We introduce Voyager, the first LLM-powered embodied lifelong learning agent in Minecraft that continuously explores the world, acquires diverse skills, and makes novel discoveries without human intervention. Voyager consists of three key components: 1) an automatic curriculum that maximizes exploration, 2) an ever-growing skill library of executable code for storing and retrieving complex behaviors, and 3) a new iterative prompting mechanism that incorporates environment feedback, execution errors, and self-verification for program improvement. Voyager interacts with GPT-4 via blackbox queries, which bypasses the need for model parameter fine-tuning. The skills developed by Voyager are temporally extended, interpretable, and compositional, which compounds the agent's abilities rapidly and alleviates catastrophic forgetting. Empirically, Voyager shows strong in-context lifelong learning capability and exhibits exceptional proficiency in playing Minecraft. It obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster than prior SOTA. Voyager is able to utilize the learned skill library in a new Minecraft world to solve novel tasks from scratch, while other techniques struggle to generalize. We open-source our full codebase and prompts at https://voyager.minedojo.org/.
核心贡献 · Key contributions
首个由大语言模型驱动的具身终身学习智能体，在 Minecraft 中无需人工干预即可探索世界并获取技能。 First LLM-powered embodied lifelong learning agent in Minecraft, exploring and acquiring skills without human intervention.
自动课程最大化探索，根据智能体状态和进展提出任务。 Automatic curriculum that maximizes exploration by proposing tasks based on agent state and progress.
不断增长的技能库，存储和检索可执行代码的复杂组合行为。 Growing skill library of executable code for storing and retrieving complex, compositional behaviors.
迭代提示机制，利用环境反馈、执行错误和自我验证改进代码。 Iterative prompting mechanism using environment feedback, execution errors, and self-verification for code improvement.
超越先前最优：3.3 倍独特物品、2.3 倍移动距离、15.3 倍更快解锁科技树里程碑。 Outperforms prior SOTA: 3.3x more unique items, 2.3x longer distances, 15.3x faster tech tree milestones.
零样本泛化到新世界中的新任务，利用已学技能库。 Zero-shot generalization to novel tasks in new worlds using learned skill library.
局限 · Limitations
缺乏视觉感知，依赖纯文本 GPT-4 API，限制空间任务。 No visual perception; relies on text-only GPT-4 API, limiting spatial tasks.
GPT-4 API 成本高，比 GPT-3.5 贵 15 倍。 High cost of GPT-4 API; 15x more expensive than GPT-3.5.
代码生成和自我验证偶尔失败，例如幻觉出不存在的物品。 Occasional failures in code generation and self-verification, e.g., hallucinating non-existent items.
自动课程可能因幻觉提出无法完成的任务。 Automatic curriculum may propose unachievable tasks due to hallucinations.
需要高级 Mineflayer API，无法直接与基于像素的低级方法比较。 Requires high-level Mineflayer API; not directly comparable to low-level pixel-based methods.