Voyager is an open-ended embodied AI agent built by MineDojo (led by researcher Jim Fan at NVIDIA) that plays Minecraft autonomously using GPT-4 as its reasoning backbone. Published in May 2023 (arXiv 2305.16291) and open-sourced under the MIT license, it introduced a new paradigm for lifelong learning agents that improve continuously without human intervention or model fine-tuning. The system is built around three core components. First, an automatic curriculum that proposes new exploration goals calibrated to the agent's current inventory and world state, always pushing toward unexplored territory. Second, an ever-growing skill library where each skill is stored as executable JavaScript code, indexed by an embedding of its description, enabling retrieval in similar future situations and composition of complex behaviors from simpler ones. Third, an iterative prompting mechanism that generates code for a task, executes it in Minecraft, collects environment feedback and execution errors, and feeds them back into GPT-4 for refinement over multiple rounds. Voyager interacts with GPT-4 exclusively via blackbox API queries, requiring no gradient updates or fine-tuning of the underlying model. This makes it model-agnostic in principle, though the published results are based on GPT-4. The agent runs against a live Minecraft Java Edition instance via the Mineflayer JavaScript API. In benchmark evaluations, Voyager obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key technology tree milestones up to 15.3x faster than prior state-of-the-art approaches. Crucially, it can transfer its learned skill library to a brand-new Minecraft world and solve novel tasks from scratch, demonstrating genuine generalization. As of May 2026 the paper has over 1,641 citations on Semantic Scholar, and the work directly influenced subsequent NVIDIA research including Eureka and GR00T. Key features: - Automatic curriculum that proposes goals based on current inventory and world state - Ever-growing skill library of executable code snippets indexed by semantic embeddings - Iterative prompting loop: generates code, runs it, collects errors, refines with GPT-4 - No model fine-tuning required: operates entirely via GPT-4 blackbox API queries - Skill transfer to new Minecraft worlds without retraining - Open-source under MIT license with full code and prompts on GitHub (MineDojo/Voyager) - Integration with Minecraft Java Edition via the Mineflayer JavaScript API
Free and open source (MIT license). Requires your own OpenAI API key for GPT-4 queries; costs depend on OpenAI API usage.