Tool-assisted speedrun


A tool-assisted speedrun or tool-assisted superplay is generally defined as speedrunning a game in an emulator with the goal of creating a theoretically perfect playthrough. As the name implies, a TAS is not performed by an actual human being, but rather by a program or a piece of software that delivers frame-perfect optimized controller input to complete the game in the fastest way possible. The script that delivers these inputs is provided by the TAS author, who would use their knowledge of the game's mechanics and various tools built into the emulator to optimize a speedrun until no more improvements can be identified. Tools used to this end include using savestates and branches, slow-down and going frame-by-frame, creating macros and scripts to perform automated actions, and so on. At the extreme end of this endeavor, means such as disassembly and brute-forcing can be used.
The idea is not to make gameplay easier for players, but rather to produce a playthrough at a level of gameplay impractical for a human player. As such, rather than being a branch of e-sports focused on practical achievements, tool-assisted speedrunning concerns itself with research into the theoretical limits of the games and their respective competition categories, and producing content with an emphasis on entertainment value — such as by including tricks and stunts that would otherwise be prohibitively difficult to incorporate.

History

1999–2001

The term was coined during the early days of Doom speedrunning, during which the first of these runs were made. When Andy "Aurican" Kempling released a modified version of the Doom source code that made it possible to record demos in slow motion and in several sessions, it was possible for the first players to start recording tool-assisted demos. A couple of months afterwards, in June 1999, Esko Koskimaa, Peo Sjoblom and Joonatan Donner opened the first site to share these demos, "Tools-Assisted Speedruns".
Like many other tool-assisted speedrun communities, the maintainers of the site stressed the fact that their demos were for entertainment purposes rather than skill competitions, although the attempt to attain the fastest time possible with tools itself became a competition as well. The site became a success, updating usually several times a week with demos recorded by its maintainers and submitted by its readers. After a short while, when version 2.03 of Lee Killough's Marine's Best Friend Doom source port was released, it became even easier for people to record these demos, adding the functionality of re-recording without having to replay the demo until it reached the point where the player wanted to continue.
The site was active until August 10, 2001, when Jonathan Donner posted a news message stating that their site would be an archive from now on, and pointing towards The Doomed Speed Demos Archive, a site mainly for non-assisted speedruns, of which the author agreed to take over the posting of tool-assisted speedruns. Although popularity dwindled since then, built demos have still been submitted as late as November 2005, and are usually made with PrBoom.

2003–present

In 2003, a video of a Japanese player named Morimoto completing the NES game Super Mario Bros. 3 in 11 minutes and performing stunts started floating around the Internet. The video proved to be controversial, as not many people knew about tool-assisted speedruns at the time, especially for the NES. As the video was not clearly labelled as such, many people felt they had been cheated when they found out it was done using an emulator. The video, however, inspired Joel "Bisqwit" Yliluoma to start a website called NESvideos, which was dedicated to tool-assisted speedruns for the NES. At first it hosted videos only for the NES, but as the community grew, its members added the features required for tool-assisted speedrunning into emulators for other systems. The name of the site was later changed to TASVideos. As of May 2020, TASVideos is the largest English-language web community that produces and hosts tool-assisted speedruns; the site holds 4161 complete speedruns, of which 2213 are the fastest of their kind.
Tool-assisted speedruns have been made for some notable ROM hacks as well as for published games. In 2014 a speedrunning robot, TASBot, was developed, capable of performing TAS runs via direct controller input.
A joke personification of tool-assisted speedruns, called TAS-san, has become popular among Japanese Internet users. Tool-assisted speedruns uploaded to sites like Nico Nico Douga, YouTube or TASVideos may be described as a new world record by TAS-san, who is said to have the superhuman memory and reflexes needed to execute such a speedrun in real time.

Method

Creating a tool-assisted speedrun is the process of finding the optimal set of inputs to fulfill a given criterion — usually completing a game as fast as possible. No limits are imposed on the tools used for this search, but the result has to be a set of timed key-presses that, when played back on the actual console, achieves the target criterion. The basic method used to construct such a set of inputs is to record one's input while playing the game on an emulator, all the while saving and loading the emulator's state repeatedly to test out various possibilities and only keep the best result. To make this more precise, the game is slowed down. Initially, it was common to slow down to some low fraction of normal speed. However, due to advances in the field, it is now expected that the game is paused during recording, with emulation advanced one frame at a time to eliminate any mistakes made due to the urgency.
The use of savestates also facilitates another common technique, luck manipulation, which is the practice of exploiting the game's use of player input in its pseudo-random number generation to make favorable outcomes happen. Using a savestate from before some event, it is possible to experiment with small input variations until the event has the desired outcome. Depending on the game and event, this can be a very time-consuming process, at times requiring much backtracking, and can as such take up a large portion of the total time spent making a tool-assisted speedrun. Examples of luck manipulation include making the ideal piece drop next in Tetris, or getting a rare item drop the first time one kills an enemy in an action game.
A rarely used tool is brute-force searching for ideal inputs by making a computer play the game, trying all possible inputs. In theory, this process could find the ideal set of inputs for any game, but since the space of all possible inputs grows exponentially with the length of the sequence, this is only viable for optimizing very small portions of the speed run. Instead, a heuristic algorithm can be used. Although such an approach does not guarantee a perfect solution, it can prove very effective for solving simple puzzle games.
Another rarely used technique is disassembling the game executable. By exposing the game logic, this enables the player to manipulate luck without trial and error, or reveal obscure bugs in the game engine. A more common, related technique, is to monitor the memory addresses responsible for certain effects to learn when or how they change. Memory watching is supported by most emulators used on TASVideos.org.
All these techniques involve direct interaction with the game state in ways not possible without emulation, but the final result, the set of inputs that makes up the speedrun, does not depend on such manipulation of the state of the emulated machine. The tool use in tool-assisted speedrunning is therefore different from the sort of state manipulation that tools like Gameshark provide, since such manipulation would not be expressible as a sequence of timed inputs.

Re-recording emulators

Tool-assisted speedrunning relies on the same series of inputs being played back at different times always giving the same results. In a manner of speaking, the emulation must be deterministic with regard to the saved inputs. Otherwise, a speedrun that was optimal on one playback might not even complete it on a second playback. This loss of synchronization, or "desync", occurs when the state of the emulated machine at a particular time index no longer corresponds with that which existed at the same point in the movie's production. Desyncs can also be caused by incomplete savestates, which cause the emulated machine to be restored in a state different from that which existed when it was saved. Desyncs can also occur when a user attempts to match inputs from an input file downloaded from TASVideos and fail to match the correct enemy reactions due to bad AI or RNG.
Problems with emulation, such as nondeterminism and incomplete savestates, are often only discovered under the precise frame-by-frame conditions of tool-assisted speedrunning. Emulator developers often do not give speedrunning issues high priority because they have little effect on regular gameplay; consequentially the community has forked several emulators to make them suitable for the task. These include Snes9X improvement, Gens rerecording, VBA rerecording and Mupen rerecording. If a forked emulator is used to produce a TAS, playback on the normal, unmodified version of the emulator will usually result in a desync.
Emulators that currently feature the tools necessary to create tool-assisted speedruns include the Arcade emulator MAME, the NES emulator FCEUX, the Super NES emulator Snes9x, the Genesis emulator Gens, the Game Boy Advance emulator VisualBoyAdvance, the Nintendo 64 emulator Mupen64, the GameCube and Wii emulator Dolphin, the Nintendo DS emulator DeSmuME, the Sega Saturn emulator Yabause, the PlayStation emulator PSXjin, and several others for these and other platforms.
In 2012, there was a release by TASVideos.org which is an all-in-one emulator called Bizhawk. Due to the success of some of the cores that are built into the emulator, the team are phasing out some of their older emulators towards and the end of the year and the team encourage TASers who were working on Nintendo 64 and PSX projects for submissions on their website to move to Bizhawk.

Relation to unassisted runs

Tool-assisted speedruns are timed in a distinct category from unassisted runs, for reasons of fairness. In unassisted runs, a difficult path is often avoided in favor of a safer, but slower one, in order to avoid risks such as dying and having to start over, failing a trick and wasting more time, or failing a setup for a difficult trick. Depending on the game, tool-assisted speedruns can surpass their unassisted counterparts by a few seconds to entire hours. For an example of a highly optimized real-time run, the fastest Super Mario Bros. TAS currently stands at 04:57, while the fastest unassisted run stands at 4:57.

Timing conventions

Tool-assisted runs are timed by input, i.e. from game power-on to the last input necessary to reach the ending scene and/or the game credits. Any introductory cutscenes, game-loading screens, and trailing dialogues after the last boss battle are included in the final times. The times are exact, a level of precision that is not possible with unassisted runs because it cannot be determined from a recording when exactly the input ended. Speed Demos Archive and Twin Galaxies measure only the length of the gameplay proper, and begin timing when the player gains control of the character and ends timing when the player loses it. These differences in timing conventions can result in seemingly discrepant times between unassisted and tool-assisted runs. For example, a Super Mario Bros. speedrun by Andrew Gardikis, a 4:58 by SDA timing, seems to be only 0.69 seconds slower than a TAS of 4 minutes and 57.31 seconds by HappyLee, but his run actually contains 5 minutes and 1 second of input starting from power-on.

Unassisted runs faster than their assisted counterparts

Because tool-assisted speedruns often take more time to create than unassisted speedruns, discovery of a time-saving trick may lead to a situation of the fastest unassisted speedrun being faster than its tool-assisted counterpart.
From August 13 to 21, 2007, the fastest unassisted speedrun of Pokémon Blue was 4 minutes faster than the best TAS due to a new trick that allowed walking through walls. On August 21, however, a TAS was submitted that was 20 minutes faster than the unassisted run.
From January 12, 2020, the fastest unassisted speedrun of Donkey Kong Country was 810 milliseconds faster than the best TAS due to a new trick that allowed Diddy Kong to grab a DK barrel and the throw the DK barrel near a hidden barrel at the beginning of the last level, which skips the entire last level and level-finish animation as well.
Some games may produce beneficial glitches if the inserted cartridge is manipulated, which may not be reproduced on an emulator for a TAS. One of the most famous examples is , where lifting the side of the cartridge may allow the player to walk through solid walls.
However, due to potential benefits for either kind of speedrunning, it is not uncommon for speedrunners of both types to collaborate. Unassisted speedrunners can provide their expertise on the subject and receive new points of reference in return. A number of unassisted speedrunners have also made complete TASes, and vice versa.

Degree of glitch abuse

One of the most important differences between a tool-assisted and unassisted run is the use of glitches in the game. Though glitch use is often prevalent in unassisted runs, tool-assisted speedruns often make much heavier use of them. This may in part be because the majority of glitches are very difficult to exploit without frame-precision and re-recording. In some cases a trick relies not only on precise timing, but on several variables in memory also having a specific state, which would be nearly impossible to recreate in real time and without detailed knowledge of the game program.

Entertainment factor

These differences also lead to different expectations from tool-assisted and unassisted speedruns. Taking damage when doing so does not save time and/or is not required may look sloppy in a tool-assisted run, while being hit by the occasional hard-to-avoid enemy in a relatively long unassisted speedrun would not prevent the runner from holding their world record title. After the advent of frame-advance, frame-precise movement has also come to be expected, the lack of which may be characterized as sloppy play. Another difference is in the standards of use of waiting time in the speedrun: in situations where it is not possible to make the game move faster, and the player has to wait, such as in autoscrolling or any other areas of a game in which the runner does not have control over the speed, the runner is advised in TASVideos guidelines to do something entertaining for the viewers. An example of this is the gathering of 99 extra lives in the autoscrolling sections of the famous Super Mario Bros. 3 speedrun. In unassisted runs, players usually would not risk dying and having to start over to entertain the viewer.
Runs that prove unentertaining may get rejected for publication, even if the run itself is technically optimized. A bad game choice may contribute to a lack of entertainment. In this context, a "bad game" may represent a goal choice that does not demonstrate the merits of tool-assistance, so choosing a different goal may alleviate this issue. In other cases, such as the Excitebike TAS by Thomas Seufert, a previously unpopular game had achieved notable entertainment boost due to massive improvements brought into play by increased tool-assisted precision.
When someone submits a finished movie file of their input data for publication on the TASvideos website, the audience will vote on if they find the movie entertaining or not. According to their website, movies that stick with their site rules and have an 80% Yes Vote rate is a sign to say that the audience are interested in the movie and is more likely to be accepted or obsolete the current published movie and have the movie published on their website.

Verification of unassisted speedruns

Because tool assisted speedruns can account for all aspects of the game code, including its inner workings, and press buttons precisely and accurately, they can be used to help verify whether an unassisted speedrun record is legitimate. In 2018, Todd Rogers' record for Dragster was removed from Twin Galaxies and the Guinness World Records after a TAS experiment proved that his time, 5.51 seconds, was impossible to achieve even in a TAS.
Some players fraudulently recorded speedruns, either by creating montages of other speedrun or altering the playing time, posting them as TAS or RTA.
One of the best-known cases is Billy Mitchell, who had his Donkey Kong and Pac-Man Guinness records revoked in 2018, considering that he used an emulator. However, in 2020, the organization reinstated Mitchell's records after further investigation.
In 2018, Todd Rogers' record for Dragster was removed from Twin Galaxies and Guinness records after an experiment showed that his 5.51-second time was impossible to achieve, even in a TAS.
Another fraudulent RTA case is Badabun, where Tavo Betancourt streamed a Super Mario Bros. speedrun, finishing it at 05:12. Later, it was discovered that he was only pretending to play a series of speedruns from other youtubers. Youtuber Karl Jobst, after analyzing the video, called it "the worst fake speedrun on YouTube". The stream has been parodied by several youtubers of the genre, including Kosmic, holder of the current RTA world record for Super Mario Bros..

Notable speedruns

In the context of tool-assisted speedrunning, many common terms, usually neologisms, have been created. These terms are necessary to understand most general discussions about the phenomenon. This list covers the most ubiquitous terminology. Note that some words may have a different typical meaning outside of the lexicon of tool-assisted speedrunning; for example, frame applies to movies as well as to video games, but only the latter has relevance in this case.
;Arbitrary Code Execution
;Category
;Emulator
;Input
;Input file
;Frame
;Frame advance
;Glitch
;Hex editing, Binary editing
;Lag
;Luck manipulation, RNG manipulation
;Re-recording
;ROM
;Savestate
;Slow motion
;TAS
;Timeattack