This text introduces Voyager, an LLM-powered embodied lifelong learning agent in Minecraft. Voyager is designed to continuously explore the world, acquire diverse skills, and make novel discoveries without human intervention. It consists of three key components: an automatic curriculum for maximizing exploration, a skill library for storing and retrieving complex behaviors, and an iterative prompting mechanism for program improvement. Voyager interacts with GPT-4 via blackbox queries, bypassing the need for model parameter fine-tuning. Empirically, Voyager demonstrates strong in-context lifelong learning capability and exceptional proficiency in playing Minecraft. It outperforms prior SOTA methods in terms of obtaining unique items, traveling longer distances, and unlocking key tech tree milestones. Additionally, Voyager is able to generalize its learned skills to solve novel tasks in new Minecraft worlds.
| Signal | Change | 10y horizon | Driving force |
|---|---|---|---|
| Introduction of Voyager, an LLM-powered embodied lifelong learning agent | From manual exploration and learning to autonomous exploration and learning | More advanced and capable autonomous agents | Advancements in language models and AI technology |
| Voyager’s automatic curriculum maximizes exploration | From limited exploration to optimized exploration | More efficient and comprehensive exploration | Improved algorithms and AI capabilities |
| Voyager’s skill library stores and retrieves complex behaviors | From limited skill storage and retrieval to comprehensive skill library | More diverse and extensive skill repertoire | Enhanced storage and retrieval mechanisms |
| Voyager’s iterative prompting mechanism improves program execution | From limited program improvement to iterative program improvement | More accurate and efficient program execution | Iterative feedback and self-verification mechanisms |
| Voyager interacts with GPT-4 via blackbox queries | From manual model parameter fine-tuning to bypassing the need for fine-tuning | More efficient and streamlined interactions with language models | Improved integration of AI systems |
| Voyager exhibits exceptional proficiency in playing Minecraft | From limited proficiency to exceptional proficiency in gaming | Higher gaming performance and efficiency | Enhanced learning algorithms and skill acquisition |
| Voyager’s skills are temporally extended, interpretable, and compositional | From limited skill capabilities to advanced and versatile skills | More flexible and adaptable skill development | Improved learning algorithms and skill composition techniques |
| Voyager shows strong in-context lifelong learning capability | From limited lifelong learning capability to strong capability | More effective and continuous learning in various contexts | Enhanced learning algorithms and adaptive mechanisms |
| Voyager utilizes the learned skill library to solve novel tasks | From limited generalization to improved generalization of skills | Better transfer of skills to new tasks | Enhanced skill library and generalization algorithms |
| Voyager serves as a starting point to develop powerful generalist agents | From specialized agents to more versatile generalist agents | More capable and adaptable AI agents | Advancements in AI research and development. |