跳到主要内容Skip to main content

← 返回AI 论文← Back to AI Papers

Agent 与工具Agents & tools

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

问题:在开放世界 Minecraft 里,agent 能不能无人干预地持续探索、积累技能并完成新任务。
证据位置:摘要给出三组件:自动课程、可执行代码技能库、带环境反馈/执行错误/自验证的迭代提示;通过 GPT-4 黑盒调用,不微调模型;报告 3.3× 独特物品、2.3× 移动距离、15.3× 更快科技树里程碑。
作者自陈局限:单一 Minecraft 环境;依赖 GPT-4 API;技能库是可执行代码,权限与安全未讨论;开放世界长期泛化仍有限。
它没主张的:不是通用机器人 agent;不是真实世界部署;不是无需模型调用。
Problem: in open-world Minecraft, can an agent explore, accumulate skills and solve new tasks without human intervention?
Where the evidence sits: the abstract lists three components — automatic curriculum, an ever-growing skill library of executable code, and iterative prompting with environment feedback, execution errors and self-verification; it calls GPT-4 as a black box without fine-tuning; it reports 3.3x more unique items, 2.3x longer travel and up to 15.3x faster tech-tree milestones.
What the authors say about limits: single Minecraft environment; depends on GPT-4 API; the skill library is executable code with no discussion of permissions or safety; long-term open-world generalisation remains limited.
What it does not claim: not a general robot agent; not a real-world deployment; not free of model calls.

关键参数Key specs

证据位置 / evidence
摘要:三组件与 3.3×/2.3×/15.3×
环境 / environment
Minecraft
复现状态 / reproduction
未复现(只读摘要)

官方原文怎么说What the source says

the first LLM-powered embodied lifelong learning agent in Minecraft
1 https://arxiv.org/abs/2305.16291
an automatic curriculum, an ever-growing skill library of executable code, and a new iterative prompting mechanism
2 https://arxiv.org/abs/2305.16291
It obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster
3 https://arxiv.org/abs/2305.16291

资料与许可Source and licence

资料时间Source checked
2026-11-19
许可 / 类型Licence / kind
arxiv-preprint
许可说明Licence evidence
https://arxiv.org/abs/2305.16291
类目Category
Agent 与工具Agents & tools

论文里的数字怎么读How to read the numbers

先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards

去官网看原页Open the official page →

← 返回AI 论文← Back to AI Papers