Reproduce “Scale-Plan: Scalable Language-Enabled Task Planning for Heterogeneous Multi-Robot Teams” 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

Piyush Gupta, Sangjae Bae, Jiachen Li, David Isele

Scale-Plan combines language-model task interpretation with a reusable PDDL-derived action graph to select only the actions and objects relevant to a heterogeneous multi-robot task. On the 49-task MAT2-THOR benchmark, it achieves 78% task completion, 85% goal recall, and 94% action executability, improving over the strongest comparison system by 25, 16, and 9 percentage points, respectively, though at a planning cost of 62.27 seconds on average.

Key insights

  • Action-domain filtering reduces the planning space before detailed language-based synthesis, while preserving dependencies through symbolic graph search.
  • The approach improves complete task success more than raw executability: ablations show 78% task completion with filtering versus 65–67% without it, especially on complex tasks.
  • Filtering does not solve all grounding and affordance problems; failures still arise from object localization, holding-state conflicts, closed receptacles, and incomplete recovery after execution errors.

Summary by alphaXiv.

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