Voyager

LLM-powered autonomous agent that learns and explores in Minecraft without human input.

4.8 (5)

İnceleyen Daniel Nikulshyn·Güncellendi Mayıs 2026

Research Open Source Reinforcement Learning Game AI Minecraft LLM Agents Autonomous Agents

Genel Bakış

Voyager is a research project that uses large language models to drive an autonomous agent inside Minecraft. The agent sets its own goals, writes executable code to act in the world, and incrementally builds a library of reusable skills as it plays. It combines an automatic curriculum for open-ended exploration, an iterative prompting loop that refines code through environment feedback, and a growing skill library that lets the agent tackle progressively harder tasks. Over time, Voyager unlocks new tech tree milestones, gathers diverse items, and traverses more terrain than prior Minecraft agents. Voyager is primarily of interest to AI researchers, game AI developers, and hobbyists exploring embodied agents, lifelong learning, and LLM-driven decision making in open-world environments.

Temel özellikler

Automatic curriculum for goal generation
Iterative prompting with environment feedback
Growing skill library of executable code
LLM-driven planning and reasoning
Open-ended exploration in Minecraft
Research-oriented, open-source implementation

Kullanım senaryoları

Benchmark LLM agents in Minecraft

Researchers can evaluate LLM-driven autonomous agents on open-ended Minecraft tasks, comparing tech tree progress, item diversity, and exploration against prior baselines.

Study lifelong skill acquisition

Use Voyager's growing skill library and automatic curriculum to investigate how agents accumulate reusable code-based skills over long horizons without human supervision.

Prototype game AI behaviors

Game AI developers can experiment with LLM-driven planning and iterative code refinement to create autonomous NPCs that set goals and adapt via environment feedback.

Hands-on learning for hobbyists

Hobbyists exploring LLM agents can run Voyager to see transparent, inspectable code actions and learn how prompting loops and curricula drive open-ended exploration.

Artılar ve eksiler

Artılar

Open-ended, lifelong learning without human intervention
Builds a reusable skill library that compounds over time
Strong benchmark performance versus prior Minecraft agents
Transparent, code-based actions are easy to inspect

Eksiler

Requires access to a capable LLM API, which can be costly
Limited to Minecraft as the environment
Setup and tuning can be technically involved
Performance depends heavily on prompt and model quality

İncelemeler

4.8

5 puandan ortalama.

İnceleme bırakmak için giriş yap.

Sofia Lindqvist

Use it every day

Honestly didn't expect to like it this much. Growing skill library of executable code is exactly what I needed, and builds a reusable skill library that compounds over time. I do wish performance depends heavily on prompt and model quality, but I reach for it almost every day now and it just clicks.

Marcus Bell

Skeptical, then convinced

I went in skeptical — most tools in this space overpromise. It actually delivers on iterative prompting with environment feedback, and open-ended, lifelong learning without human intervention caught me off guard. Performance depends heavily on prompt and model quality is why this isn't a perfect score, still, I'd recommend giving it a real trial.

Liam O’Connor

Use it every day

Honestly didn't expect to like it this much. Iterative prompting with environment feedback is exactly what I needed, and strong benchmark performance versus prior Minecraft agents. but I reach for it almost every day now and it just clicks.

Nadia Petrova

Compared a few options

Evaluated this against two competitors. Where it wins: iterative prompting with environment feedback and builds a reusable skill library that compounds over time. On balance the feature set — especially automatic curriculum for goal generation — justifies the 5 stars for our use case.

Aisha Khan

Solid for our team

We rolled this out across the team last quarter and open-ended, lifelong learning without human intervention. Automatic curriculum for goal generation fits neatly into how we already work, and iterative prompting with environment feedback removed a step we used to do by hand. Performance depends heavily on prompt and model quality, which is the main caveat, but it has held up under daily use.

Sorular

Henüz soru yok — ilk soruyu sen sor.

Soru sor