2023 · NVIDIA, Caltech, UT Austin, Stanford, ASU · Intermediate
Voyager: An Open-Ended Embodied Agent with Large Language Models
Voyager autonomously explores Minecraft and builds a reusable library of executable skills.
Direct answer
What does Voyager: An Open-Ended Embodied Agent with Large Language Models contribute?
Voyager autonomously explores Minecraft and builds a reusable library of executable skills.
Background
The agent uses an LLM for curriculum generation, code-based skills, and iterative self-verification. It stores successful programs in a skill library and composes them for increasingly complex tasks.
Problem
The work addresses a central constraint in Embodied Agents: building systems that learn useful representations or actions while remaining general enough to transfer beyond a single demonstration or environment.
Core idea
Voyager autonomously explores Minecraft and builds a reusable library of executable skills.
Architecture and method
The agent uses an LLM for curriculum generation, code-based skills, and iterative self-verification. It stores successful programs in a skill library and composes them for increasingly complex tasks.
- Automatic curriculum
- Executable skill library
- Iterative prompting and verification
Results and impact
Voyager became a clear example of open-ended embodied learning driven by language-model planning and tool use.
Prerequisites
- Language models
- Planning
- Program synthesis
Recommended reading order
Read the explanation above, review the related topic pages, then use the primary-source links below to inspect the abstract, figures, experiments, and released implementation.
Primary sources
External links are provided after the context needed to evaluate the work.
Follow-up research
Related papers and concepts
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
GR00T N1 uses a dual-system architecture for language reasoning and continuous humanoid control.
V-JEPA 2: Self-Supervised Video Models for Physical Planning
V-JEPA 2 learns predictive video representations that support visual understanding and zero-shot robot control.
Ego4D: Unscripted First-Person Video Dataset
Ego4D is a large first-person video dataset for studying what people see, remember, manipulate, and anticipate.
PARTNR: Planning and Reasoning Tasks in Human-Robot Collaboration
PARTNR is a large benchmark for household human-robot collaboration with natural-language tasks.
Common questions
Frequently asked questions
What is the main idea of Voyager: An Open-Ended Embodied Agent with Large Language Models?
Voyager autonomously explores Minecraft and builds a reusable library of executable skills.
Why is Voyager: An Open-Ended Embodied Agent with Large Language Models important?
Voyager became a clear example of open-ended embodied learning driven by language-model planning and tool use.
What should I learn before reading Voyager: An Open-Ended Embodied Agent with Large Language Models?
Recommended prerequisites are Language models, Planning, Program synthesis.