Sokoban puzzle generator
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

造倉庫 ぞうそうこ Zōsōko Zosoko

Sōkoban puzzle generator

Sōkoban is a Japanese transport puzzle in which the player pushes boxes around a warehouse. See wikipedia for more information about Sōkoban.

This library will generate random Sōkoban puzzles for use in games. A player is included so you can quickly get a feel for the runtime of the algorithm and the library's features.


Build in eclipse, then execute with:

  • java -cp bin net.bytten.zosoko.player.Main -box-lines

Also requires my gameutil library, which you can get the source for on github.


  • -width=W Sets the width of generated puzzles.
  • -height=H Sets the height of generated puzzles.
  • -boxes=N Sets the number of boxes and goals in generated puzzles.
  • -unbounded Allow the player to walk around outside the bounds of the puzzle.
  • -test Always generates the puzzle designed for testing the renderer and controller.
  • -limit-goal-experiments=N Limit the number of combinations of goals the generator tries for each map.
  • -limit-depth=N Limit the depth of the farthest state search.
  • -time-limit=N Limit how long (in milliseconds) to search.
  • -box-lines Sets the scoring metric to purely box-lines rather than the default based on the paper.

The best results seem to be given with widths and heights that are multiples of 3 and when -box-lines is given.


  • Blue squares are boxes.
  • The red circle is the player.
  • Use the arrow keys to move the player and push boxes.
  • Press R to reset the puzzle.
  • Press F5 to generate a new puzzle.


  • Needs significant clean-up.
  • Profiling indicates the slowest part is PuzzleMap.getPlayerSpacePartition(). The use of this can be eliminated entirely by implementing the suggestions in PlayerCloud, or maybe there is a way to speed it up?
  • After that is Vec2I.add, which must mean that there are no further bottlenecks.
  • See if making FarthestStateFinder.go return a set of states would make the default scoring metric work better.


The paper this library is based on: