Definition
Sokoban-style pushing is a mechanic in which the player pushes objects across a grid to bring them onto designated squares. The player has exactly two verbs, walk and push; pulling is not available. Each turn changes only the player's position and the position of the pushed object. The level is solved when that arrangement satisfies the stated goal.
The core of the mechanic is not the small verb count but the irreversibility of a single move. A pushed box cannot be taken back, so the board divides into arrangements that are still solvable and arrangements that no longer are. Nothing stops the player from continuing after crossing that line, so failure arrives quietly. As the board above shows, a box shoved against a wall can never be moved back the other way. Finding those deadlocks before they happen is the actual play.
Origins and lineage
The mechanic first appears in Sokoban, published by Thinking Rabbit in 1982 and designed by Hiroyuki Imabayashi. The Japanese title 倉庫番 means "warehouse keeper," and the romanised form Sokoban became the name of the whole genre abroad. A hobby version for the NEC PC-8001 is reported to predate it in 1981, with the commercial release arriving from 1982 on the PC-8801 and other machines.
Games of that period lived almost entirely inside arcade grammar, where reflexes carried the reward. Sokoban stepped outside it and offered a board with no real-time pressure, no enemies and no score display. Sokoban (1982) covers that context in detail. In an interview in the December 1983 issue of the magazine LOGiN, Imabayashi described Sokoban as "sort of like tsume shogi," the Japanese chess-problem form. Little else survives in which he explains his design intent, so that remark is one of the few footholds.
The lineage branched almost at once. HAL Laboratory's Eggerland (1985) changed the pushed object into an enemy turned into an egg, and settled the one-room format (Eggerland (1985)). Adventures of Lolo (1989) followed in the same line (Adventures of Lolo (1989)), and Atlus's Puzzle Boy, released abroad as Kwirk (1989), added turnstiles that rotate ninety degrees when pushed (Puzzle Boy / Kwirk (1989)). How these games travelled abroad as a verb rather than under their own names is traced in names are taken, verbs survive.
Variations
The most economical variation changes the nature of the pushed object without adding a verb. Sokobond (2014, Alan Hazelden and Lee Shang Lun) makes the pushed objects atoms; adjacent atoms bond and then move as one molecule, so irreversibility shifts from position to shape. A Good Snowman Is Hard To Build (2015) grows the snowball with every push, and in Snakebird (2015) and Can of Wormholes (2023) the pusher itself is a body that stretches and shrinks.
Other games push the object further toward the abstract. Baba Is You (2019, Hempuli) makes the pushed objects the word blocks that state the rules (rule rewriting). Patrick's Parabox (2022, Patrick Traynor) puts boards inside boxes so a push chains across nested levels (recursion and nesting). Void Stranger (2023) replaces the verb with carrying floor tiles, and withholds undo besides. In all three the operation stays "push the adjacent thing." Moving that single dimension is the method examined in the verb of pushing and when fewer verbs make a richer game.
A third direction adds hidden state to the pushed object. Stephen's Sausage Roll (2016, increpare) gives a rolling sausage a cooked-face state, and Pipe Push Paradise (2018) handles the shape and height of the pushed pipe in three dimensions. In Bonfire Peaks (2021) the carried crate doubles as footing, and Isles of Sea and Sky (2024) opens the pushing surface into a wide island and mixes in exploration. A Monster's Expedition (2020) shows the opposite choice: it is built so that no arrangement can become unsolvable, stripping irreversibility out of the pushing mechanic entirely.
Design pitfalls
The first pitfall is state explosion. In a walking maze the state is only the player's position, but in a pushing puzzle the whole arrangement of boxes is the state, and each added box multiplies the possibilities. Deciding whether an arbitrary Sokoban level is solvable is in fact NP-hard (Fryers and Greene 1995; Dor and Zwick 1999) and further shown to be PSPACE-complete (Culberson 1997; Hearn 2006). For a player that means the board cannot be read to the end; for a designer it means the board must be arranged so that most bad moves look bad early.
The second pitfall is teaching deadlock in the wrong order. Failure in a pushing puzzle is invisible at the moment it occurs, so an unexplained dead end is remembered only as unfairness. Early levels should be arrangements in which no deadlock is possible, instilling only the premise that a push cannot be undone; the first small deadlock belongs later, once that premise is settled. How much of a dead end to show is the subject of legible failure, and the choice between granting undo and punishing mistakes is treated in the ethics of undo. Many modern games ship undo as standard precisely to move that teaching burden outside the mechanic.
The third pitfall is buying difficulty with floor space. A larger board lengthens the solution but thins the thinking, and the work drifts toward brute force. Why closing a puzzle into a single screen works is set out in designing within one screen. Procedural generation falls into the same hole: solvable levels are easy to produce in bulk, levels containing one insight are not (handmade levels, machine-made levels). The same hardness is now used to measure planning in AI systems (Monti et al. on measuring planning with corridor Sokoban).
How to tell it apart
What separates this from constrained grid movement is whether anything is pushed at all. Constrained grid movement only attaches conditions to how the player may move, so the state is essentially the player's position. In Sokoban-style pushing the arrangement of objects is the state, and that arrangement carries irreversibility from the start. Rule rewriting and recursion and nesting are best read as higher mechanics sitting on top of the pushing grammar. A useful test: if the game still works with pushing removed, it is a different mechanic; if it collapses, it is a variation of this one.
The difference from physics simulation and magnetism and gravity is whether quantities are continuous. Movement here is discrete, one square at a time, with no momentum and no friction. Rolling is a close neighbour with a different character: a rolling object gains orientation as hidden state, yet its position can usually be returned to (English Country Tune, the verb of rolling). Rolling creates difficulty by adding state; pushing creates it by removing reversibility.
Further reading
On this site: the verb of pushing (the central essay on this mechanic) / Sokoban (1982) (its first appearance) / the philosophy of Hiroyuki Imabayashi / how to get faster at logic puzzles: Sokoban-style tips (the solver's side) / the ethics of undo / legible failure / the lineage of meta-puzzles / recursion as puzzle grammar / when the control scheme decides the difficulty
Primary sources off-site: Wikipedia (English): Sokoban (release year, designer, and the citations for each complexity result), Wikipedia (Japanese): 倉庫番 (port history), Thinking Rabbit 1983 Developer Interview (a translation of the December 1983 LOGiN issue, and the source of the "tsume shogi" remark). First-hand material on the design intent of Imabayashi and Thinking Rabbit is scarce, and this entry marks where the evidence stops.
相邻机制
Constrained Grid Movement / Rule Rewriting / Recursion & Nesting / Physics Simulation
