Any House Any Task: Scalable Long-Horizon Planning for Abstract Human Tasks

ABSTRACT

Open world language conditioned task planning is crucial for robots in large-scale households. A key challenge remains scalability; performance degrades with increasing environment size and plan length. We propose Any House Any Task (AHAT), a planner optimized for long-horizon tasks given ambiguous human instructions. AHAT utilizes an LLM to map instructions and scene graphs into PDDL subgoals, which are solved via symbolic reasoning. We introduce TGPO, a novel RL algorithm that integrates external correction of reasoning traces. AHAT achieves significant gains over state-of-the-art methods, particularly in tasks characterized by brief instructions but requiring complex, multi-room execution plans.