Reproduce “Neuro-Symbolic Learning for Long-Horizon Task Planning Under Complex Logical Constraints” 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

Qiwei Du, Zitong Zhan, Shaoshu Su, Bowen Li, Yi Du, Zhipeng Zhao, Taimeng Fu, Sebastian Scherer and 2 more

iFlax trains an object-scoring model using feedback from the reduced planning problems that its own predictions create, and pairs it with three recovery strategies for when pruning stalls symbolic search. Across three planning benchmarks, it improves solve rates and planning time over prior learned-pruning methods; the system is also demonstrated in simulation and on a mobile-manipulation robot.

Key insights

  • Training the scorer on plans produced under its own pruning decisions addresses the mismatch between offline supervision and deployment-time search.
  • Object pruning and search recovery are complementary: the scorer reduces the search space, while recovery helps prevent early scoring errors from blocking planning and learning.
  • The approach targets high-level symbolic planning; it relies on predefined PDDL domains and does not jointly optimize continuous motion feasibility.

Summary by alphaXiv.

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