Voyager
Continually grows an executable JavaScript skill library in Minecraft via environment execution traces and a GPT-4 critic that decides task success.
The loop
- Proposes
A code-generating LLM writes asynchronous Mineflayer JavaScript control skills to accomplish curriculum subgoals.
- Judged by
The Minecraft simulator returns execution errors and observations; a second GPT-4 instance acting as critic decides whether the task succeeded.
- Steers
Verified skills are indexed into a persistent vector skill library and composed as subroutines for harder tasks.
Headline result
Unlocked 3.3x more unique items, traveled 2.3x longer distances, and reached key tech-tree milestones up to 15.3x faster than prior agent baselines.
What this does not prove
Relies on high-level Mineflayer pathfinding APIs rather than raw pixel-to-motor actuation.
How it generalizes the harness
A GPT-4 critic deciding task success gates admission into a growing, composable library of reusable JavaScript control programs.
Classification
| What it proposes | Program |
|---|---|
| Generative model | GPT-4 code generator and curriculum planner |
| Algorithmic authority | Mineflayer simulator runtime and a GPT-4 critic (self-verification) |
| Feedback | Verifier errors |
| What gets updated | Population of candidates and prompt context |
| When it runs | In continuous operation |
| Validity | Empirical (measured) |
| Evidence | Code released |
| Lineage | Original system |