#Terraria AI Development

1 messages · Page 1 of 1 (latest)

fossil patio
#

I'm curious if anyone would like to work on developing a Terraria AI. This is something I've been wanting to do for a while and would likely need some help with due to the difficulty of the large exploration space of terraria.

Initially, I considered a reinforcement learning approach for basic tasks. The major problem here will be in training, it is quite hard to do reinforcement learning in the Terraria environment. One example of a developer's documentation about the difficulties: https://github.com/dkoleber/TerrarAI/wiki/Why-Terraria-isn't-the-best-RL-environment
It's very hard to train on lots of data for reinforcement learning because you can't quite run Terraria at 1000x speed like in other implementations of reinforcement learning.
Simulation of the game itself could be useful here, but something that would also take quite a bit of work that I'd need some help with.

Secondarily, I'm looking at iterative prompting using LLMs. This had some sucess in Minecraft with 'Voyager' https://voyager.minedojo.org/. They basically describe the environment to GPT-4 and have it write code to execute different actions using the MineDojo framework.

For this, a fairly large framework or API would probably be required, which I'm excited to look into.

Any suggestions, feedback, or general interest would be appreciated, just reply or DM!

GitHub

A platform for reinforcement learning in Terraria. Contribute to dkoleber/TerrarAI development by creating an account on GitHub.

cursive abyss
#

I'm curious if anyone would like to work on developing a Terraria AI

Your post goes on details but can you share what do you mean with "a Terraria AI"? I've not heard that term before.

fossil patio
cursive abyss
#

oh, very nice

#

are agents allowed in the game?

vast stirrup
#

Yeah, going the gym route right sounds like the best right now, and leveraging LLMs to help get that started sounds like a great idea!