$Re^2$: Reflective Rule Induction and Rule-Guided Refinement for Embodied Planning

Yang Chen, Hong-Jie You, Jie-Jing Shao, Xiao-Wen Yang, Ming Yang, Yu-Feng Li, Lan-Zhe Guo

Keywords: Embodied Planning, Vision-Language Models, Neuro-Symbolic Methods, Test-Time Adaptation

Abstract:

Embodied planning requires agents to translate high-level language instructions into executable behavior under physical, logical, and temporal constraints.
Although recent LLM- and VLM-based agents have shown promising reasoning abilities, their performance in embodied environments remains limited by partial observability, evolving world states, and strict action preconditions, making semantically plausible plans often fail during execution.
Moreover, interaction feedback is typically used only for local correction rather than accumulated as reusable procedural knowledge for future decisions.
To address this limitation, we propose $Re^2$, a closed-loop framework for test-time improvement in embodied planning through interaction-driven procedural memory.
$Re^2$ converts execution feedback into reusable rules, selectively reactivates task-relevant knowledge for subsequent planning, and continuously refines rule utility based on downstream outcomes.
Each rule is represented in dual form, combining natural-language guidance with symbolic structure for retrieval and refinement.
We evaluate $Re^2$ in both text-based and multimodal embodied planning settings.
$Re^2$ achieves a state-of-the-art overall score of 81.36 on EAI and improves the average performance on EB-Habitat from 33.0 to 51.3 (+55.5%) over the reflective baseline.
These results show that closed-loop procedural memory provides an effective mechanism for improving executability and robustness in embodied planning.

Submission Number: 20