Voyager: An Open-Ended Embodied Agent with Large Language Models

How you could get an agent to learn skills, and become smarter

A friend turned me on to this paper, which came out in 2023, and used GPT-4 to train Minecraft agents to do various tasks (see examples and details here).

Essentially, they setup a curriculum that tracked how a human would progress in Minecraft (first mine wood, then create a crafting table from it, then fight a zombie, etc.) and let it learn how to do that particular task. Once it had learned the task, the task was stored in a skill library. Then the agent was then asked to do something new, it could learn a new task AND/OR retrieve a skill learned prior from its library.

The LLMs' ability to rely on the skill library made it significantly able to cover more ground, and accomplish more things, with the LLM calling upon the skill library to function as a subroutine when pursuing some other goal.

Note that this has not been replicated since using newer models, and I’m curious to see how much more effective the current state of the art is than 2 years ago.