Home‎ > ‎

Push Dataset

he data here is from the following publication:
Unsupervised Learning for Physical Interaction through Video Prediction
NIPS, 2016. [arXiv] [PDF]

This data set contains roughly 59,000 examples of robot pushing motions, including one training set (train) and two test sets of previously seen (testseen) and unseen (testnovel) objects.

  • Feature descriptions and counts -- This sheet describes the meanings of the features in the push dataset and lists the number of examples with each feature.

The Example protobufs are stored in Tensorflow RecordWriter format.  The files are named:

The script download_listing.sh linked at the bottom of the page will use curl to download all (137GB) of the data files.
$ ./download_listing.sh push_listing.txt ./push_dataset

Convert Images to GIFs

The script, grab_train_images.py, can be used to convert the images from each training example in the dataset into an animated gif.

The code assumes that the training data is in the directory "push/push_train/".

It requires TensorFlow (v0.11) and moviepy.  Later versions of TensorFlow should require minimal modifications, if any.  The moviepy package can be installed with 'sudo pip install moviepy'.

The script can be run via 'python grab_train_images.py'.

Some example gifs that the script produces are:

License: This data is licensed by Google Inc. under a Creative Commons Attribution 4.0 International License.

Laura Downs,
Sep 2, 2016, 11:36 AM
Laura Downs,
Mar 16, 2017, 10:46 AM
Laura Downs,
Aug 11, 2016, 11:28 AM