19. The VGDL landscape

Status: research summary (reference only)

This entry condenses a research conversation on 2026-09-22. Papers and links were gathered by the assistant then. On 2026-09-29 the arXiv pages for the GVGAI survey (1802.10363) and Griddly (2011.06363) and the schaul/py-vgdl and GAIGResearch/GVGAI repositories were confirmed to exist; the other links were not re-checked.

What a VGDL is

A domain-specific, declarative language for the rules and layout of 2D sprite games, compact enough to be generated and simulated automatically. Academic VGDL has two files, a game description and an ASCII level map. The game description has four parts:

Part	Purpose
SpriteSet	Hierarchical list of object types; children inherit parents’ rules and can override them
LevelMapping	Maps characters in the level file to sprite types
InteractionSet	Rules that fire only on collision between two sprites
TerminationSet	Win and lose conditions

There is no general scripting: behavior comes from picking built-in sprite classes and writing collision rules. There is no dedicated Wikipedia article; the closest coverage is under General game playing, and the similarly named Game Description Language is a different, older language.

Compared with other approaches

Approach	Idea	Relation to this project
GDL (Stanford, Genesereth)	Logic-based (Datalog-like), agents reason over the rules; very verbose because nothing is built in	Opposite philosophy: XMLGameEngine leans on a large built-in vocabulary
RBG (Regular Boardgames)	Rules as regular expressions, for finite deterministic turn-based games	Sibling for board games
Ludi / Ludii	“Ludemic” decomposition of game components	Sibling for board games
PuzzleScript (Lavelle)	Declarative find-and-replace rules over a grid; turn-based; rules are not limited to collisions	Closest spiritual cousin
Imperative engines (Unity, Unreal, GML)	Full control flow	Unlimited expressiveness, no generic reasoning about a game
Griddly (GDY, YAML)	C++-core rewrite of the GVGAI idea for speed	The successor that mostly replaced VGDL in AI research

Papers and code

Origin:

• Ebner, Levine, Lucas, Schaul, Thompson, Togelius. “Towards a Video Game Description Language.” Dagstuhl Follow-Ups, 2013. Dagstuhl
• Schaul. “A Video Game Description Language for Model-based or Interactive Learning.” IEEE CIG, 2013. PDF. Source: schaul/py-vgdl.

GVGAI competition era:

• Perez-Liebana et al. “General Video Game AI: a Multi-Track Framework…” IEEE Trans. on Games, 2019. arXiv:1802.10363. The best single survey.
• Khalifa, Green, Perez-Liebana, Togelius. “General Video Game Rule Generation.” arXiv:1906.05160.
• Torrado et al. “Deep Reinforcement Learning for General Video Game AI.” CIG 2018. Code: GVGAI_GYM.
• The physics and macro-actions extension, IEEE CIG 2017 (IEEE Xplore), one of the few times the language itself was extended.
• Frameworks: GAIGResearch/GVGAI (Java), gvgai.net, Python 2.0 rewrite rubenvereecken/py-vgdl.

Applications and successors:

• Broll et al. “AtDelfi: Automatically Designing Legible, Full Instructions For Games.” arXiv:1807.04375.
• “Planning from video game descriptions.” arXiv:2109.00449.
• “Game Generation via Large Language Models.” arXiv:2404.08706 (an LLM writing VGDL directly).
• Bamford, Huang, Lucas. “Griddly: A platform for AI research in games.” AAAI 2021. arXiv:2011.06363; code Bam4d/Griddly; docs griddly.readthedocs.io.
• PuzzleScript: puzzlescript.net; the study of what gives it power, ScriptButler (CWI, 2023): PDF.
• Regular Boardgames: AAAI paper.

Why the field stayed small

• It was built for one job: an unlimited supply of benchmark games for AI agents. Success shows up as papers using GVGAI, not as papers extending VGDL.
• A closed vocabulary means a closed language: a genuinely new dynamic must be coded into the engine, not written in the description.
• Each generation is reimplemented under a new name (GDL to RBG and Ludii, VGDL to Griddly and MiniHack), so citations restart from zero and the field looks smaller than the idea.

“No loops or conditionals”: does it make sense?

The conditionals and loops are implicit, not absent: the engine’s per-tick evaluation is the loop and pattern matching on collisions or grid state is the conditional. This is a production-rule or term-rewriting system, the same paradigm as cellular automata, Datalog and regex-based rule languages. The trade-off is explicit in the literature: expressiveness against compactness and generability. VGDL chose compactness so search and evolution can mutate thousands of games; a Turing-complete layer would kill that. It stops making sense where the sprite and interaction vocabulary runs out. XMLGameEngine makes the same bet, with exprtk carrying per-rule arithmetic rather than general control flow, and pays the same price (01, 12).

Second batch: alternatives

• Current state. The research VGDL is declarative and text-based, built for two-dimensional arcade-style games: sprites, interactions, terminations, a level layout. It became a foundation for general game-playing agents and, more recently, for generating games with language models.
• How the XML differs. A structured element tree with schema validation, opt-in parts per object, and a growing verb set; similar spirit, more explicit structure.
• Related structure. Entity-component systems are a near neighbor; the attraction is the same (parts a game can opt into).
• A question worth writing up. How small can a vocabulary be and still express a class of games; related formal versions ask for the minimal rule set that distinguishes one game from another.

Sources

• “Video game description languages: overview and research” (2026-09-22).
• Second batch: “VGDLs and Technology” (2026-05-09), “VGDL Papers and Projects” (2025-04-24), “Declarative Pong XML redesign” (2026-09-23).