Voyager

Voyager uses an LLM backbone to autonomously decompose high-level Minecraft objectives into executable sub-tasks, then learns and caches successful skill implementations as reusable code modules. The system maintains a dynamic skill library that grows over time, allowing the agent to compose previously-learned skills to solve novel problems without retraining. This creates a cumulative learning loop where each solved task expands the agent's capability repertoire for future challenges.

Voyager decomposes complex Minecraft objectives into hierarchical subtasks by prompting an LLM with the current world state, available skills, and task description. The LLM generates intermediate goals and execution strategies, which are then grounded into concrete action sequences. The planner dynamically adjusts the decomposition based on execution feedback, re-planning when subtasks fail or when the environment changes unexpectedly.

Voyager maintains a searchable library of learned skills as executable code modules, indexed by semantic descriptions. When planning, the system retrieves relevant skills using embedding-based similarity search or LLM-guided retrieval, then composes them into execution plans. New skills are generated by the LLM, executed in the environment, and added to the library if successful. The library persists across episodes, enabling cumulative learning.

Voyager generates executable code (Python or Minecraft commands) from LLM outputs, executes it in a sandboxed Minecraft environment, and captures execution results (success/failure, observations, errors). Feedback from execution is fed back into the LLM planning loop to refine strategies. This creates a tight feedback loop where code generation, execution, and learning are interleaved.

Voyager constructs a structured representation of the Minecraft world state including entity positions, block types, inventory contents, and agent status. This state is encoded into natural language descriptions and/or structured data that can be consumed by the LLM planner. The observation system continuously monitors the environment and updates state representations, enabling the agent to react to dynamic changes.

When a generated skill fails or produces suboptimal results, Voyager uses execution feedback to iteratively refine the skill code. The LLM analyzes failure modes, generates improved versions of the skill, and re-executes in the environment. This process repeats until the skill succeeds or a maximum iteration limit is reached. Successful refined skills are added to the library for future reuse.

Voyager can pursue complex, long-horizon objectives (e.g., 'build a house') by decomposing them into intermediate milestones and tracking progress toward each milestone. The system monitors whether milestones are achieved and adjusts the plan if progress stalls. This enables the agent to maintain focus on distant goals while handling short-term failures and replanning.

Voyager can be configured to pursue tasks in a curriculum order, starting with simpler objectives and progressing to more complex ones. The system tracks success rates and adjusts task difficulty based on agent performance. Easier tasks help the agent build foundational skills that transfer to harder tasks, creating a natural learning progression.