Craftium: Creating Efficient Environments for Open-Ended and Embodied Agents Beyond Gridworlds

ICLR 2025 Conference Submission1122 Authors

Keywords: reinforcement learning, environment, embodied, open-ended, continual learning, meta reinforcement learning

TL;DR: Craftium is a framework for creating Micraft-like 3D environments for autonomous agents in general. It is flexible and rich while being easy to use and computationally efficient.

Abstract: Advancements in open-ended and embodied AI require highly adaptable and computationally efficient environments. Yet, existing platforms often lack the flexibility, efficiency, or richness necessary to drive progress in these areas. Research in fields related to open-endedness, such as unsupervised environment design and continual reinforcement learning, usually defaults to simplistic 2D grid environments, as more complex alternatives are either too rigid or computationally expensive. Conversely, in embodied AI, the field relies on fully featured video games like Minecraft, which are rich in content but computationally inefficient and offer limited customization for creating new tasks. This paper introduces Craftium, a framework based on the open-source Minetest game engine, providing a highly customizable, easy-to-use, and efficient platform for building rich Minecraft-like 3D environments. We showcase environments of different complexity and nature: from simple reinforcement learning tasks to a vast world with many creatures and biomes, along with a customizable procedural task generator. Conducted benchmarks show that Craftium substantially improves the computational cost of Minecraft-based frameworks, achieving +2K steps per second more.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1122