MineLlama
Llama with Retrieval Augmented Generation
as A Decision Maker in
Minecraft
Abstract
MineLlama is a novel localized Large Language Model (LLM) framework designed to enhance
decision-making
in the sandbox game Minecraft, without relying on external APIs or ex-tensive datasets.
MineLlama operates on a two-tier system:
- a planning level that generates a sequence of interdependent subgoals to achieve a given objective
- an executing level that determines and implements the immediate actions based on the current state and the subgoals.
Relationship graph generation
The framework of executing module
Images and Videos
Belows are the results of several representive tasks achieved by MineLlama.
Case 1 : Cooking a beaf
Case 2 : Creating a stone pickaxe
Case 3 : Creating a wooden pickaxe