Submission #5243: Weatherton's N64 Mario Kart 64 in 20:33.32

Mario Kart 64 has no unlockable cups, courses or characters. Players can turn on the game and immediately play the hardest engine class (150cc) with the best characters. There is one unlockable mode though; Extra. If you play Grand Prix (GP) mode and get a gold trophy on all four cups, you unlock Extra mode. The credits also roll (although this happens even if you play only Special Cup). Extra mode is just mirrored versions of every course so they don't actually differ in any meaningful way from the other modes with one exception, Toad's Turnpike has traffic driving against you rather than with you. Unfortunately, Extra mode is only at the 100cc engine class. So, Extra is actually a slower and easier mode than 150cc. The only other unlockables are course ghosts in Time Trials.

So, by playing in GP mode I can unlock gold trophies (which are stored to the cartridge), 150cc is the fastest and most challenging engine class and upon getting gold trophies on all cups, I unlock the only hidden mode in the game. By finishing on Special Cup I also get to see the credits. Two player GP is also an option and it may be faster (note that both players must finish every race, so you can't just use one player to jump off of). However, the strategies would differ enough to warrant a separate branch as players could trade places to always have access to better items and could hit one another.

Thus, to be completely clear what the goals are of this run, the branch name is '1 player GP 150cc all cups'. The order of the words is also consistent with the menuing in the game. An alternate title could be '1 player unlock extra mode' but this is the same number of characters while providing less information about the mode being played.

I've deleted the branch name to be consistent with the published SMK run. I also began a discussion in the Mario Kart 64 thread about possible branched.

I first played Mario Kart 64 on February 10, 1997 (when it was released in North America) and completed the game on 150cc soon thereafter. The game could be completed quickly, but there was a depth to it that I was drawn to. With four player and time trials, I found myself consistently playing the game year after year. In 2002 I joined the Mario Kart 64 Players' Page, encouraged to do so by my friend David Hehn and invited in with open arms by the webmaster, Kevin Booth. In 2004, I began experimenting with using emulators and save states to create superplay style movies. My friend Martin McMullan and I used standard emulators (without input recording) and screen capture software to record segmented movies, from save state to save state. The results were unlike anything we had seen before. We were using CPUs as strategic launch pads to hop around the levels in fun and surprising ways. In November of 2005, the world suffered a loss when Martin took his life. I was devastated to lose a friend and I commented that I would like to someday finish a 150cc TAS in his honor. In March of 2007 I collaborated with Comicalflop and Michael Fried to complete Luigi Raceway in 37"18. My contribution came on the last lap where I jumped off of Wario, spun around the finish line and landed back on the road. At this time, I was still not completely comfortable with recording input in a movie file and using even basic tools to improve my output. I decided to practice so that my skills would improve while having a bit of fun. I completed three test runs: Star Cup (April 9, 2007); Special Cup (April 25, 2007) and Flower Cup (April 11, 2008). I even submitted the Special Cup run to TASVideos as an Any % completion but it was sadly rejected. I was now more comfortable with building a TAS, but I had not yet built my capabilities to the level necessary to achieve consistent excellence in such a challenging game.

On June of 2008, Peter T."Xenos" began TASing Mario Kart 64. He immediately demonstrated a mastery of the game with an ability to produce nearly perfect input (as demonstrated in his Choco Mountain non-shortcut fast lap completion) while also thinking outside the box and developing new strategies. Xenos noticed that toggling the control stick during a slide increased his speed slightly and he used this to beat Michael Fried's very strong Mario Raceway fast lap time. He then went on to demonstrate creative applications of red shells on Luigi Raceway and Mario Raceway and even found a new shortcut on Kalimari Desert that required a CPU jump. All this happened in just a few months and then he was mostly on to other projects. His precise runs, careful analysis and creativity were an inspiration for me though.

In August of 2009 I began working on improving Luigi Raceway. It's perhaps fitting that the 150cc TAS should start with one of the most challenging courses in the game. I was able to beat Xenos' 29"71 TAS, initially just barely with a 29"47. I continued working and also received some good ideas on the TASVideos forums from Kyman about the initial landing. After several versions, I achieved the 25"06 that is in this TAS. That run was something I was (and am) really proud of. I had to work really hard for it and without any assurance of success. In the end, I produced a TAS that was precise, surprising and fast. This single video has had several rounds of viral sharing on Facebook and Twitter that generated millions of views (for other people who didn't credit me). If I were to try it again now, with everything I've learned and all of the tools I've built I'm sure I could beat it, but not by a huge amount. It provided a pretty strong foundation for me to produce an all cups TAS.

In January of 2014, TASVideos was set to present the first "TAS Block". Micro500 had built an N64 bot and was looking for games that would console verify to show off. Mario 64 worked and so did Mario Kart 64. It just so happened that I live near where AGDQ is held each year. DwangoAC and micro500 realized this and asked me if I would like to commentate on some of my Mario Kart 64 runs. Here is the Mushroom Cup video which is the same version as the final submitted run. Earlier we had also shown the old Special Cup test run.

At AGDQ 2015, I didn't have any completed cups ready but I pitched a "Choco Mountain 4-Player Gauntlet" concept using the 4-player capabilities of his bot to set up a challenging scenario of items and boulders to later be played by real players. We played this at AGDQ 2015. Making the movie required lag compensated multi-track recording lua script and special build of BizHawk that micro500 made.

At AGDQ 2016, we showed Star Cup, which I completed just in time. The AGDQ 2016 deadline motivated me to focus again on developing my TAS after a long period of focusing more on the Lua script. We will probably show another cup or two at AGDQ 2017 so keep an eye out for that if you're interested in live commentary. For more detailed commentary though, look below.

I have worked with Beck Abney and abitalive to develop a Game Resources page in which we've documented aspects of Mario Kart 64 that are important to Tool Assisted Speedrunning. Go there for more extensive descriptions of relevant strategies and documentation of relevant game data. Please note that the page is (and will continue to be) a work in progress. For the purpose of this submission text, I will summarize a some relevant information here.

A MT is a small boost awarded after a player performs a complicated drift process used to build the boost. The MT is a core advanced driving technique and is a primary differentiator between beginner and advanced players. It can be done around corners as well as on straightaways (a "Straight Stretch Mini Turbos" or SSMT). A MT has seven unique phases which can be performed "frame perfect". The longer a player spends on each step beyond what is necessary, the more speed they lose while building the MT boost and the greater overall distance they travel. Thus, for a TAS it is important to understand the absolute limits of this technique. The following table summarizes the MT phase requirements:

Note that optimal MT spacing is situation dependent. Generally some spacing after untilting is beneficial (as some speed is lost while generating a new MT) but there are cases (around corners) in which a new MT is worth generating before even starting a glide turn.

Although generally it is best to avoid getting airborne (as time off the ground is time not accelerating), getting airborne can be useful especially for achieving shortcuts. Although you can jump in this game, it's not a very large jump. Big air can be achieved through the use of items, by jumping off of terrain or by jumping off of other players (in the case of GP mode, CPU players). When performing a "CPU jump" the larger the weight difference between characters, the further the lighter character can jump. A boost is necessary to obtain big air while not spinning out. Particularly large jumps are possible when the heavier character pushes the lighter character into the ground which provides a very large upwards boost to the lighter character when they are ejected. There are three lightweights in the game and all three have equal stats for time trials. The lightweights are by far the best characters to choose in nearly every situation and they are also the best choice for getting large CPU jumps. Yoshi and Peach are both the same weight and Toad weighs 22% less. Weight only matters when characters collide with one another.

The random number generator in Mario Kart 64 is based upon the same function used in Mario 64, which pannenkeok2012 describes very well in one of his youtube videos. Although the RNG is called many times per frame, presumably for things like generating exhaust and dust, it is also called by the item roulette process. When a player hits an item box, they have a period of a few seconds during which they can select an item by hitting the Z button. Thanks to the great work that abitalive and Beck Abney did in reverse engineering the random item function just a few months ago, we now know how the item function works.

The randomness for the item selection works as follows:
Formula: (RNGValue + ABRFrames + RaceFrames + LastItem) % 100

RNG Value: Returned value of random number call giving a random number from 0 to 99
A/B/R Buttons: Counter goes up each frame (29.97fps) that on of these buttons is pressed. Can go up by 1, 2, or 3 depending on how many of them are pressed at once (or more if using more than 1 controller).
Race Frames: Counter that goes up each frame (59.94fps) during a race (including startup with Lakitu and points screens)
Last Selected Item: Last item table number that was produced from the formula above

For additional details on this, see: Beck's item RNG calculator and abitalive's item RNG assembly code with comments.

Because of the way the item RNG works, it is possible to cycle the RNG rotation by using the A, B, and R buttons when they otherwise don't affect the driving (before the race starts, for example). So, I can use frames of A, B and R or "ABR frames" before a race to ensure that I get the item I want from on the first possible frame for the first item box (which is sometimes useful). After the first item, I am more limited on my ability to manipulate the ABR counter because it can take on 100 different values but I can only increase it without affecting speed while I'm in the air. So, for each MT I can either release R while in the air (which changes the count by -5 in most situations) or I can add B while in the air (which has no noticeable affect for brief periods but does slightly alter physics many frames later). My biggest weapon for item manipulation is the frame I press Z, which is why my Lua script contains an item bot (thanks in large part to micro500).

In May of 2014, Sean Sullivan was tracing the item rng code to try to help me understand what was going on and to determine whether the item I would get could be predicted. His conclusion was that predicting the item would be very challenging (as you could conclude from Beck and abitalive's later findings, in fact, it's not possible in BizHawk because we can't follow register values at the requisite sub-frame levels). On 5/28/2014 though, Sean did come across a large set of tables, 100 bytes long that seemed to list item values that were returned from the item RNG function. With this finding, I quickly copied the values from the ROM and built a table to calculate the values in the first eight columns. For the first time, we had the actual probabilities (assuming each row was as likely as the next) of receiving items. Mark Jones found that this first table was for GP mode, the next was for CPUs (which I confirmed by altering the values) followed by 2-player versus, 3-player versus, 4-player versus and finally battle mode.

abitalive later made a ROM hack to fill the expansion pack with output from the item RNG function. After running for several hours, he had millions of results and confirmed that all values are equally likely. Having said all of this, it's only possible for RTA runners to control the item RNG function in very limited ways, none of which are useful. For example, Beck was able to consistently get the same item by tapping through the menus as quickly as possible (including selecting Mario) and then holding the A button through the entire race setup so that Mario would consistently spin out and drive straight to the items. He had to wait for the item roulette to expire and give him and item. All of this was necessary to ensure he could get the same result from the item RNG function. In normal situations, an RTA runner would have no way to know what item they would get (beyond the probability for the place they are in). Knowing the item probabilities is extremely helpful and they are summarized in the following tables:

The full lookup table is helpful when manipulating ABR frames and this is summarized in a table on the Game Resources page.

In 2014, after AGDQ, I transitioned from Mupen to BizHawk (Mupen64Plus). I was convinced to do this by adelikat and micro500 with the promise of Lua scripting. On Mupen, I had had to use external tools to monitor memory addresses and the closest thing to scripting I was beginning to do was to use Auto Hotkey to play back basic input. With Lua, I was able to create an on-screen HUD with a variety of direct measures and calculations. I was also able to store optimal MTs of various types in a catalog and recall them into an input queue to be fed into the game. Over time, I added additional features including an item bot and waypoint markers.

The initial foundation for feeding input into the game was written by adelikat and I collaborated with micro500 on the item bot, though he wrote most of the code for it. Micro500 also wrote the recording functionality and has generally been a sounding board and teacher for me as I fumble through writing code. Beck Abney found the (very helpful) rivals memory address and added it to the script.

The current Lua scripts that partially automate optimal maneuvers and provide helpful information while TASing Mario Kart 64 are available at https://github.com/weatherton/BizHawkMarioKart64.

This year I began experimenting with TAStudio and encountered several crashing issues (which I believe have now largely been addressed by Feos). I did use TAStudio for updating the menuing to use 4 controllers and it was very helpful for that purpose. I look forward to using it more in the future.

As you can see, the item RNG odds are highly manipulated in my favor throughout the run. We run out of named numbers to use to describe such a large number directly.

When comparing this run to RTA, it is important to know that the times are about 2.5 minutes different. RTA runs start timing when Mushroom Cup is selected, power cycle the moment they finish the fourth race of a cup and stop timing the moment they finish Rainbow Road. In this TAS, timing starts at power-on, the game is power cycled the frame after the gold trophy is stored and the timing stops on the last input required to advance to the end credits (over a minute after the second-to-last required input).

While making the 4-player Choco Mountain Gauntlet TAS, I realized that all four controllers can be used while menuing. Because of the game's implementation of polling, if all four controllers press buttons they are read in order. With four controllers, I have enough sub-frame polls to complete every section of the menus in one frame a piece. This means that I can move the cursor to the entry I want and select and confirm it all in one frame. The one exception is the character select screen where only controller one is functional. Controller one is also the only controller that is functional while in 1 Player GP mode. It is not possible to use the other controllers to manipulate RNG or pause the game or anything.

Although all 16 races in this submission are faster than any previous completion, no run is perfect. Over the extended development time of the TAS, I improved my TASing capabilities, switched to BizHawk and iteratively developed a robust Lua script (with contributions from micro500, Beck Abney and adelikat). Our knowledge of the game also improved over this period. Relevant knowledge includes an understanding of how CPU rivals are chosen, how item RNG works (for both the player and the CPUs) and how to directly manipulate it, and how more ways to cancel collisions to achieve large jumps. If an improvement is worked on, Toad should be chosen for Mushroom cup as this will set Bowser as the primary rival. With Bowser as the rival, I could jump off of him on lap 3 of LR directly to the finish line and I would not have to wait as long for him as I did for Wario on lap 3 of KD. KTB could, in general, be better executed as I somewhat rushed this level so that I could see whether the lap skip on KD was possible with Yoshi.

I could have used ABR frames to get some items a few frames earlier. I also have a new strategy on Royal Raceway involving red shell item tumble cancellation. Speaking of cancellation, there is a new set of cancellation jumps that we are just beginning to understand. It is possible to cancel tumbles caused by certain obstacles by spinning out. This includes Chain Chomps (confirmed) and (theoretically) the KD train and TT vehicles. I say theoretically because it's still not been done but it seems plausible because those obstacles put you in the same state as the chomp. There is even one type of spin-out jump that I have personally experience during a 4-player match in October of 2015 in which I triggered the "Large feather jump" state that is still in the game code. The feather item was in the pre-release versions of the game and can be seen in some screenshots. The large version of the feather jump triggers such a massive jump that when I triggered it I jumped halfway across the Rainbow Road level. This has never been replicated but I did it by hitting a Chain Chomp and that state activator behaves in an unusual way so it's plausible that this could be reproduced. Keep in mind that this may only be possible on the three levels I mentioned before (KD, TT and RRd) as they are the only levels with this type of obstacle. Still, this large feather jump is similar to the Mario 64 TTC Upward $1000 Bounty because it has only happened once that we know of it requires a similar "bit flip" type of change to memory.

The members of:

Anyone who has provided encouragement to me on the forums, Twitch or chat.

Noxxa: Judging.

Noxxa: Here, we have a run that has been in the making, in various shapes and iterations, for many years. This TAS follows a history of almost a decade of TASing this game. The end result shows. The submission text excellently describes every trick available, every item manipulated, and every step performed, and the TAS itself is just a work of art. All of the various and precise tricks used make this run a marvel to watch. This is what years of building up a TAS looks like in the end. Accepting to Stars.

Spikestuff: You guys realize that Noxxa actually wrote "Accepting to Stars". You don't need to keep writing that it should be stared after it already got the Star.... Publishing.