Voyager: An Open-Ended Embodied Agent with Large Language Models
About
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/.
Key insights
What problem does VOYAGER solve in Minecraft?
VOYAGER is the first LLM-powered embodied lifelong learning agent in Minecraft that explores, acquires skills, and makes discoveries without human intervention via an automatic curriculum, a skill library, and iterative prompting.
How does VOYAGER differ from prior LLM agents?
Unlike ReAct, Reflexion, and AutoGPT, VOYAGER adds a skill library and automatic curriculum. It obtains 3.3× more unique items and unlocks wooden tools 15.3× faster across three trials in MineDojo.
What are the practical uses of VOYAGER's findings?
The skill library transfers to unseen tasks in new worlds and also boosts AutoGPT, demonstrating plug-and-play reusable code skills for LLM-driven embodied agents in open-ended environments.
What limitations do the authors acknowledge?
The GPT-4 API is 15× more expensive than GPT-3.5, the agent still gets stuck or fails self-verification, and the curriculum hallucinates unachievable tasks or invalid game mechanics.