Hello!
Thank you very much for providing this dataset. I have a question.

Do we assume that the gripper width is always at the constant opening of 0.08 for each grasp? Because I cannot find in the provided files any array describing the width for each of the 2000 grasps per object. I found only /gripper/configuration, which is a scalar. Did I miss something?

Thank you in advance!

where to download the mesh files of the 8827 objects?
The dataset is comprised of 8836 HDF5 files， can not find the corresponding meshes....

When the ShapeNetSems are unzipped, there is only the Models folder,so I run

python scripts/acronym_generate_scene.py --mesh_root models/ --objects grasps/1Shelves_1e3df0ab57e8ca8587f357007f9e75d1_0.011099225885734912.h5 grasps/1Shelves_1e3df0ab57e8ca8587f357007f9e75d1_0.011099225885734912.h5 --support data/examples/grasps/Table_99cf659ae2fe4b87b72437fd995483b_0.009700376721042367.h5 --show_grasps

than get

Traceback (most recent call last):
File "/home/user/acronym/scripts/acronym_generate_scene.py", line 123, in
main()
File "/home/user/acronym/scripts/acronym_generate_scene.py", line 73, in main
object_meshes = [load_mesh(o, mesh_root_dir=args.mesh_root) for o in args.objects]
File "/home/user/acronym/scripts/acronym_generate_scene.py", line 73, in
object_meshes = [load_mesh(o, mesh_root_dir=args.mesh_root) for o in args.objects]
File "/home/user/acronym/acronym_tools/acronym.py", line 375, in load_mesh
obj_mesh = trimesh.load(os.path.join(mesh_root_dir, mesh_fname))
File "/home/user/anaconda3/envs/acr/lib/python3.9/site-packages/trimesh/exchange/load.py", line 110, in load
) = parse_file_args(file_obj=file_obj,
File "/home/user/anaconda3/envs/acr/lib/python3.9/site-packages/trimesh/exchange/load.py", line 605, in parse_file_args
raise ValueError('string is not a file: {}'.format(file_obj))
ValueError: string is not a file: models/meshes/1Shelves/1e3df0ab57e8ca8587f357007f9e75d1.obj

Was there a step I missed？

Hello

I am a big fan of your work! Thank you for releasing your code.

For my research project, I would like to generate grasp annotations for my own mesh.

Particularly, I wonder to which reference frame is the orientation and position of the grasp annotation.

If I were to use it for real world grasping by training some NN model like ContactGraspNet or M2T2, should it be with respect to the robot base frame? or should it be with respect to the camera frame?

Thank you so much in advance for your help and guidance.

I would like to ask about the "grasps/h5" h5 format file, what tools can be quickly generated?

NVlabs/contact_graspnet: Efficient 6-DoF Grasp Generation in Cluttered Scenes
https://github.com/NVlabs/contact_graspnet

I am studying this project, the data set format needs h5 file for training, so I do not know how to generate h5 file at present

Hi,

Thank you for your excellent work!
I would like to generate some physically feasible grasps for our own objects. Could you give some hints for the implementation based on Nvidia Flex? Thank you

Beat,
Rui

Hi,
I have some questions regarding the following fields in each of the grasp files:

/grasps/qualities/flex/object_in_gripper Dataset {2000}

/grasps/qualities/flex/object_motion_during_closing_angular Dataset {2000}

/grasps/qualities/flex/object_motion_during_closing_linear Dataset {2000}

/grasps/qualities/flex/object_motion_during_shaking_angular Dataset {2000}

/grasps/qualities/flex/object_motion_during_shaking_linear Dataset {2000}

For the 'object_in_gripper', 'object_motion_during_closing_angular', and 'object_motion_during_closing_linear' fields, why are they single array of 2000 numbers instead of something like a 2000 SE3 matrices? How can these numbers be translated to describe the full position and orientations of a 3D object?

For the 'object_motion_during_shaking_angular' and the 'object_motion_during_shaking_linear' field, are these the rotation in radian and mm as described in the paper during shaking?

I understand that maintaining a dataset is a lot of work. I really appreciate this open-source dataset and the effort you put into it. Any response would be appreciated.

Hey :)

I am encountering an error when loading meshes from the h5 files:

line 371, in load_mesh
    mesh_fname = data["object/file"][()].decode('utf-8')
AttributeError: 'str' object has no attribute 'decode'

In Python3, strings are already unicode objects, so dropping .decode('utf-8') resolves the error for me.

Can you give some better instructions regarding the mesh creation with manifold?
I am following your steps and leads to segmentation fail.
Is the shapenetsem file models-OBJ.zip the one needed or am I doing sth wrong?

Why is the cup under the table？

Hi. In this dataset, what does the "volume" represent for?
The reason why I asked this question is because there are some objects have negative volume. For example,
Table_2e7a2d47040a0d4e35836c728d324152_0.009612910954917283.h5 with volume: -6.49676e-05
Barstool_d4fd2dd6e3407f0f5a2aa419fded589d_0.006393724755823298.h5 with volume: -5.2156e-06
Bench_6f5cb68e41ab4356ad5067eac75a07f7_0.011921571439527373.h5 with volume: -2.68208e-05
Basket_b06aa40e972e1de86d431f3195ee60_0.003267646646604986.h5 with volume: -1.41226e-05

This is more of a feature request. It seems like some people are struggling to convert the original meshes even with the instructions in the README, including myself (besides some errors for some meshes, my conversion script gets stuck for some reason).
Would be nice if there can be a bash script to reproduce the conversion process you guys did.

This is the script I'm using:

for fn in ~/data/shapenetsem/models-OBJ/models/*.obj
do
  rm temp.watertight.obj

  timeout -sKILL 10 ./manifold "$fn" temp.watertight.obj -s
  
  if [ $? != 0 ]; then
    echo "skipping object"
  else
    out="$(basename $fn)"
    ./simplify -i temp.watertight.obj -o ~/data/shapenetsem/models-OBJ/simplified/$out -m -r 0.02
  fi
done

Hey,
I have an issue with the scene generation. Everything seems to work fine, except the valid grasps are not displayed when generating a scene(--show_grasps is set). I am working with Ubuntu 20.04 and conda environment. Do you have an idea why it is not working? Thanks!

In section Render and Visualize Observations, the second example should start with acronym_render_observations.py instead of acroacronym_render_observations.py.

Hi,

I am trying to access https://www.shapenet.org/ but seems like the server is down because there is no response. Is there a way around this so that I can get access to the ShapeNetSem dataset?

Thank you.

I wanted to try your examples, but ran into a few problems:
(i am on Ubuntu 18 using Anaconda)

The error occurs when running your example for acronym_visualize_grasps.py

$ acronym_visualize_grasps.py --mesh_root data/examples/ data/examples/grasps/Mug_10f6e09036350e92b3f21f1137c3c347_0.0002682457830986903.h5
Traceback (most recent call last):
  File "/tmp/acronym-env/bin/acronym_visualize_grasps.py", line 7, in <module>
    exec(compile(f.read(), __file__, 'exec'))
  File "/tmp/acronym/scripts/acronym_visualize_grasps.py", line 31, in <module>
    from acronym_tools import load_mesh, load_grasps, create_gripper_marker
  File "/tmp/acronym/acronym_tools/__init__.py", line 24, in <module>
    from .acronym import *
  File "/tmp/acronym/acronym_tools/acronym.py", line 28, in <module>
    import trimesh.path
  File "/tmp/acronym-env/lib/python3.7/site-packages/trimesh/path/__init__.py", line 9, in <module>
    from .path import Path2D, Path3D
  File "/tmp/acronym-env/lib/python3.7/site-packages/trimesh/path/path.py", line 34, in <module>
    from . import polygons
  File "/tmp/acronym-env/lib/python3.7/site-packages/trimesh/path/polygons.py", line 3, in <module>
    from shapely import ops
ModuleNotFoundError: No module named 'shapely'

it seems shapely is missing from your requirements.

After installing it with pip install shapely I get this error:

Traceback (most recent call last):
  File "/tmp/acronym-env/bin/acronym_visualize_grasps.py", line 7, in <module>
    exec(compile(f.read(), __file__, 'exec'))
  File "/tmp/acronym/scripts/acronym_visualize_grasps.py", line 74, in <module>
    main()
  File "/tmp/acronym/scripts/acronym_visualize_grasps.py", line 55, in main
    obj_mesh = load_mesh(f, mesh_root_dir=args.mesh_root)
  File "/tmp/acronym/acronym_tools/acronym.py", line 375, in load_mesh
    obj_mesh = trimesh.load(os.path.join(mesh_root_dir, mesh_fname))
  File "/tmp/acronym-env/lib/python3.7/posixpath.py", line 94, in join
    genericpath._check_arg_types('join', a, *p)
  File "/tmp/acronym-env/lib/python3.7/genericpath.py", line 155, in _check_arg_types
    raise TypeError("Can't mix strings and bytes in path components") from None
TypeError: Can't mix strings and bytes in path components

I get similar errors for your examples for acronym_generate_scene.py and acronym_render_observations.py

Hi, thanks so much for the fantastic dataset. It has clean implementations and diverse training samples.

I'm using Acronym for a grasp planning project with offline reinforcement learning. I'm wondering whether it's possible to open-source the Flex implementation for the data generation? I noticed there is a repo for deformable objects, but using the original implementation may be helpful to avoid the covariant shift issue.