Voyager

A GPT-4-driven, lifelong-learning agent in Minecraft that sets its own goals and builds up a library of skills by writing code.

Voyager was released in May 2023 by Guanzhi Wang, Linxi Fan, Yuke Zhu, Anima Anandkumar, and colleagues at NVIDIA, Caltech, UT Austin, Stanford, and Arizona State University, later accepted at TMLR, with code open-sourced. It is an agent that runs in Minecraft without fine-tuning any model parameters, working purely by calling GPT-4. It has three core components: an automatic curriculum that proposes the next exploration goal based on the current state; a skill library that stores validated, successful behaviors as executable code, retrievable and composable by description; and iterative prompting, which feeds environment feedback, execution errors, and self-checks back to GPT-4 to revise the code. Because skills accumulate as code, they are easy to reuse and this also mitigates catastrophic forgetting. Compared with prior methods, Voyager obtains 3.3 times more distinct items, travels 2.3 times farther, and unlocks key tech-tree milestones up to 15.3 times faster, making it a representative example of the 'large model plus code skill library' style of embodied agent.

ExampleAfter learning to craft a workbench and a wooden pickaxe, Voyager stores that code in its skill library; dropped into a brand-new Minecraft world, it directly retrieves and composes these skills to tackle new tasks.

Also called

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

Related

Embodied Agent · Large Language Model · Minecraft Environments · Code as Policies · Curriculum Learning · Continual Learning

Sources

Voyager (arXiv 2305.16291)
Voyager 项目主页 (Chinese)
Voyager (OpenReview, TMLR)

As of

2024-03