Paper - Voyager: An Open-Ended Embodied Agent with Large Language Models logo

Paper - Voyager: An Open-Ended Embodied Agent with Large Language Models

Free

An open-ended embodied agent that learns and explores Minecraft autonomously using LLMs.

FreeFree tier
Type
Open Source

About Paper - Voyager: An Open-Ended Embodied Agent with Large Language Models

Voyager is the first LLM-powered embodied lifelong learning agent designed for the Minecraft environment. It uses OpenAI's GPT-4 to continuously explore the world, acquire diverse skills, and make novel discoveries without human intervention. The system features three key components: an automatic curriculum that maximizes exploration, an ever-growing skill library of executable code for storing and retrieving complex behaviors, and an iterative prompting mechanism that incorporates environment feedback, execution errors, and self-verification for program improvement. Empirically, Voyager obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster than prior state-of-the-art methods. The full codebase and prompts are open-source.

Key Features

Automatic curriculum that maximizes exploration in Minecraft
Ever-growing skill library of executable code for storing and retrieving complex behaviors
Iterative prompting mechanism that incorporates environment feedback, execution errors, and self-verification
Powered by GPT-4 via blackbox queries without model fine-tuning
Temporally extended, interpretable, and compositional skills
Alleviates catastrophic forgetting through skill compounding
Open-source full codebase and prompts on GitHub

Pros & Cons

Pros
  • Achieves 3.3x more unique items, 2.3x longer travel distances, and 15.3x faster tech tree progression than prior SOTA
  • Learns without human intervention or fine-tuning of model parameters
  • Skills are interpretable, reusable, and composable
  • Demonstrates strong in-context lifelong learning capability
  • Open-source codebase and prompts available for reproducibility
Cons
  • Relies on GPT-4 API which is expensive and proprietary (blackbox queries)
  • Currently restricted to the Minecraft environment only
  • Performance depends on the underlying LLM's capabilities and API availability
  • May require significant computational resources for running experiments

Best For

Autonomous exploration and skill acquisition in MinecraftResearch in lifelong learning and embodied agentsBenchmarking LLM-based agent capabilities in open-ended environmentsLearning from scratch in new Minecraft worlds using previously acquired skills

FAQ

What is Voyager?
Voyager is the first LLM-powered embodied lifelong learning agent that explores Minecraft, acquires diverse skills, and makes novel discoveries autonomously using GPT-4.
How does Voyager work?
It uses three components: an automatic curriculum for exploration, a skill library of executable code, and an iterative prompting mechanism that uses environment feedback and self-verification to improve programs.
Is Voyager open-source?
Yes, the full codebase and prompts are open-source and available on GitHub (link in the paper).
What are the performance improvements reported?
Voyager obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks tech tree milestones up to 15.3x faster than prior state-of-the-art methods.
Does Voyager require parameter fine-tuning?
No, it interacts with GPT-4 via blackbox queries, bypassing the need for model parameter fine-tuning.