROOT
                        Directory used for loading meshes. (default: .)
  --show_scene          Show the scene and camera pose from which observations
                        are rendered. (default: False)

Examples

This will show RGB image, depth, image and segmentation mask rendered from a random viewpoint):

acronym_render_observations.py --mesh_root data/examples/ --objects data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 --support data/examples/grasps/Table_99cf659ae2fe4b87b72437fd995483b_0.009700376721042367.h5

Same as above but also visualizes the scene and camera position in 3D:

acronym_render_observations.py --mesh_root data/examples/ --objects data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5 --support data/examples/grasps/Table_99cf659ae2fe4b87b72437fd995483b_0.009700376721042367.h5 --show_scene

Using the full ACRONYM dataset

1. Download the full dataset (1.6GB): acronym.tar.gz
2. Download the ShapeNetSem meshes from https://www.shapenet.org/
3. Create watertight versions of the downloaded meshes:
   1. Clone and build: https://github.com/hjwdzh/Manifold
   2. Create a watertight mesh version assuming the object path is model.obj: manifold model.obj temp.watertight.obj -s
   3. Simplify it: simplify -i temp.watertight.obj -o model.obj -m -r 0.02

For more details about the structure of the ACRONYM dataset see: https://sites.google.com/nvidia.com/graspdataset

Citation

If you use the dataset please cite:

@inproceedings{acronym2020,
    title     = {{ACRONYM}: A Large-Scale Grasp Dataset Based on Simulation},
    author    = {Eppner, Clemens and Mousavian, Arsalan and Fox, Dieter},
    year      = {2020},
    booktitle = {Under Review at ICRA 2021}
}