sea-bass / ycb-tools

Tools to download models from the YCB dataset and use them with the Gazebo simulator.

YCB Tools

Tools to download models from the YCB dataset and use them with the Gazebo simulator.

By Sebastian Castro, 2020-2021

Python Setup

You only need to install a few Python packages to get these tools running.

pip install numpy pillow scipy shapely trimesh

Downloading YCB objects

You can download the models using a variant of the download script provided on the YCB web page. However, this script has been modified to work with Python 3.

python download_ycb_dataset.py

You can configure a few options in the script, including choosing which objects and model types to download. However, the default options will get all of the YCB object models and may take a few minutes to download.

Using YCB Object Models in Gazebo

After you have downloaded the YCB models, you can run the following script.

python3 create_ycb_sdf.py

This by default uses the ./models/ycb and ./templates/ycb folder, assuming you are running the script from the top-level folder of this repo. You can always modify this if you want:

python3 create_ycb_sdf.py --template-folder /path/to/templates/ycb --ycb-folder /path/to/models.ycb

There is also a --downsample_ratio argument set in this script which uses rejection sampling to randomly remove faces from the mesh to get a subset of faces. This downsampled mesh is used only for collision detection, so it will not affect the visuals, but it should make simulation faster if you don't need perfect collision fidelity (i.e., if you're not trying to do manipulation). The Google 16k meshes are fairly high-resolution, so this should not have a huge impact, but if you want the full meshes used for collision, you can leave this parameter at its default value of 1.

python3 create_ycb_sdf.py --downsample-ratio 0.5

NOTE: A few of the models in the dataset do not have either of the google_16k or tsdf meshes available, so these will not work with Gazebo.

Finally, you need to make sure the folder containing your models is included in your GAZEBO_MODEL_PATH environment variable. We recommend setting this in your ~/.bashrc file as follows.

export GAZEBO_MODEL_PATH=$GAZEBO_MODEL_PATH:$REPO_BASE_FOLDER/models/ycb

where $REPO_BASE_FOLDER is the folder containing this README file.

ycb-tools's Issues

The {image, mask} pair

Hi there! Thanks for the code! I've used it and downloaded the dataset. However, I could not find the mask of the objects. Could you please give suggestion where I can get the data if I want to use this for a object segmentation? Many thanks!

Sliding in some objects

Hi, any idea on why the objects slide?, or better, how to stop it?

We been using this data-set for a project and implemented a similar approach using trimesh to generate the objects properties for the SDF like you did, but some of them start sliding or having some jittering in Gazebo and we haven't found out why.

(We used a different method for the collision models with Meshlab and its quadric edge collapse decimation process, and also got sliding problems.)

Also thanks for uploading this code, it was a pain for us when we had to do it.

conversion of ycb to ignition gazebo models (sdf format)

Hi there,

The conversion, if I understand it correctly, happens on the basis of the template file. If I change the file with another template supported by the ignition gazebo , will I get the compatible model that ignition gazebo supports?

Thanks

unable to load YCB texture map image to pybullet

I obtained object mesh, mtl files and texture map images from YCB dataset

However, I get this following error when loading it with pybullet:
b3Printf: Unsupported texture image format
any resolution for this?
Also, noting that using these works fine with blender.