sketchy

Sketchy data

This is a dataset accompanying the paper Scaling data-driven robotics with reward sketching and batch reinforcement learning. If you use this dataset in your research please cite

@article{cabi2019,
  title={Scaling data-driven robotics with reward sketching and batch reinforcement learning},
  author={Serkan Cabi and
    Sergio G{\'o}mez Colmenarejo and
    Alexander Novikov and
    Ksenia Konyushkova and
    Scott Reed and
    Rae Jeong and
    Konrad Zolna and
    Yusuf Aytar and
    David Budden and
    Mel Vecerik and
    Oleg Sushkov and
    David Barker and
    Jonathan Scholz and
    Misha Denil and
    Nando de Freitas and
    Ziyu Wang},
  journal={arXiv preprint arXiv:1909.12200},
  year={2019}
}

See example data

There is a small amount of example data included in this repository. To examine it, run the following commands from the repository root (i.e. one level up from this folder):

python3 -m venv .sketchy_env
source .sketchy_env/bin/activate
pip install --upgrade pip
pip install -r sketchy/requirements.txt
python -m sketchy.dataset_example --show_images

For an example of loading rewards for episodes see reward_example.py.

Download the full dataset

Run ./download.sh path/to/download/folder to download the full dataset. The full dataset requires ~5.0TB of disk space to download, and extracts to approximately the same size.

You can edit download.sh to download subsets of the data.

Once the dataset has been downloaded it can be extracted wtih ./extract.sh path/to/download/folder.

Named subsets

We provide several named subsets of the full dataset, which can be easily downloaded on their own. See download.sh for a description of the subsets that are provided.

The episodes in each of these named subsets are identified by a tag in the metadata. If you would like to curate your own subset you can download the metadata file and inspect the ArchiveFiles table (see below) to figure out which archive files contain the episodes you want.

Dataset Contents

The dataset is distribted as a metadata file (metadata.sqlite) and a collection of archive files (with names ending in .tar.bz2).

The metadata file contains information about the episodes, including annotated rewards for a subset of the episodes.

Each archive file contains several episode files, which have names like 10000313341320364033_b615a417-ce34-41a8-8411-2a1ce3f3bd07.

Each episode file is a tfrecord file, containing a sequence of timesteps for a single episode.

Each timestep is a tf.train.Example proto containing features corresponding to the observations and actions from a particular point in time.

Metadata

The metadata file, metadata.sqlite, is a sqlite database containing metadata describing the contents of the files in the dataset.

The following sections describe the important metadata tables. You can find the full schema by running

sqlite3 metadata.sqlite <<< .schema

Episodes

• EpisodeId: A string of digits that uniquely identifies the episode.
• TaskId: A human readable name for the task corresponding to the behavior that generated the episode.
• DataPath: The name of the episode file holding the data for this episode.
• EpisodeType: A string describing the type of policy that generated the episode. Possible values are:
  • EPISODE_ROBOT_AGENT: The behavior policy is a learned or scripted controller.
  • EPISODE_ROBOT_TELEOPERATION: The behavior policy is a human teleoperating the robot.
  • EPISODE_ROBOT_DAGGER: The behavior policy is a mix of controller and human generated actions.
• Timestamp: A unix timestamp recording when the episode was generated.

EpisodeTags

• EpisodeId: Foreign key into the Episodes table.
• Tag: A human readable identifier for some aspect of the episode (e.g. which object set is used).

RewardSequences

• EpisodeId: Foreign key into the Episodes table.
• RewardSequenceId: Distinguishes multiple rewards for the same episode.
• RewardTaskId: A human readable name of the task for this reward signal. Typically the same as the corresponding TaskId in the Episodes table.
• Type: A string describing the type of reward signal. Currently the only value is REWARD_SKETCH.
• Values: A sequence of float32 values, packed as a binary blob. There is one float value for each frame of the episode, corresponding to the annotated reward.

ArchiveFiles

• EpisodeId: Foreign key into the Episodes table.
• ArchiveFile: Name of the archive file containing the corresponding episode.

Episodes

Each episode file is a tfrecords file containing a sequence of timesteps, encoded as tf.train.Example protos.

Each episode file contains a single episode, and each timestep within an episode contains all of the observations and actions associated with a that timestep as a single tf.train.Example. Within each episode file the timesteps are temporally ordered, so reading a file from beginning to end will visit all of the timesteps from the episode in the order they occurred.

Observations and actions occur at 10Hz.

Timesteps

Each timestep is a collection of observations and actions. Actions stored with a timestep correspond to actions taken in response to the observations they are stored with.

For a description of the shapes and types of the timestep data, see the data loader in sketchy.py.

Dataset Metadata

The following table is necessary for this dataset to be indexed by search engines such as Google Dataset Search.

property	value
name	Sketchy
url	https://github.com/deepmind/deepmind-research/tree/master/sketchy
sameAs	https://github.com/deepmind/deepmind-research/tree/master/sketchy
description	Data accompanying [Scaling data-driven robotics with reward sketching and batch reinforcement learning](https://arxiv.org/abs/1909.12200).
provider	property value name DeepMind sameAs https://en.wikipedia.org/wiki/DeepMind
citation	https://identifiers.org/arxiv:1909.12200