Code and data for 'The Grand Atari Challenge dataset' paper.

Cite the Atari Grand Challenge dataset as:

V. Kurin, S. Nowozin, K. Hofmann, L. Beyer, and B. Leibe. The Atari Grand Challenge Dataset. arXiv preprint arXiv:1705.10998, 2017.

There are several options for you:

• the whole dataset (screenshots + trajectories) for all five games
• trajectories only for all five games
• per-game screenshots+trajectories:
  • Space Invaders
  • Q*bert
  • Ms. Pacman
  • Video Pinball
  • Montezuma's Revenge

In order to use the data in a convenient way, we release the small lib agc. Installation is easy, just run this in the directory root.

pip3 install -e .

Then you can use the library in your python code:

import agc.dataset as ds
import agc.util as util

# DATA_DIR is the directory, which contains the 'trajectories' and 'screens' folders
dataset = ds.AtariDataset(DATA_DIR)


# dataset.trajectories returns the dictionary with all the trajs from the dataset
all_trajectories = dataset.trajectories

You can find more examples in the notebooks folder.

P.S.

We've done some fixes after the initial release and removed the trajectories with final zero score from the data, so, we have smth like 3 hrs less of replays as we claimed in the paper. Though, we already have enough data to release a second version, so, dataset_v2 will be here soon. We plan to release the data and update the preprint after it.

upd: v2 is released. Preprint update is in progress.

P.P.S

In case you found a bug or just in trouble, email me [email protected] =)