Model Learning to Solve Minecraft Tasks

Yarin Benyamin, Argaman Mordoch, Shahaf S. Shperberg, Roni Stern

Keywords: Minecraft, Model Learning, Automated Planning

Abstract:

Minecraft is a sandbox game that offers a rich and complex environment for AI research.
Its design allows defining diverse tasks and challenges for AI agents, such as gathering resources and crafting items.
Previous works have applied both Reinforcement Learning (RL) and Automated Planning methods to accomplish different tasks in Minecraft.
RL methods usually require a large number of interactions with the environment, while planning methods requires a model of the domain to be available. Creating planning domain models for Minecraft tasks is arduous.
Algorithms for learning a domain model from observations exist, yet have mostly been used on planning benchmarks.
In this work, we explore the use of such algorithms for solving Minecraft tasks.
We focus on the task of crafting a wooden pogo stick and explore different ways to represent states in this domain. Then, propose an agent that learns domain models from observations --- either generated by an expert or collected online --- and uses them with an off-the-shelf domain-independent planner.

Submission Number: 6