Reproduce “MinePlanner: A Benchmark for Long-Horizon Planning in Large Minecraft Worlds” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

William Hill, Ireton Liu, Anita De Mello Koch, Damion Harvey, Nishanth Kumar, George Konidaris, Steven James

MinePlanner introduces a benchmark for long-horizon planning in object-dense Minecraft worlds, designed to challenge state-of-the-art planning systems with thousands of irrelevant objects. It demonstrates that current domain-independent planners universally fail on its medium and hard difficulty tasks, exposing critical scalability limitations in existing planning approaches.

Key insights

  • The 'object-dense' problem is a fundamental bottleneck, as state-of-the-art classical planners exhibit catastrophic failure in environments with a large number of objects, particularly during PDDL translation/grounding or search.
  • Minecraft effectively serves as a realistic and accessible proxy for real-world planning challenges, allowing researchers to focus on long-term symbolic reasoning by abstracting low-level visual and control complexities.
  • Existing task scoping methods and current lifted planning approaches proved insufficient to handle the scale and object density of the MinePlanner benchmark, indicating a need for more advanced abstraction and relevance determination techniques.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.