6.8k
Stars
+83
Stars/month
0
Releases (6m)
Star Growth
+11 (0.2%)
Overview
Voyager 是首个基于大语言模型的具身终身学习智能体，专门设计在 Minecraft 环境中进行开放式探索和技能获取。该系统由三个核心组件构成：最大化探索的自动课程生成器、用于存储和检索复杂行为的可执行代码技能库，以及结合环境反馈、执行错误和自我验证的迭代提示机制。Voyager 通过黑盒查询与 GPT-4 交互，无需模型参数微调。它开发的技能具有时间扩展性、可解释性和组合性，使智能体能力快速复合增长并缓解灾难性遗忘问题。在实验评估中，Voyager 展现出强大的上下文终身学习能力，在获取独特物品数量上比之前最先进方法高出 3.3 倍，旅行距离长 2.3 倍，解锁关键技术树里程碑的速度快达 15.3 倍。该系统能够在新的 Minecraft 世界中利用学习到的技能库从零开始解决新任务，展现出其他技术难以匹敌的泛化能力。
Deep Analysis
Key Differentiator
First LLM-powered embodied agent that continuously explores, acquires skills as reusable code, and demonstrates lifelong learning in an open-world environment — 3.3x more unique items and 15.3x faster milestone achievement vs prior SOTA
⚡ Capabilities
- • LLM-powered embodied lifelong learning agent in Minecraft
- • Automatic curriculum for maximizing exploration
- • Ever-growing skill library of executable code
- • Iterative prompting with environment feedback and self-verification
- • Task decomposition for complex multi-step goals
- • Checkpoint resume for continuing learning sessions
🔗 Integrations
OpenAI GPT-4Minecraft (Fabric mods)Node.js (Mineflayer)Azure
✓ Best For
- ✓ Research on embodied AI agents and lifelong learning
- ✓ Studying LLM-driven autonomous exploration and skill acquisition
✗ Not Ideal For
- ✗ Production game AI or commercial Minecraft bots
- ✗ Teams without access to GPT-4 API budget
Languages
PythonJavaScript
Deployment
Local (requires Minecraft instance)Research environments
⚠ Known Limitations
- ⚠ Requires a running Minecraft instance with specific Fabric mods
- ⚠ Depends on GPT-4 API which can be expensive for long exploration sessions
- ⚠ Research prototype; not designed for production game AI
- ⚠ Complex setup requiring Python, Node.js, and Minecraft configuration
Pros
- + 首创的 LLM 驱动具身学习架构，实现了真正的开放式探索
- + 可解释和可组合的技能库，支持复杂行为的持久存储和复用
- + 无需模型微调，通过黑盒 API 调用即可获得强大性能
Cons
- - 严重依赖 Minecraft 环境，限制了在其他领域的应用
- - 需要复杂的安装配置过程，包括 Python、Node.js 和 Minecraft 实例设置
- - 依赖 GPT-4 API 调用，可能产生较高的运行成本
Use Cases
- • 自主游戏 AI 代理开发和测试
- • 具身人工智能和终身学习算法研究
- • 复杂环境中的自动化任务执行和技能积累实验
Getting Started
1) 安装 Python ≥3.9 和 Node.js ≥16.13.0，然后克隆仓库并通过 pip install -e . 安装 Python 依赖；2) 在 voyager/env/mineflayer 目录下安装 Node.js 包和编译 TypeScript 代码；3) 按照官方教程设置 Minecraft 实例和必要的 Fabric 模组，配置好后即可运行 Voyager 智能体