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

**Authors:** Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, Anima Anandkumar  
**arXiv:** [2305.16291](https://arxiv.org/abs/2305.16291)  
**Date:** 2023-05-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16291v1/

## Key points

- Voyager is the first LLM-powered embodied lifelong learning agent in Minecraft
- It autonomously explores its environment and acquires skills without human intervention
- Components include an automatic curriculum for exploration, expanding skill library with executable code, and an iterative prompting mechanism
- Interacts with GPT-4 through blackbox queries without fine-tuning model parameters
- Skills developed are temporally extended, interpretable, and compositional to accelerate abilities while mitigating forgetting
- Empirical results show exceptional lifelong learning capabilities and proficiency in Minecraft gameplay
- Outperforms prior techniques by obtaining more unique items, traveling longer distances, and achieving milestones faster
- Can apply learned skills in a new world to tackle novel tasks effectively
- Authors have open-sourced their codebase and prompts at https://voyager.minedojo.org/

## Summary

The paper "Voyager: An Open-Ended Embodied Agent with Large Language Models" introduces Voyager as the first LLM-powered embodied lifelong learning agent in Minecraft. This groundbreaking agent autonomously explores its environment and acquires a wide range of skills without human intervention. It comprises three essential components: an automatic curriculum for exploration, an expanding skill library containing executable code for complex behaviors, and an innovative iterative prompting mechanism to enhance program performance. Voyager interacts with GPT-4 through blackbox queries, eliminating the need for fine-tuning model parameters. The skills developed by Voyager are temporally extended, interpretable, and compositional, accelerating its abilities while mitigating catastrophic forgetting. Empirical results demonstrate Voyager's exceptional in-context lifelong learning capabilities and proficiency in Minecraft gameplay. Compared to prior techniques, it obtains 3.3 times more unique items, travels 2.3 times longer distances, and achieves key tech tree milestones up to 15.3 times faster. A standout feature is its ability to apply the learned skill library in a new world to tackle novel tasks from scratch—a feat that other methods struggle with effectively. The authors have generously open-sourced their complete codebase and prompts at https://voyager.minedojo.org/, providing a valuable resource for further research and development in this domain.