Skip to content
This repository contains levels for boxoban, a box-pushing puzzle game inspired by Sokoban.
Branch: master
Clone or download
Latest commit d5ca561 Jan 24, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
hard Initial commit Dec 14, 2018
medium Initial commit Dec 14, 2018
unfiltered Initial commit Dec 14, 2018 Initial commit Dec 14, 2018
LICENSE Initial commit Oct 26, 2018
LICENSE.txt Initial commit Dec 14, 2018 Update Jan 23, 2019


This repository contains levels for boxoban, a box-pushing puzzle game inspired by Sokoban.

If you use this dataset in your work, please cite the following:


author = {Arthur Guez, Mehdi Mirza, Karol Gregor, Rishabh Kabra, Sebastien Racaniere, Theophane Weber, David Raposo, Adam Santoro, Laurent Orseau, Tom Eccles, Greg Wayne, David Silver, Timothy Lillicrap, Victor Valdes},
title = {An investigation of Model-free planning: boxoban levels},
howpublished= {},
year = "2018",

Questions regarding the dataset can be directed to Theophane Weber (


The aim is to push the boxes on top of the targets. Each level has four boxes and four targets. The player can push a box as long as nothing (another box, a wall) is behind it.

Puzzles are grouped into sets of one thousand puzzles, each set encoded as a text file. Each puzzle is a 10 by 10 ASCII string which uses the following encoding: '#' for wall, '@' for the player character, '$' for a box, and '.' for a goal position.

Within a file, puzzles are separated by a semicolon and a puzzle number. There are three levels of difficulties fo puzzles: 'unfiltered', 'medium', and 'hard'.

The unfiltered levels are separated into train (900000 puzzles), validation (100000 puzzles), and test (1000 puzzles).

Unfiltered puzzles are generated according to the procedure described in 'Imagination-Augmented Agents for Deep Reinforcement Learning' (Racaniere et. al, proceedings of NeurIPS, 2017). The unfiltered test is the one used by Orseau et al. in 'Single-Agent Policy Tree Search With Guarantees' (proceedings of NeurIPS, 2018).

The medium and hard datasets generation procedure is described in 'An investigation of model-free planning' (Guez, Mirza, Gregor, Kabra et. al, arxiv, 2018). The medium set is composed of train (450000 puzzles) and validation (50000 puzzles); the hard set has 3332 levels.


This is not an official Google product.

You can’t perform that action at this time.