简体中文 · English

Voyager

Implements an automatic curriculum, execution feedback/self-verification, and a reusable executable skill library that grows over time. It does not reproduce the Minecraft world locally.

Paper information

Field	Details
Paper	Voyager: An Open-Ended Embodied Agent with Large Language Models
Institution	NVIDIA / Caltech / UT Austin / Stanford / ASU
First public release	2023-05-25
Author code	Released code.
Local adapter or method key	voyager
Local reproduction code	src/auto_research/agent_research/

Original paper summary

Background and main changes

Open-world agents need to keep choosing novel tasks, retain successful skills, and repair programs from environment errors. Voyager uses GPT-4 to generate curricula and code as its action space. Successful programs are indexed by description in a skill library, then retrieved and composed for new tasks.

flowchart LR
    W["World state and exploration progress"] --> C["Automatic curriculum"]
    C --> G["Generate executable program"]
    L["Retrieve from skill library"] --> G
    G --> X["Execute in environment"]
    X --> V["Error feedback and self-verification"]
    V -->|Failure| G
    V -->|Success| L

Key figure from the paper

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

\[ g_t=\operatorname{Curriculum}(s_t,\mathcal S_t),\quad z_t=\operatorname{Generate}(g_t,s_t,\operatorname{Retrieve}(\mathcal S_t,g_t)), \quad \mathcal S_{t+1}=\mathcal S_t\cup\{z_t\}\ \text{if Verify}(z_t)=1. \]

Offline and online results

In Minecraft, the paper reports 3.3× as many unique items, 2.3× exploration distance and up to 15.3× faster tech-tree milestones. The library transfers to unseen tasks in new worlds. No production A/B test.

Local reproduction

Run a fixed 120 PlanBench-mini episodes. New task families trigger generation, execution and self-verification, with success saved to the skill library. Later tasks with the same procedure retrieve it directly.

Metric	Voyager
Joint success	1.0000
Average cost	1.1200
Skills created / reused	12 / 108
Verification retries	12

auto-research agent-eval --method voyager --benchmark planbench-mini \
  --episodes 120 --memory-size 24 --seed 42

Stable metrics: classic-agent-mini-suites-seed42.json.

Scope

Retains curriculum, skills, verification and cross-episode reuse. Local skills are structured tool plans, not Mineflayer JavaScript. Minecraft, GPT-4 and embedding retrieval were not run.