Sokoban is Japanese for warehouse keeper and a traditional video game.
The game is a transportation puzzle, where the player has to push all boxes in the room on the storage locations/ targets.
The possibility of making irreversible mistakes makes these puzzles so challenging especially for Reinforcement Learning algorithms, which mostly lack the ability to think ahead.
The repository implements the game Sokoban based on the rules presented DeepMind's paper Imagination Augmented Agents for Deep Reinforcement Learning.
The room generation is random and therefore, will allow to train Deep Neural Networks without overfitting on a set of predefined rooms.
| Example Game 1 | Example Game 2 | Example Game 3 |
|---|---|---|
 |
 |
 |
pip install gym-sokobangit clone [email protected]:mpSchrader/gym-sokoban.git
cd gym-sokoban
pip install -e .Checkout the examples on how to use an external gym environment.
Every room consists of five main elements: walls, floor, boxes, box targets, and a player. They might have different states whether they overlap with a box target or not.
| Type | State | Graphic | TinyWorld |
|---|---|---|---|
| Wall | Static |  |
 |
| Floor | Empty |  |
 |
| Box Target | Empty |  |
 |
| Box | Off Target |  |
 |
| Box | On Target |  |
 |
| Player | Off Target |  |
 |
| Player | On Target |  |
 |
The game provides 9 actions to interact with the environment. Push and Move actions into the directions Up, Down, Left and Right. The No Operation action is a void action, which does not change anything in the environment. The mapping of the action numbers to the actual actions looks as follows
| Action | ID |
|---|---|
| No Operation | 0 |
| Push Up | 1 |
| Push Down | 2 |
| Push Left | 3 |
| Push Right | 4 |
| Move Up | 5 |
| Move Down | 6 |
| Move Left | 7 |
| Move Right | 8 |
Move simply moves if there is a free field in the direction, which means no blocking box or wall.
Push push tries to move an adjacent box if the next field behind the box is free. This means no chain pushing of boxes is possible. In case there is no box at the adjacent field, the push action is handled the same way as the move action into the same direction.
Finishing the game by pushing all on the targets gives a reward of 10 in the last step. Also pushing a box on or off a target gives a reward of 1 respectively of -1. In addition a reward of -0.1 is given for every step, this penalizes solutions with many steps.
| Reason | Reward |
|---|---|
| Perform Step | -0.1 |
| Push Box on Target | 1.0 |
| Push Box off Target | -1.0 |
| Push all boxes on targets | 10.0 |
Every time a Sokoban environment is loaded or reset a new room is randomly generated. The generation consists of 3 phases: Topology Generation, Placement of Targets and Players, and Reverse Playing.
To generate the basic topology of the room, consisting of walls and empty floor, is based on a random walk, which changes its direction at probability 0.35. At every step centered at the current position, a pattern of fields is set to empty spaces. The patterns used can be found in Figure 2.
Figure 2: Masks for creating a topology
During this phase, the player including all n box targets are placed on randomly chosen empty spaces.
This is the crucial phase to ensure a solvable room. Now Sokoban is played in a reverse fashion, where a player can move and pull boxes. The goal of this phase is to find the room state, with the highest room score, with a Depth First Search. For every room explored during the search is a room score is calculated with the equation shown below. The equation is a heuristic approach to evaluate the difficulty of the room. BoxSwaps counts the number of times a player changes the box to pull. BoxDisplacement is the Manhattan Distance between a specific box and its origin box target. As long as at least one box is on a target the RoomScore is always 0.
Sokoban has many different variations, such as: Room Size, Number of Boxes, Rendering Modes, or Rules.
Besides the regular Sokoban rendering, each configuration can be rendered as TinyWorld, which has a pixel size equal to the grid size.
To get an environment rendered as a tiny world just add tiny_ in front of the rendering mode. E.g: env.render('tiny_rgb_array', scale=scale_tiny). Scale allows to increase the size of the rendered tiny world observation. Using scale in combination with the rendering modes, human or rgb_array, does not influence the output size.
Available rendering modes are:
| Mode | Description |
|---|---|
| rgb_array | Well looking 2d rgb image |
| human | Displays the current state on screen |
| tiny_rgb_array | Each pixel describing one element in the room |
| tiny_human | Displays the tiny rgb_array on screen |
The available room configurations are shown in the table below.
| Room Id | Grid-Size | Pixels | #Boxes | Example | TinyWorld |
|---|---|---|---|---|---|
| Sokoban-v0 | 10x10 | 160x160 | 3 |  |
 |
| Sokoban-v1 | 10x10 | 160x160 | 4 |  |
 |
| Sokoban-v2 | 10x10 | 160x160 | 5 |  |
 |
| Sokoban-small-v0 | 7x7 | 112x112 | 2 |  |
 |
| Sokoban-small-v1 | 7x7 | 112x112 | 3 |  |
 |
| Sokoban-large-v0 | 13x11 | 208x176 | 3 |  |
 |
| Sokoban-large-v1 | 13x11 | 208x176 | 4 |  |
 |
| Sokoban-large-v2 | 13x11 | 208x176 | 5 |  |
 |
| Sokoban-huge-v0 | 13x13 | 208x208 | 5 |  |
 |
Please note that the larger rooms might take some time to be created, especially on a laptop.
Besides the regular game of Sokoban, this repository implements or will implement variations, which might make the game easier or more complicated. Except noted differently the variations do not implement a Tiny-World version.
| Variation | Summary | Expected Difficulty | Example | Tiny World | Status | Details |
|---|---|---|---|---|---|---|
| Fixed Targets | Every box has to be pushed on the target with the same color. | More difficult |  |
Yes | implemented | ReadMe |
| Multiple Player | There are two players in the room. Every round one of the two players can be used. There is no order of moves between the two players. | More difficult |  |
Yes | implemented | ReadMe |
| Push&Pull | The player can not only push the boxes, but also pull them. Therefore, no more irreversible moves exist. | Easier | |
Yes | implemented | ReadMe |
| Boxoban | Uses by DeepMind pregenerated Sokoban puzzles. | Similar | |
Yes | Implemented | ReadMe |
If you are using this repository for your research please cite it with the following information:
@misc{SchraderSokoban2018,
author = {Schrader, Max-Philipp B.},
title = {gym-sokoban},
year = {2018},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/mpSchrader/gym-sokoban}},
commit = {#CommitId}
}
Feel free to get in touch with me to talk about this or other projects. Either by creating an issue or mail me on LinkedIn.
If you reached the end and liked the project, please show your appreciation by starting this project.
Feel free to contribute to this project by forking the repo and implement whatever you are missing. Alternatively, open a new issue in case you need help or want to have a feature added.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent