Speedrunning, a community centred around the practice of playing games as fast as possible and often under constraints, is a rich and vibrant part of modern games culture. It is also an area largely untouched by game AI research, despite it having its own culture of technical innovation and research. In this paper, we provide an overview of modern speedrunning, describe research areas already pursued by the community, and report on exploratory work investigating how AI research might be applied to simple speedrunning processes. We close by suggesting new research questions this space reveals for us, and suggesting ways that we can peacefully and respectfully work alongside and as a part of the speedrunning community.
(a) The player dashes up from the lfoor below towards the corner. (b) They pass through the spikes from below, touching the floor. (c) They can then jump of the spikes and dash across.
Speedrunning is the art and science of playing games quickly [
Traditional speedrunning emerged in the 1990s, as games began tracking player times and ofering target times to beat. DOOM ofered target times set by the developers for each level and also allowed players to record their performance and store it as a small file containing only player inputs, meaning they could be shared online easily, and other players could review and watch them in-game 1.
Today, speedrunning is a major subculture within videogames. The most popular website for
organising and sharing speedruns is speedrun.com2, which hosts over 5.4 million speedruns, with 2.4
million registered users. The community hosts regular charity marathons such as Games Done Quick3,
which has raised over $50 million by livestreaming speedruns and soliciting donations. Speedrunners
are often engaged in game development too, either being invited as testers for games, or as a part of the
design community for mods, fangames and standalone titles. It is increasingly common for games to
directly support speedrunners through custom features, such as in-game timers and other quality-of-life
features that make running a game easier [
The speedrunning community also has an interest in technical innovation and problem-solving.
Toolassisted speedruns, randomisers and glitch hunting (all of which we describe at length later) are aspects
of speedrunning that have led to the design of custom tools and hacks that incorporate ideas from
procedural generation, program analysis, verification and more [
Despite the incredible growth of speedrunning over the past three decades, its enormous impact on the culture of games and game development, and its passionate technical community, speedrunning has seen little exploration by AI researchers. In this paper, we outline the opportunities we see in this space for AI researchers to engage with the speedrunning community, to make contributions (without detracting from or devaluing the spirit of speedrunning), and to find exciting new research questions that challenge existing AI techniques. To support this, we also report on exploratory work we undertook to probe diferent approaches to writing AI speedrunners for the ‘fantasy console’ game engine PICO-8 4, and identify areas of future work that could build on this in the future.
This section introduces some speedrunning concepts, in particular how categories work, the technical problems and challenges involved in speedrunning, and related work in the area. We also contrast the traditional idea of game-playing AI with the challenge of speedrunning, to highlight the gap in research. 1Original demo files for DOOM runs made in 1994 can still be found online: http://cd.textfiles.com/doomcompanion/ 2https://www.speedrun.com/ 3https://gamesdonequick.com/ 4https://www.lexalofle.com/pico-8.php Due to the relatively small number of existing papers on speedrunning (which we discuss) this section is longer than usual, to provide a good grounding for all readers.
A typical speedrun will target a specific game (for modern games, sometimes a specific version of a
specific game) and aim to complete it in the fastest time possible. ‘Complete’ is usually defined as
seeing the credits roll at the end of the game, or the last moment that control is taken away from the
player. This traditional format of speedrunning is sometimes referred to as an ‘Any%’ (pronounced ‘any
percent’) run – that is, reaching the end by any means. This is known as a category of run. Games
commonly have many categories of speedrun within them, each with diferent goals and constraints
that the community has deemed meaningfully diferent and interesting [
Speedrun categories are defined by a community on a case-by-case basis. Although a single person can define a category, categories typically only become oficial once multiple runners express an interest in competing in it. A common category is ‘100%’, which usually requires the player to reach the end of the game having completed most optional objectives. The 100% speedrun for Super Metroid requires the player to finish the game with all items collected 5. Some games will have categories that are specific to a particular game. Super Mario 64, for example, has categories for 120 Stars (which is every star in the game, analogous to a 100% category) as well as 70 Stars, 16 Stars, 1 Star, and 0 Stars6. The number and variety of categories a game needs is negotiated by the community and shifts over time.
Categories can also change the way in which the game is played, putting restrictions on the game or the player. Common category modifiers require the runner to avoid using certain techniques, such as ‘No Out Of Bounds’ which prevents the runner from using any techniques which push themselves outside of the level’s ‘bounds’ (which can mean diferent things in diferent games; usually the 2D or 3D space the player is intended to be in)7. Categories can also change the game itself. For games with procedural or random elements, a common category is a ‘Seeded’ run where all players play on the same random seed, meaning that nondeterministic elements are fixed and the same for all runners 8. This changes the nature of the speedrun, mitigating some of the randomness, enabling the community to collectively plan strategies, and placing more emphasis on the execution of a strategy.
The speedrunning community is not just made up of speedrunners – many diferent people make contributions to the community. Some of those are developers of tools that help speedrunners practice or investigate games. For example, the introduction of a cutscene removal mod for Final Fantasy IX reduced the length of the speedrun by several hours, making it more viable for speedrunners to practice and increasing interest in the category9.
One of the best-known examples of technical innovation in the speedrunning scene is in tool-assisted
speedruns or TAS [
The development of new TAS strategies also reveals exploits in a game that speedrunners may be able to use elsewhere, or eventually learn to execute manually. Because TAS runs are often made and 5https://www.speedrun.com/supermetroid 6https://www.speedrun.com/sm64 7For example, Portal: https://www.speedrun.com/portal 8For example, Minecraft: https://www.speedrun.com/mc 9https://www.speedrun.com/f9 tweaked frame-by-frame, they can pay attention to minor details and find exploits that would not be discovered through normal play. This is related to the act of glitch hunting in speedrun communities, where players explore aspects of a game they believe may be vulnerable to manipulation. This can involve trying strategies that have caused glitches in other games (since games often share engines or implementations).
It is important to emphasise that TAS runs are not guided by AI search – they are made manually, and tweaked by hand to find the optimal route through a game. Their superhuman execution means that extremely unusual feats can be achieved, such as a TAS of Super Mario World which can use pixel- and frame-perfect tricks to manipulate registers and write code into the memory of the Super Nintendo console, then execute it. Another TAS author was able to find a route that played four diferent Super Mario games (1, 2, 3, and Lost Levels) simultaneously, using the same inputs to all four games, and completing all four games at the exact same moment10. This delight in pushing both hardware and software to do improbable things is a good example of how creative and fun the speedrunning community can be.
Randomisers are modifications to a game that change static game elements into dynamic ones. Common types of randomiser swap around enemy types or locations, or swap the locations of items. More extreme randomisers can change where doors or exits lead to, what the goal of the game is, or how the game looks or sounds. Randomisers generally fall into two categories: playability-preserving and -non-preserving. In the latter case, the randomiser makes no claim that any particular use of the randomiser is a completable game, or that it adheres to a particular dificulty level. These are often used for entertainment value outside of speedrunning.
In speedrunning, randomisers are more often playability-preserving. These randomisers come with internal logic, often customisable, that changes how the randomisation afects the game. For example, in the Hollow Knight randomiser, there is specific logic to ensure that unlockable abilities are never placed behind obstacles that require that ability to traverse (more generally: they ensure that keys are never behind the locks they open). More complex randomisers can take into account specific speedrunning techniques, meaning that they can place some items in out-of-reach areas if the speedrunner has stated they know how to use glitches or tricks to access them. A screenshot of a Hollow Knight randomiser can be seen in Fig. 3.
The most prominent example of speedrunning being covered in game AI research is the 2022 paper by
Mawhorter et al. [
Game AI research has already extensively examined the problem of playing a game automatically. One
of the most famous examples of this is Robin Baumgarten’s A* Mario bot, which has become a recurring
example of game AI research creating impressive gameplay [
This practice of designing AI agents to play games difers from the task of speedrunning, even in
the sense of a TAS run, in several important ways. First, research into such agents is normally more
interested in tackling a general problem rather than a specific game. The AI Platformer competition,
which Baumgarten’s bot entered and won, was not challenging entrants to complete a specific Super
Mario Bros. level, or even game, from start to finish; instead, it was interested in the general task of
controlling a game character in a platforming environment similar to Super Mario Bros. This means such
agents do not use strategies specific to a particular game, and in most cases are not playing on specific
hardware or software either. The PAC-MAN AI competitions that ran in the 2010s, for example, could
not reach the game’s infamous ‘kill screen’ because they were run on a researcher’s reimplementation
of the game [
Game-playing AI agents often have their set of available actions restricted in some way. Researchers’
goals when creating AI agents are usually either: to provide a realistic opponent for humans; to mimic
average human performance for player modelling; or to compete with and exceed human performance.
In all three cases, there is an expectation that the agent will generally avoid using glitched or unusual
behaviour, and changes are sometimes made to the AI agent to reflect this. In solving Sokoban puzzles,
for example, the best AI solvers ignore player control and instead abstract the game to be a problem of
moving boxes instead [
Existing work in game AI related specifically to speedrunning is rare. In [
Speedrunning is a vibrant community of people that has existed for decades, composed of huge numbers of players playing similarly huge numbers of games. It is important for us as AI researchers to engage respectfully with this space, and to focus on clear communication about our goals, both to establish what is interesting about this space for researchers, and to clarify to the existing community what we hope to bring to the space. We discuss engagement with the speedrunning community as a point of future work as well.
We propose that a core goal for work in this space is the implementation of game-playing agents that can perform speedrunning tasks such as running games (seen or unseen) and glitch discovery. This is distinct from traditional game-playing AI agents, and brings with it its own unique challenges and research problems. Aside from contributions to speedrunning, these problems and challenges also ofer a platform to push AI research further too, just as game-playing AI competitions have in the past.
Speedruns are typically run on specific versions of original game software. These are sometimes
emulated, but they are always based on real versions of the game released by its original developer. In
contrast, game-playing agents are usually developed in a specialised research environment (such as
machine learning ‘gyms’ [
Game-playing agents are also usually provided with an action space which defines all possible actions within the game, from which they can select a move at a particular timestep. These action sets may be simplified to focus the search towards legal moves, which might block an agent from either discovering or using glitches in their solution. For example, in speedruns for Resident Evil 7, the graphics settings menu must be accessed to change the Frames Per Second cap multiple times during a run. To our knowledge, no game-playing AI agent has ever had access to game settings as part of its action set.
The rate at which an AI agent updates is usually chosen based on a reasonable reaction window to avoid the agent being polled for decisions at too high a rate, which would slow down performance. However, many speedrunning glitches require rapid sequences of inputs or require inputs to be sent at a specific frame in the game’s execution (so-called ‘frame-perfect’ tricks). Super Mario Bros. runs at 59.97 frames per second, but the AI Mario Competition polls the agent 25 times per second - meaning the agent would not be able to perform a frame-perfect trick in the original game.
Using known glitches presents one set of problems for researchers, but discovering new glitches is its
own challenge. This has an overlap with tools for quality assurance and testing for games, which has
been tackled by both AI researchers [
Another key challenge in glitch discovery is the gap between identifying a glitch, and finding a use for it. Glitches are often discovered accidentally, and their utility is not always obvious to speedrunners. Glitches can also be discovered but considered too dificult to replicate or perform. For example, many tricks used by speedrunners of Super Mario Bros. were considered to be ‘TAS-only’ upon their discovery; that is, they required such precise inputs that they were only feasible for an automated tool-assisted speedrun. While this is still true of some tricks, others have become possible either through the increased skill ceiling of speedrunners, or by alternative methods being discovered for triggering them [18].
Thus, glitch discovery is a thorny problem for AI researchers: it relies on writing AI agents that perform in ways that run counter to most existing game-playing AI; it requires deep exploration to decide whether a glitch is useful (or even potentially useful); and it involves player/runner modelling to determine if it is feasible for humans to perform. Of course, an AI system does not need to perform all of these tasks independently to be useful – AI techniques for glitch discovery may act as tools for existing human glitch hunters to use.
In the development of this paper, we prototyped three agent-based setups using the PICO-8 game engine as a target, to begin to probe some of these questions and identify the challenges present in developing AI tooling for speedrunning. PICO-8 is a ‘fantasy console’, meaning it has been designed as a game engine with artificial limitations that evoke an earlier era of game development. These limitations (a) Celeste Classic. (b) Celeste. include a fixed palette size, a maximum length for the game’s code, and a fixed memory for sprites, game maps, music, and sound.
We chose PICO-8 as the target for our case study for several reasons. First, all games written with
PICO-8 are open source, which gave us a lot of options for exploring approaches to instrumentation,
and reading input and output from the games. PICO-8 is a great platform for AI researchers to target in
general for this reason. Second, PICO-8 includes games with a significant speedrunning community,
particularly Celeste, now known as Celeste Classic, a 2015 platformer made by Maddy Thorson and Noel
Berry [
To better understand some of the challenges and research questions involved in working with speedrunning techniques and games, we explored three diferent approaches for implementing an AI agent framework capable of interfacing with PICO-8 games. Our motivating challenge was to build an agent capable of performing a spike jump in Celeste Tech Training. A spike jump is a technique used in all versions of Celeste that takes advantage of the fact that spikes only kill the player if they are moving towards them. The player can jump ‘through’ spikes from below, allowing them to touch the lfoor, refresh their air-dash capabilities, and then jump again into the air. Figure 1 shows three stages of a spike jump in Celeste Tech Training.
One reason that leads researchers to use specially-made environments for AI research is that most games, even open-sourced games, are not designed for supporting AI agents. For example, many search-based AI algorithms require a forward model, or some way of rapidly resetting the game state. We began first by exploring whether such an implementation would be possible within a PICO-8 game itself. Writing an agent within the same project as the game it targets allows for speedrunning agents or related tech to be implemented and transferred around more easily, and allows more direct access to game features.
We modified a version of Celeste Tech Training (CTT) with the features needed to implement a search-based AI player, to see how feasible this is to do within a PICO-8 game. PICO-8’s size restrictions means that any new code we add is competing for space with the game code itself. For some games, including Celeste, developers have pushed the limits of the platform’s size restrictions so much that almost no space remains to add our custom code in. PICO-8 does support combining multiple ‘carts’ together (a ‘cart’ or cartridge is a self-contained program with its own spriteset, map, code, and so on) however carts remain self-contained and so we cannot load code from an AI agent cart while still executing code on the main game cart. For this reason, any custom API for an AI player would need to be as minimal as possible.
We first added in a virtual controller which bypassed the normal inputs for PICO-8. This was relatively simple – the cart maintains a map of PICO-8’s six buttons, and we implement functions for pressing, holding, and releasing buttons. During the update loop, these virtual button presses are converted into real button presses in the game’s logic. This allowed us to hardcode a spike jump using virtual inputs, shown in Figure 1.
We attempted to add in save states, which are useful for rapidly resetting the game during search. PICO-8 has no native support for save states, although its memory is entirely accessible and editable. PICO-8 does have support for copying large chunks of memory out to other locations – in particular, we can use a command called cstore() to save the contents of the game’s memory to another cartridge file, or memcpy() to store segments of a game’s memory into unused higher memory addresses. However, we cannot access the Lua program stack or other kinds of memory data relating to the execution of PICO-8 itself, which means although we can store a lot of information about the game state, we are unable to inject the state back into the game from within PICO-8. This makes full memory save states infeasible from within PICO-8, as far as we understand; although many other platforms exist where it may be feasible to write a speedrunning agent within the game itself, especially older platforms for which strong emulation exists [20].
If we cannot reliably implement AI agents within a PICO-8 game, then we will need an external tool to interact with and change the PICO-8 game state directly. Celia is a tool built in Lua for creating TAS runs of PICO-8 games, designed by PICO-8 user gonengazit11, which includes a number of useful features including automatically sending inputs from a TAS file. This allowed us to write an external script in Lua to generate and execute TAS files, using a modified version of CTT which could be instantly reset to the spike jump test level. After resetting the cartridge, our system would use Celia to play a randomly-generated TAS, then evaluate the result using an objective function. To simplify the search, our TAS generator only added inputs on every fifth frame, although for a full-scale system it would be simple to expand this to every frame.
This approach used purely random generation for the TAS files as a proof of concept, but it was able to successfully rediscover the spike jump mechanic after a short series of runs. Writing an AI agent in Lua using Celia has the advantage of sharing a platform/language with PICO-8, which means that while it isn’t as self-contained as an agent written into the cartridge itself, it is nonetheless quite portable and uses only a single language. However, constraining ourselves to Lua means we miss out on the opportunity to support a wider range of agents and users, and limit access to certain tools and libraries.
It is also worth noting that rediscovering an existing glitch is not the same as glitch hunting. Suppose we suspected that there were other ways to cross a spike pit in 1. In order to force a search-based agent to discover these, we would need to forbid or discourage existing techniques like the spike jump. It is hard to specifically exclude a technique as it may be possible to perform the technique in many ways, with diferent timings, or with other junk inputs mixed in. One way to search for new glitches might simply be to change the objectives. For example, if our objective function incorporates time, we might only be interested in TAS files that reach the end faster than the spike jump TAS does. Another would be to design situations that are not known to be beatable, an approach used in automated game design research [21].
Our third system took the usage of Celia and built it into a fully external system, written in Python. Python is a very popular language, particularly with AI researchers, and so this is also a good approach for building a widely-accessible framework for running events such as competitions. Our goal here was to investigate how a completely external program could use Celia to perform a search within a PICO-8 game.
Celia supports advancing and rewinding by a single frame, which we can use to more fully search the state space in a more granular way. This allowed us to build a system with a forward simulation of the game which precisely executes the game code and rolls it back using Celia’s frame skip. This enables a greedy search of the state space to look for inputs that make progress in the game, find new states, or avoid death. More importantly, because this system manipulates individual frames rather than loading a full TAS, it does not need a reset point to be added in as it does not reset the cartridge repeatedly to test new routes. This makes it a lot more eficient and closer to existing AI frameworks that can test an action and explore diferent parts of the state space. Celia can also output screenshots of the game state, which could be used for vision-based evaluations as well. Our implementation did not use this; instead, it was searching based on novelty. The source code can be found here: https://anonymous.4open.science/r/pico-speedrunning/README.md
This approach was tested on three separate, open-sourced PICO-8 games retrieved from the PICO-8 BBS, the ‘bulletin board system’ that serves as the platform’s community and game-sharing network12: GET OUT of this Dungeon13, treeboi_test14, and Witch Loves Bullets15. With all three games, randomly made TAS files were successfully generated, loaded, and played in the Celia software. The original code of each game was not modified – thus, this system can allow for speedrunning, glitch discovery, and agent replay on any PICO-8 game.
Two out of three of the games tested currently have an extremely small community of speedrunners on the Speedrun.com leaderboards. The third game, treeboi_test, was the most recently made game and chosen at random from the BBS page. For future work, we would like to develop an artificial agent to search for novel game states via screenshots using keypresses simulated with Celia. Ideally, this 12https://www.lexalofle.com/bbs/?cat=7 13https://www.lexalofle.com/bbs/?tid=31538 14https://www.lexalofle.com/bbs/?tid=142653 15https://www.lexalofle.com/bbs/?tid=35109 system can identify new tricks for speedrunning or optimize agent movements by discovering glitches to reach quicker game state times. The completion times of the agent can be evaluated against existing leaderboard times or even pave the way for a new speedrunning community for a PICO-8 game.
The authors of this paper include speedrunners, game developers and fans of speedrunning, and this paper represents our attempt to earnestly engage with an area of games culture that we believe is full of interesting potential and fun challenges, and that means a lot to us personally. However, particularly in the current climate, it is important for us as AI researchers to engage respectfully and openly with communities we wish to work alongside. Therefore, it is an important point of future work for all of us interested in this space to build connections and dialogue with communities of speedrunners and understand what interests them, what worries them, and how we can conduct our research responsibly. This is particularly relevant given that ‘AI’ is a charged term currently, and so being clear that research in the area does not need to involve LLMs and can be done locally, with no data, and little energy use, for example, may be crucial in initial dialogues.
In particular, the initial response to the concept of ‘AI for Speedrunning’ might be one of concern, as it is seeking to automate something that is fundamentally about the joy of human skill and performance. However, we see AI’s contribution to this area as being closer to TAS or ‘R&D’ (what some speedrunners call people who look for glitches). This is a way to support the curiosity of the speedrunning community with new tools and techniques that they can use and modify themselves. In the same way that speedrun randomisers are closely aligned with the goals of the procedural generation community, we hope that the common ground between game-playing AI research and speedrunning will yield exciting new ways to explore, break, and possibly even play, games of all kinds.
One topic not touched in our case studies is the challenge of routing a speedrun. Routing is the decisionmaking process concerned with which tasks a player should complete to finish a speedrun, how they should be completed, and in what order. Sometimes this is simply a question of ordering: for example the Dark Souls ‘All Bosses’ category requires that all bosses are defeated, so routing involves simply choosing which is the fastest order to fight them in 16. However, routing may also involve deciding whether some tasks should be performed at all. For example, in the Final Fantasy XII Any% speedrun, the speedrunner must acquire a certain amount of gil (the game’s currency) to spend on particular items during the run17. How they acquire this gil is completely open to the runner – they could fight monsters, do quests, or sell items. Here, routing is a task selection problem: given a set of possible tasks, what is the most eficient combination of tasks that achieves a particular goal (such as having an amount of money)?
Over time, the most optimal routes will be adopted as standard by speedrunners. However, these can change if new routes are tested and found to be faster, or – more commonly – something else changes in the speedrun that afects the rest of the route. For example, in the Bioshock speedrun, changes to the route were made after a skip was found that resulted in the player not acquiring a certain skill. Although the skip saved time overall, it required a new route and approach to certain levels to take into account the fact that the player no longer had the same skills to use.
Routing is a very complex problem in speedrunning, particularly for large games with many diferent tasks to perform and no fixed order they must be completed in. This problem maps naturally onto problems in computer science related to scheduling, planning, and search – many of which are already studied by game AI researchers, as we cited earlier in this paper. There remains a lot of unexplored space 16https://www.speedrun.com/darksouls 17https://www.speedrun.com/f13 here in terms of identifying and formalising routing problems, solving them eficiently, and explaining and visualising solutions for users.
Traditionally, AI competitions have been a powerful way to attract researchers to work collaboratively
on a problem, pool resources and create open-source example work for education and further research
[
A straightforward competition format would be an Any% speedrun of a game, although some kind of incentive or constraint would be needed to dissuade researchers from submitting an agent that simply runs the equivalent of a TAS (a randomiser would be one example of this, as this keeps most aspects of the game consistent but changes the specific challenge encountered). Alternatively, small perturbations could be introduced that are enough to derail pre-written TAS routes, but not so significant that they change the nature of the game. A final alternative, which would also be a positive way to engage with game development and speedrun communities, would be to center the competition around a base game such as Celeste Classic, but have the final competition test be on an unknown mod or variant of the game (many fangames have been made using Celeste Classic’s mechanics, but with new levels). Commissioning developers directly to produce new levels would also be a nice way to combine our communities.
Glitch discovery is another potential competition format. This could work as a bounty-type format (discovering new glitches to bypass a certain problem or area in a game, with no specific competition phase) or a more traditional competition where a set number of known glitches exist and the agent is given an amount of time to find as many as they can. Since glitch hunting can benefit from generality of approach, it may not be necessary to specify the game in advance (perhaps the genre – such as platformer – or the basic controls). Glitch hunting is a counterintuitive challenge where, by definition, setting challenges based on existing glitches may not adequately test agents’ ability to find entirely new classes of glitch that are not known to exist yet.
Speedrunning is a vibrant community of both research and practice, and has been in existence for longer than any videogame AI conference or journal. It is much more than simply playing games quickly it’s a space where unusual forms of play are celebrated and discovered, where game design is pushed beyond its breaking point, and where incredible interdisciplinary work takes place. AI researchers are yet to properly engage with this space, and it holds myriad exciting new challenges and problems for them to tackle, including a very diferent approach to game-playing than game AI has historically focused on.
In this paper, we discussed some of the basic properties of modern speedrunning and how this diferentiates itself from most of the problems that game AI researchers have tackled in the past. We reported on our work investigating PICO-8 as a platform for speedrunning, and used this to highlight some future work that could potentially be carried out in this area, particularly in organising future events that researchers can focus on together, such as competitions. Above all, we want to ensure that researchers approach this community with an open mind and a respectful manner, and build relationships with the community who already know this space so well, and who have worked so hard to build it into what it is today. We hope this paper encourages others to take a look in this area, and join us in exploring the future potential of research in this area.
No Generative AI tools were used in the preparation of this paper.
The authors wish to thank the speedrunning community who have influenced our thinking in so many diferent ways and led us to writing this paper. Thanks to the reviewers for their kind feedback and suggestions. This work began at Dagstuhl Seminar 24261: Computational Creativity for Game Development. Thanks to the seminar organisers and to Schloss Dagstuhl for facilitating this.
Filippo Carnovalini is funded by the European Union, HORIZON- MSCA-2022-PF-01 Project ID 101108690 (CALIOPE). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the REA can be held responsible for them. dificulty and engagement using ai players, Proceedings of the ACM on Human-Computer Interaction 5 (2021). [17] C. Muratori, Walk monster, 2012. URL: https://caseymuratori.com/blog_0005. [18] H. Kemps, Super mario bros. speedrun record shattered by a matter of milliseconds, 2016. URL: https://www.vice.com/en/article/super-mario-speedrun-record-shattered-by-milliseconds/. [19] rarelikeaunicorn, Astra and the new constellation, 2023. URL: https://rarelikeaunicorn.itch.io/astra. [20] J. Franušić, K. Tuite, A. Smith, Playable quotes for game boy games, in: Proceedings of the 18th
International Conference on the Foundations of Digital Games, 2023. [21] M. Cook, S. Colton, A. Raad, J. Gow, Mechanic miner: reflection-driven game mechanic discovery and level design, in: Proceedings of the 16th European Conference on Applications of Evolutionary Computation, 2013. [22] gonegazit, Celia - a PICO-8 TAS framework, 2022. URL: https://github.com/gonengazit/celia. [23] C. Salge, M. C. Green, R. Canaan, J. Togelius, Generative design in minecraft (gdmc): settlement generation competition, in: Proceedings of the 13th International Conference on the Foundations of Digital Games, 2018. [24] M. Čertický, D. Churchill, K.-J. Kim, M. Čertický, R. Kelly, Starcraft ai competitions, bots, and tournament manager software, IEEE Transactions on Games 11 (2019).