JARVIS-1: A Multimodal AI Agent that Excels at Minecraft Tasks

Minecraft, the popular sandbox video game, has been used as a testbed for artificial intelligence (AI) research, as it offers a rich and complex environment that challenges the AI agents to learn, plan, and act. However, most of the existing AI agents are limited in their capabilities and performance, especially when it comes to completing diverse and long-horizon tasks.

JARVIS-1’s multimodal language model consists of three components: a vision encoder, a text encoder, and a symbolic encoder. The vision encoder processes the pixel-level information from the game screen and extracts high-level features. The text encoder processes the natural language instructions or queries from the user and encodes them into semantic representations. The symbolic encoder processes the symbolic information from the game state, such as the inventory, the health, the position, etc., and encodes them into logical representations.

The multimodal language model then fuses the information from the three encoders and generates a latent representation, which captures the current game situation and the user’s intention. Based on this representation, JARVIS-1 can plan and execute a sequence of actions that can achieve the desired goal. JARVIS-1 can also learn from its own experience and improve its performance over time, by using reinforcement learning and self-play techniques.

The researchers evaluated JARVIS-1 on over 200 Minecraft tasks, ranging from simple ones, such as collecting wood or building a house, to complex ones, such as crafting a diamond pickaxe or exploring a dungeon. They compared JARVIS-1 with prior versions of AI agents, such as JARVIS-0 and JARVIS-0.5, and found that JARVIS-1 outperformed them on all tasks, especially on the complex and long-horizon ones.

JARVIS-1 achieved over 90% success rate on a variety of tasks, such as mining, farming, fishing, cooking, etc. It also achieved 12.5% success rate on the very challenging task of crafting a diamond pickaxe, which requires finding and mining diamonds, crafting a wooden pickaxe, a stone pickaxe, and an iron pickaxe, smelting iron ore, and combining the materials. This is a significant improvement over JARVIS-0 and JARVIS-0.5, which achieved 0% and 2.5% success rate, respectively.

JARVIS-1 represents impressive progress in building AI agents that can act in complex environments and handle diverse and dynamic tasks. The researchers hope that JARVIS-1 can inspire new applications and research directions in AI, as well as foster new collaborations and communities among AI enthusiasts and Minecraft players.