Neura Market
Tweet logo

Tweet

Free

First lifelong learning agent that plays Minecraft purely in-context

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)
LinksX

About Tweet

Voyager is the first lifelong learning agent that plays Minecraft purely in-context, powered by GPT-4. It continuously improves by writing, refining, committing, and retrieving code from a skill library. The agent uses an iterative prompting mechanism that incorporates game feedback and execution errors to refine programs, a skill library of code to store and retrieve complex behaviors, and an automatic curriculum to maximize exploration. Voyager rapidly becomes a seasoned explorer, obtaining 3.3× more unique items, traveling 2.3× longer distances, and unlocking key tech tree milestones up to 15.3× faster than prior methods. The project is fully open-sourced with code and paper available.

Key Features

Iterative prompting mechanism: incorporates game feedback, execution errors, and self-verification to refine programs
Skill library: stores, retrieves, and composes code for complex behaviors via vector database
Automatic curriculum: proposes suitable exploration tasks based on agent's skill level and world state
Uses GPT-4 for code generation and refinement
Fully open-source code, paper, and models

Pros & Cons

Pros
  • Open-source with full code and paper available
  • Achieves significantly better performance than prior methods (3.3x items, 15.3x tech tree speed)
  • Lifelong learning without gradient descent—uses in-context code execution
  • Modular skill library enables compounding capabilities over time
Cons
  • Requires GPT-4 API, which incurs cost and latency
  • Currently focused on Minecraft environment only
  • Relies on a specific JavaScript API (Mineflayer) for interaction

Best For

Autonomous exploration and skill acquisition in MinecraftResearch in lifelong learning and open-ended AI agentsCode generation and refinement for game playingBenchmarking AI agents in complex environments

FAQ

What is Voyager?
Voyager is a lifelong learning agent that plays Minecraft using GPT-4. It continuously improves by writing, refining, and retrieving code from a skill library, all in-context without gradient descent.
Is Voyager open-source?
Yes, the project is fully open-sourced. Code is available on GitHub (github.com/MineDojo/Voyag…) and the paper on arXiv.
How does Voyager learn?
Voyager uses three components: iterative prompting with game feedback and errors to refine code, a skill library to store and retrieve successful programs, and an automatic curriculum to propose tasks based on the agent's skill level and world state.
What environment does Voyager run in?
Voyager is designed for Minecraft, using a JavaScript API (Mineflayer) to interact with the game.