wang2023-voyager paper

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

Guanzhi Wang et al. (8 authors; NVIDIA, Caltech, UT Austin, Stanford, UW Madison)

Transactions on Machine Learning Research, 03/2024 · 2023-05-25

notes by Claude Fable 5 · retrieved 2026-08-07

The founding exemplar of skill-library evolution: verification before persistence, frontier-aware task proposal, and skills indexed by purpose — demonstrated against notably handicapped baselines.

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

The canonical skill-library paper (NVIDIA-led, TMLR 03/2024, first posted May 2023): a GPT-4 agent in Minecraft that proposes its own tasks, writes executable code to accomplish them, and banks verified programs in an ever-growing library — “lifelong learning” implemented entirely in context and files, no weight updates. Read from 2026 its architecture is strikingly familiar, because the modern coding-agent stack inherited it: skills as description-indexed executable procedures, retrieval by embedding, environment feedback loops, an LLM critic gating what gets persisted. Three components carry everything: an automatic curriculum (GPT-4 proposes the next task from the agent’s current state and its completed/failed task history, under a standing goal of “discover as many diverse things as possible” — in-context novelty search, frontier-aware so tasks stay challenging but feasible), a skill library (each mastered skill is a JavaScript program over Mineflayer APIs, keyed by the embedding of its natural-language description, retrieved top-5 by similarity to the current plan; complex skills compose simpler ones), and an iterative prompting mechanism (up to four rounds of generate → execute → refine on environment feedback and interpreter errors, ending when a separate GPT-4 self-verification critic confirms task success — only then is the program committed to the library).

Results

Against re-interpreted ReAct, Reflexion, and AutoGPT baselines in MineDojo: 63 unique items in 160 prompting iterations (3.3× the best baseline), wooden-tool tech-tree level unlocked 15.3× faster, stone 8.5×, iron 6.4×, and Voyager alone reaches diamond (one of three runs). In a fresh world with cleared inventory, Voyager solves all four unseen tasks; no baseline solves any — and bolting Voyager’s skill library onto AutoGPT lifts it from zero to partial success, showing the library is a portable asset, not an entangled component. A limitations-section demo (Fig. 10) shows one capability the loop alone lacks: with a human supplying visual critique or curriculum, Voyager builds complex 3D structures (nether portal, house) — spatial feedback its text-only self-verification cannot provide. The ablations rank the parts: removing the curriculum costs 93% of discovered items, removing self-verification costs 73% (the largest of any feedback type), no skill library means plateau in later stages, and GPT-3.5 in place of GPT-4 for code generation costs 5.7× — capability of the base model gates everything else.

What the design gets right (2026 reading)

Three decisions proved durable. (1) Skills are code with a natural-language index — temporally extended, interpretable, composable, and immune to catastrophic forgetting because they live outside the model; this is the direct ancestor of skills in today’s coding agents. (2) Nothing persists without verification — the critic, not the generator, decides what enters the library; the ablation showing self-verification as the single most valuable feedback type is the paper’s most transferable empirical result. (3) The curriculum reads the frontier — next tasks come from current state plus the completed/failed record, not a fixed list; progress compounds because new skills build on banked ones.

Assessment