MineDojo/Voyager
Voyager
Voyager is an autonomous embodied agent and lifelong learning framework that uses a large language model to explore virtual environments. It functions as a code-based action controller, translating natural language instructions into executable scripts to interact with its surroundings.
The system features an automatic curriculum generator that creates sequences of exploration goals to discover new items and behaviors without human intervention. It maintains a skill library manager that stores learned behaviors as reusable code fragments, which can be composed to execute complex tasks.
The framework incorporates a lifelong learning loop that utilizes feedback-driven code refinement and iterative prompting to debug programs using environment error messages. It also performs automated task decomposition to break high-level objectives into smaller, manageable sub-goals.
Progress is managed through checkpoint-based state persistence, allowing the agent to save and resume its learning state from a specific point.
Features
- Lifelong Learning Models - Provides a lifelong learning framework that continuously discovers and stores new skills as reusable code fragments.
- Lifelong Learning Frameworks - Functions as a lifelong learning framework that continuously discovers new skills and stores them as reusable code fragments.
- Agent Skill Libraries - Maintains a structured library of learned behavioral patterns as code fragments for reuse in complex tasks.
- Autonomous Agents - Integrates LLMs with memory and tool usage to autonomously navigate and interact with virtual environments.
- Automatic Curriculum Generation - Implements an automatic curriculum generator that creates exploration goals to discover new items and behaviors without human guidance.
- Code and UI Generation - Implements the capability to generate and iteratively refine executable code based on real-time environment feedback.
- Code-Generating Agents - Uses a large language model to iteratively write and execute code scripts to control an agent in a virtual world.
- Curriculum Learning Frameworks - Provides a framework for curriculum learning to systematically discover new items and behaviors in virtual environments.
- Embodied AI Exploration - Develops systems that autonomously explore virtual worlds to discover new items and learn complex environmental interactions.
- Skill-Based Lifelong Learning - Builds a permanent library of skills and behaviors that an agent can store and reuse over long periods.
- Automated Code Refinement Loops - Iteratively modifies generated scripts based on environment error messages to debug and improve program execution.
- Error-Correction Feedback Loops - Uses recursive feedback loops to provide structural error messages for iterative correction of generated code.
- Multi-step Goal Execution - Independently plans and carries out sequences of actions to achieve complex objectives by composing learned skills.
- Natural Language Code Generators - Translates high-level natural language instructions into executable scripts to control an agent.
- Action Controllers - Functions as a code-based action controller that translates natural language instructions into executable scripts for environment interaction.
- Automatic Curriculum Generation - Generates a sequence of exploration goals to maximize the discovery of new items and environments without human intervention.
- Agentic Goal Decomposition - Utilizes LLMs to recursively break high-level objectives into actionable sub-goals and executable skills.
- Code Refinement Loops - Refines generated programs by feeding environment feedback and error messages back into an automated refinement loop.
- Agent State Persistence - Saves the agent's progress and learned skill library to disk to allow resuming lifelong learning processes.
- Machine Learning State Management - Persists the agent's learning state and skill library to checkpoints to facilitate resumption of exploration.
- Autonomous Environment Exploration - Navigates and interacts with virtual environments autonomously to acquire new skills through exploration.
- Agent Action Frameworks - Open-ended embodied agent using language models.
- Agent Environments - Open-ended embodied agent framework for Minecraft-based tasks.
- Agent Frameworks - Embodied agent framework for complex environments.
- Agent Reflection - Open-ended embodied agent that learns and reflects through exploration.
- Autonomous Agent Frameworks - Lifelong learning agent that explores and acquires skills in Minecraft.
- Embodied Agents - Open-ended embodied agent capable of autonomous exploration.
- Embodied Agents and Tools - Agent framework for open-ended learning in simulated worlds.
- Manipulation and Control - Open-ended embodied agent for Minecraft tasks.
- Agentic AI - Listed in the “Agentic AI” section of the The Incredible Pytorch awesome list.
