Minecraft's complexity and diversity as an open world make it a perfect environment to test if agents can learn, adapt, and tackle a variety of unscripted tasks. However, the development and validation of novel agents in this setting continue to face significant engineering challenges. This paper presents MineStudio, an open-source software package designed to streamline the development of autonomous agents in Minecraft. MineStudio represents the first comprehensive integration of seven critical engineering components: simulator, data, model, offline pre-training, online fine-tuning, inference, and benchmark, thereby allowing users to concentrate their efforts on algorithm innovation. We provide a user-friendly API design accompanied by comprehensive documentation and tutorials. Our project is released atthis https URL.

View on arXiv

Comments on this paper