Voyager is an open-ended embodied agent with large language models that operates in the Minecraft environment. It is the first agent of its kind to achieve lifelong learning without human intervention.
Voyager consists of three main components: an automatic curriculum, a skill library, and an iterative prompting mechanism. The automatic curriculum maximizes exploration by considering the agent’s progress and state.
It generates tasks based on the goal of discovering diverse things, similar to an in-context novelty search. The skill library stores and retrieves complex behaviors.
Each skill is indexed by the embedding of its description, allowing for easy retrieval in similar situations. Complex skills can be synthesized by composing simpler programs, enhancing Voyager’s capabilities over time and preventing catastrophic forgetting.
The iterative prompting mechanism incorporates environment feedback, execution errors, and self-verification for program improvement. It allows Voyager to learn from its mistakes and refine its skills.
Voyager interacts with GPT-4, a large language model, through blackbox queries, eliminating the need for fine-tuning of model parameters. In experiments, Voyager demonstrates strong lifelong learning capabilities and exceptional proficiency in playing Minecraft.
It outperforms previous state-of-the-art approaches by obtaining more unique items, covering longer distances, and achieving key milestones in the tech tree at a faster rate.
It also outperforms other techniques in generalizing to novel tasks in new Minecraft worlds. Voyager continually discovers new items and skills through self-driven exploration, surpassing baseline methods.
Overall, Voyager combines the power of large language models with embodied learning to create an agent capable of continuous exploration, skill development, and novel discoveries in open-ended environments like Minecraft.