Reproduce “TGPO: Trace-Guided Policy Optimization for Robot Task Planning via Verifiable Subgoal Generation” 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

Zhihong Liu, Yang Li, Renming Huang, Chendong Zeng, Cewu Lu, Panpan Cai

TGPO translates abstract household requests into grounded PDDL subgoals, which a symbolic planner uses to find feasible subplans. During training, an auxiliary model revises unsuccessful reasoning traces, and the policy generates subgoals from original and corrected attempts. On Abstract tasks, TGPO reports 70.2 ± 1.8% success, versus 22.4 ± 2.1% for SayPlan; this metric combines plan feasibility with LLM-judged task completion, not physical-world execution.

Summary by alphaXiv.

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