Craftium: Bridging Flexibility and Efficiency for Rich 3D Single- and Multi-Agent Environments

Mikel Malagón, Josu Ceberio, Jose A. Lozano

TL;DR: Craftium is an easy-to-use and flexible framework for creating rich 3D single- and multi-agent environments while substantially improving the computational cost of alternatives.

Abstract: Advances in large models, reinforcement learning, and open-endedness have accelerated progress toward autonomous agents that can learn and interact in the real world. To achieve this, flexible tools are needed to create rich, yet computationally efficient, environments. While scalable 2D environments fail to address key real-world challenges like 3D navigation and spatial reasoning, more complex 3D environments are computationally expensive and lack features like customizability and multi-agent support. This paper introduces Craftium, a highly customizable and easy-to-use platform for building rich 3D single- and multi-agent environments. We showcase environments of different complexity and nature: from single- and multi-agent tasks to vast worlds with many creatures and biomes, and customizable procedural task generators. Benchmarking shows that Craftium significantly reduces the computational cost of alternatives of similar richness, achieving +2K steps per second more than Minecraft-based frameworks.

Lay Summary:

Artificial Intelligence (AI) systems usually require vast amounts of data to learn. For example, teaching a machine to read and write involves gathering millions of text examples; similarly, computer vision demands millions of images. Thus, if we want to teach machines to operate in the real world, we have to train these systems for hundreds of years of experience. As training AIs in the real world for hundreds of years is unfeasible (and dangerous), researchers often train them in simulated (virtual) worlds. However, these simulations are too simplistic or slow to run, making them impractical for many research scenarios.

In this work, we present Craftium, a tool that allows researchers to generate very rich and complex scenarios to train and analyze AI systems. You might find Craftium visually similar to Minecraft, with huge worlds with many biomes, animals, monsters, and complex tools and artifacts. However, Craftium differentiates from Minecraft in many ways, as Minecraft is a game for humans, while Craftium is built for AI systems. For instance, Craftium allows AI agents to play more than 2,000 frames per second faster than Minecraft, which translates into much faster training times, saving time and money for researchers. Moreover, Craftium also includes several features missing in existing 3D virtual worlds, such as the ability to create completely custom scenarios and to have multiple AI agents learning and interacting together. Finally, Craftium is open, free to use, modify, and distribute for everyone.

Link To Code: https://github.com/mikelma/craftium/

Primary Area: Reinforcement Learning

Keywords: environment, reinforcement learning, agent, embodied AI, open endedness

Submission Number: 10801