We propose a new large-scale database containing grasps that are applied to a large set of objects from numerous categories.
These grasps are generated in simulation and are annotated with the standard epsilon-metric and a new physics-metric.
We use a descriptive and efficient representation of the local object shape at which the grasp is applied.
Each grasp is annotated with the proposed metrics and representation.
We use crowdsourcing to analyze the correlation of the two metrics with grasp success as predicted by humans.
The results confirm that the proposed physics-metric is a more consistent
predictor for grasp success than the epsilon-metric.
Furthermore it supports the hypothesis that human labels are not required for good ground truth grasp data.
Instead the physics-metric can be used for simulation data.
We hope that this database will increase over time and the community will contribute back. Therefore, we encourage people to subscribe to the database information service. We will send notification mails as soon as new data, code changes, or other changes related to the database are available.
Every database is stored in
the HDF5 file
format. There are existing tools to manually inspect the
database, e.g. vitables
shown on the right. In the software package we provide
scripts to obtain the raw HDF5 files of different versions
of the database.
To have a good user experience we decided to provide a
docker image which has been tested succesfully on Ubuntu
12.04/14.04 and Mac. The image ships with a pre-compiled
software package to visualize grasps stored in the
database and also store the grasp templates.
All dependencies to compile our code are described in the
docker build description, making it transparent to the
user to build our code outside of this environment.
The source code is also provided and explains how to
acquire data from the hdf5 file using our python library.
The grasp database consists of 87 categories. In the following some example objects for the 3 groups, small, medium, and large objects are shown. We want to stress that there are more models available in our database. The subset shown in the following is chosen to illustrate the variaty of models we use.
For each experiment we had pre-labeled ground truth images
as shown in the following. Ground Truth Positive and
Ground Truth Negative are examples shown to the user
during the whole experiment as illustrated on the right
hand side picture. Please click on the right hand sight
image to get to an example mechanical turk webpage.
The Ground Truth Reject pictures are used to block
mechanical turk workers, to be able to get consistent
data. Please click on the corresponding images to download
all pre-labeled images.