The Minecraft and Margarita
Alright folks, it’s been another year of you guys busting out bangers left right and center, and now it’s time for me to add a few droplets to this delicious cup of creator juice.
Before going into my main project, I’ve been messing around in Ableton these past few months and made some demos of my music. Check them out here if you’re into weird stuff: https://soundcloud.com/adrisj/sets/early-2021-demos
Now on to the nerdy crap. Disclaimer, if you aren’t big into Minecraft this will probably make no sense, sorry ‘bout that folks.
UPDATE: Finished Video
Part 1: What?
There are few things in this world that are guaranteed. Some may say it’s death and taxes. Others cite the the expansion of space time and thermodynamics, pointing to the inevitable the heat death of our universe. But on this mortal plane there exists one rule so steadfast, so certain, that it bends reality to meet its needs. This rule is without doubt the center of all online creation.
I now present to you, Rule 34.5:
If a method to beat Minecraft exists, somebody will beat it that way.
As far as I know, Randomly generated Minecraft worlds have never been beaten by a bot before. And I hope to change that.
Part 2: Context
Minecraft is obviously the most popular game. Over the last year, beating Block Game has become a viral phenomenon. Nowadays everybody and their grandmothers tune in to watch epic gamers perform this feat over and over again.
You have speedrunners valiantly pushing the game to its limits, online personalities modifying how the game works and beating it with creative limitations, and A Green Man who just got really lucky, I swear. There are hundreds of different ways of modify the difficulty and challenge of beating the game, and these different challenges have been thoroughly explored, feeling almost saturated at this point (just look up “I Beat Minecraft But” on YouTube and you’ll see what I mean).
We also have TAS runs. A TAS (or Tool Assisted Speedrun) is a speedrun where a user can finely control their inputs and use all the tools at their disposal to create a “frame perfect run”. This includes slowing time down and editing what the bot presses at each frame of the run. Minecraft TAS’s are capable of beating the game with absurd speed and performing precise tricks that would be near impossible for a human to perform. By nature, TAS runs are almost always static, in that the player inputs are always going to be the same. If a game has random elements that require reacting in different ways, a TAS effectively becomes useless. By default, Minecraft worlds are generated completely randomly, until you load a world with a seed (which is a number that determines world generation). This differentiates “Random Seed” and “Set Seed” runs, where the former involves running a brand-new never-seen-before world and the latter involves running a world you can investigate and practice in advance. TAS runs are exclusively in the “Set Seed” category since they cannot adapt to randomness.
So humans can easily tackle both random and set seed worlds, and bots can only tackle set seed worlds. But does it have to be this way? What if we could make a bot that can beat a random seed world? Such a bot would have to be capable of observing a random world and collecting all of the resources it needs, and it would have to adapt to the world around it. That’s impossible, right?
Part 3: The Bot
So there’s this Minecraft mod called Project Baritone. It’s a pathfinding bot that can be used to perform various tasks, like escaping the chaotic spawn region of the oldest anarchy server in Minecraft or doing parkour and placing blocks in mid air or using a network of bots to build a 11km by 11km structure (also on the oldest anarchy server in Minecraft). It’s incredibly powerful and a shining pillar of Minecraft Tech.
But as cool as it is, I think its true potential isn’t realized yet. With such a bot you can do so much more than just parkour or the automation of one simple task. You know where this is going.
Say you wanted a bot that mines diamonds for you. Baritone has a “mine” command built in. If you run it, baritone will mine whatever block you specify for you. But it will do exactly that: Just mine the block. It won’t make sure you have the proper tool for the job. If you run this command and don’t have any pickaxe equipped, tough luck, because the bot will just mine diamond ore with its hands and get nothing from it. A human would probably try to get an iron pickaxe first. In that case, let’s make a bot that gets diamonds, collecting an iron pick if it doesn’t already have one.
To simplify, let’s think about the steps involved in getting just a stone pickaxe. They are:
- Mine 1 log
- Craft 4 planks
- Craft a crafting table
- Place the table down
- Get 2 more logs
- Craft 8 planks
- Craft 4 sticks
- With 2 sticks and 3 planks, craft a wooden pickaxe
- Mine 3 cobblestone with the wood pick
- With 2 sticks and 3 cobblestone, craft a stone pickaxe.
Okay, that’s pretty simple. We’ve described a linear set of instructions on how to do this that anyone can (sort of) follow.
There are unfortunately 3 big problems here.
First of all, what if we ask the bot to get a stone pick and it already has a wooden pickaxe in its inventory? With this method, the bot will acquire a second wooden pickaxe, which would waste both time and resources.
The second problem is this system makes a big assumption: It assumes that the world stands still until the bot changes it. For instance, after placing down the crafting table the bot assumes it will stay there, when it could have been broken. Or, the bot may die and loses all of its items in the middle of this linear set of commands.
The third problem is that this is overly verbose. I have to know exactly how many materials are needed, and that’s a lot of numbers to keep track of in my head what if these instructions are much more complex I make a math error somewhere?)
We can fix two of these problems by describing the goal as a tree that is split into smaller tasks.
We start at the top and move down, then left to right (although right-to-left or any order in between each level works fine too). Every frame, the bot checks if a resource is acquired and traverses down the first branch that is not fulfilled. For instance, if at whatever point the bot is trying to get a stone pickaxe and it has 3 cobblestone in its inventory, it will skip the left most branch (with the wood pick) and move on to another resource (like the crafting table or 2 sticks). This fixes both problems mentioned before: It adapts to varying conditions and also can react to changes on the fly by re-checking the tree every frame.
This idea can be improved just a bit though, since we can see a lot of repetition here. Planks, sticks, and the crafting table repeat multiple times, which is repetitive and overly descriptive. If you’re a programmer, you probably know where this is going.
(Some diagram explanations: On the left side of the cobblestone is a wooden pick, indicating that a wood pickaxe is needed before mining the stone. For each pickaxe, a crafting table represents the need for a crafting table placed down somewhere nearby.)
This diagram represents the same exact tree as above, but with some improvements. For one, resources only need to be defined once and they can then be re-used. This is just like creating a function in programming, it makes it easier to design larger programs and also keeps the code clean.
Thanks to the power of abstraction, we can extend this to collecting diamonds. Here’s what a diamond collection tree looks like:
(Here we also have a small furnace icon near the iron ingot resource, indicating that iron ore needs to be smelted in a nearby furnace to get the ingot)
Notice that we only added three extra trees here to get to diamond. However, if we stuck with the “full tree” approach we would create quite a massive single tree.
A small problem is that in this stage, the bot can only work in a “Peaceful” world with no hostile mobs. If it encountered any it would probably just keep running into them. The fix for this isn’t super elegant, but my solution was to “interrupt” the user command to run survival tasks (which include performing a simple kill aura to push away enemies, running away from mobs when at low health, eating food when hungry, and dodging projectiles like arrows or fireballs).
So. With that in place, does this work? Can the bot mine diamonds and avoid enemies?
Guess that answers that question.
The bot can now mine diamonds. That was a great victory for sure, but once I got this figured out I asked myself: Can we push this idea further?
Surprisingly, yes.
Part 4: Beating the game
We need a few things.
- Go to nether
- Get blaze rods
- Get ender pearls somehow
- Locate the stronghold
- Fill in stronghold and enter the end
- Kill the dragon
Easy right? Right?
Well, let’s handle them one at a time.
Nether Portal
There are multiple ways to get to the nether. You can mine 10 blocks of obsidian with a diamond pick and build it yourself. Or you can go the speedrun route and use a bucket of water next to a lava pool. I tried both methods, and neither worked. Both methods use flowing water (to make obsidian you need water to come in contact with lava) and Baritone hates flowing water (it gets stuck). Instead we just have the bot build the portal with lava and water, but one block at a time (so no flowing liquid). Kinda like a 3D Obsidian printer.
Blaze Rods
Actually “kind of” simple. You just need to locate a nether stronghold (nether brick basically), find a blaze spawner and camp near there, killing blazes wherever they appear. Bots are pretty good at this because they can locate nether brick the moment its loaded into a chunk, which gives it an edge over a player. For combat, if the bot equips diamond armour and food it can tank a few hits before it eats some food and heals up.
Ender Pearls
Here lies an impasse: There are roughly three ways to get ender pearls: Hunting endermen, Piglin bartering, and Villager trading. Right off the bat I ignored villager trading because that would be another UI screen I’d have to sift through and the trading randomness doesn’t give me lots of confidence. Piglin Bartering was what I started with, and I made the bot collect gold, locate piglins and trade with them, but it took way too long because I was running Minecraft on version 1.16.4, which has 50% the drop rates in the version most speedrunners use, 1.16.1 (or 10% the drop rates if you’re Green Man). I ended up just biting the bullet and hunting for endermen in the overworld, since piglin trading is much more boring and equally dangerous to the run.
Locating the Portal
When throwing an eye of ender, the eye will always travel in the direction of the stronghold. However, it doesn’t say where the stronghold is located, so speedrunners often use triangulation to locate the portal which requires at least 2 eye throws (two intersecting lines converge to one point). There’s also the issue of accuracy where the human eye can’t remember the exact angle the eye traveled, so course corrections are often made along the way. This is where the bot gets to shine. A bot can always keep track of the eye’s travel direction to its exact coordinate and keep track of that information. This means the bot can create a near perfect line that the stronghold will intersect with at some point. Also, the bot can use Baritone to scan the world for blocks, so the moment the stronghold’s chunk is loaded in, the bot will know. This means that the bot only needs to throw ONE eye of ender to locate the stronghold.
Defeating the Dragon
Before entering the end, the bot sets its spawn near the end portal with a bed. It then fills in the end portal, jumps in and destroys all remaining end crystals by hand. This isn’t a big issue because the bot can effectively attack the crystals through walls, protecting itself from their blast. Once the crystals are dead, the bot attacks the dragon while it’s perching until it’s dead.
And that’s that.
Here’s an example of that in motion, minus the final part (which wasn’t developed at the time of recording)
Part 5: What’s next?
Imma make a YouTube video showcasing this puppy and will livestream this bot beating the game. It’ll be a first (I think?) in Minecraft history. Will do the video some time this week and the stream over the weekend or next week I hope.
Part 6: Shenanigans
The bot can do a few more things besides beating the game. Here’s a silly grab bag of tasks I worked on in the meantime:
- Collect signs and print the entire bee movie script in a straight line
- Run away from players, collect diamond gear, then search for players and beat em up
- Collect 400+ items, like Cake, TNT, Diamond Armour and Vines.
- A command based system where other players can use /msg to “Whisper” commands to the player.
Just look at that boy go. Anyway that’s about it, I’ve had a lot of fun making this and I need to go to bed, so it’s been fun yall.
Oh yeah, I need a picture of me with a bio again.
About me:
Flash games are my childhood
Camping is good
Send me wacky music please
I like curry
JoJo part 2 > JoJo part 3
I’m a lil dumb sometimes ngl
Thank you for coming to my Ted Talk, and good night folks.
