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

TL;DR

Voyager combines an automatic curriculum, an executable skill library, and iterative prompting to build an open-ended Minecraft agent without model parameter updates.

Why it matters

It demonstrates how environment progression, memory, and executable tools can produce compounding agent capability even before adding gradient-based RL.

Key findings

1. 01

   A persistent skill library can reuse successful behavior across tasks.

2. 02

   Automatic curricula help keep an open-ended agent near its learning frontier.

Scaling dimensions

Models, methods & benchmarks

Algorithms

Automatic curriculum

Benchmarks

Minecraft

Topics