Wang et al. (2024)

Find It

DOI

In Summary

Voyager is an LLM-driven embodied lifelong learning agent in Minecraft that explores continuously, acquires skills, and makes discoveries without human intervention. It combines an automatic curriculum that maximises exploration, an ever-growing library of executable code for storing and retrieving complex behaviours, and an iterative prompting mechanism using environment feedback, execution errors, and self-verification. It queries GPT-4 as a black box, requiring no parameter fine-tuning. The stored skills are temporally extended, interpretable, and compositional, compounding the agent's abilities and mitigating catastrophic forgetting. Empirically it obtains 3.3 times more unique items, travels 2.3 times further, and reaches key tech tree milestones up to 15.3 times faster than prior state of the art, and can carry its learned skill library into a new world to solve novel tasks that baselines cannot.

Reference