AI researchers have unveiled Voyager, an AI bot capable of autonomously exploring and expanding its capabilities within the open-world game Minecraft.
Unlike other bots, Voyager possesses the remarkable ability to write its own code through trial and error, utilizing internal conversations with the advanced GPT-4 model. This experimental system represents an embodied agent, offering a glimpse into the future of artificial intelligence.
Minecraft serves as an ideal testing ground for such systems due to its lifelike representation of the real world, simple, intuitive rules and physics, and expansive open environment. Previously, the AI-powered Minecraft simulation framework called MineDojo provided a foundation for creating and fine-tuning AI models.
Voyager takes a high-performance approach, harnessing the power of GPT-4 to dynamically rewrite its code as it progresses. Not only does Voyager train its models against Minecraft data, but it also engages in internal conversations with GPT-4 to determine appropriate actions in response to in-game events.
For instance, when faced with the appearance of skeletons during nighttime, Voyager seeks guidance from GPT-4 by asking, “What would a skilled player do when monsters are nearby?” GPT-4 advises the agent to craft and equip a sword for self-defence while attacking the skeletons. This general advice is then translated into a specific goal, such as “Collect stone and wood, craft a sword at the crafting table, equip it, and fight skeletons.”
After accomplishing these tasks, Voyager registers them in its general skill library. This library enables Voyager to avoid relearning combat techniques when faced with similar challenges in the future, such as venturing into a deep cave to find iron ore. While Voyager initially relies on GPT-4 for such skills, it transitions to the faster and more cost-effective GPT-3.5 model, which focuses on teaching skills most relevant to specific situations.
Voyager’s unique traits empower it to discover more objects, learn a wider array of skills, and explore vast areas within the Minecraft environment, surpassing the capabilities of other bots. This self-improvement capability is not designed to replace human Minecraft players but rather to pave the way for relatively simple AI models to enhance themselves based on their own “experience.” By learning from their experiences, AI agents like Voyager have the potential to execute complex tasks in both virtual and real-world environments.
The research conducted on Voyager holds tremendous promise for the field of artificial intelligence. The ability of AI agents to evolve and adapt based on their experiences opens up new avenues for advanced task execution. Further research and development are expected to refine the self-improvement capabilities of AI agents, enabling them to perform increasingly complex tasks and contribute to real-world applications.