In line with recent progress in AI Minecraft (https://voyager.minedojo.org/), will there be a public announcement of a fully autonomous AI system which can defeat the Ender Dragon before Jan 1, 2024? To establish an objective baseline of competence and consistency in achieving this task, I'll stipulate that such a system must beat the Ender Dragon in less than 150 minutes of in-game time (deaths allowed) in at least 1% of runs.
Systems which are pretrained on human data are allowed.
Clarification 6/28 10:37AM - The system must only use observational data that includes what a human player would be able to see while playing the game (including the debug screen), like the Minedojo simulator. Systems which have active access to external reference information (like a Wiki) are fine, as long as it does not include additional information on the specific world that the agent is playing in.
1,000🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ666 | |
| 2 | Ṁ530 | |
| 3 | Ṁ302 | |
| 4 | Ṁ234 | |
| 5 | Ṁ228 |
I see we are going with minedojo for criteria but this is a good sign since the 2021 run.
@Khoja The bot Baritone beat the Ender dragon in 2021 in over 3 hours, not within your 150 minute time limit. Other than time, does it fulfill your requirements? Watching the video, I infer that the bot uses chunk data as an input, so can mine directly to points of interest (similar to a human using x-ray textures, which would be cheating in a human speedrun). Is this acceptable? I don't know if Voyager is the same way.
@kenakofer the spirit of this market feels like it requires a more end to end approach. I think it would be helpful to tie the resolution criteria to the minedojo environment or equivalent, and I think it would be useful for the creator to specify what observational data are allowed.
@kenakofer Yeah, some clarification is in order. My intention making this market included systems whose observational data just includes what a human player can see while playing the game (including the debug screen). Systems which are trained on/have active access to external reference information (like a Wiki) are fine. The Minedojo simulator is the reference I'm primarily imagining.
@Khoja If you want to tie resolution to Minedojo, that's fine, but I strongly suspect Minedojo is using similar observational shortcuts (because nothing in the installation or setup implies it's doing image processing, correct me if I'm wrong), so I would recommend sticking to one criteria or the other.
Here is the modlist that Minedojo Voyager uses, just for awareness of other potential advantages:
Fabric API: Basic Fabric APIs.
Mod Menu: Used to manage all the mods that you download.
Complete Config: Dependency of server pause.
Multi Server Pause: Used to pause the server when waiting for GPT-4 to reply.
Better Respawn: Used to respawn close to your death spot
@kenakofer The observation space of the Minedojo simulator is detailed in section B1 of the paper. Beyond the RGB information and its inventory, it does have voxel information of 3x3x3 surrounding blocks, which I see as quite limited. It also has its coordinates, but this was also observable to a human player via the debug screen.
What is this?
What is Manifold?
Are our predictions accurate?
Why use play money?
Why use play money?Related questions
What is this?
What is Manifold?
Are our predictions accurate?
Why use play money?
Why use play money?