$Re^2$ Agent: Reflection and Re-execution Agent for Embodied Decision Making

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

Keywords: Embodied agent, Large language model

Abstract: Accurate and efficient decision-making is essential for robots operating in open-world embodied environments. While large language models (LLMs) have shown promise in generating task plans from instructions, existing single-shot prompting methods often fail to adapt to dynamic constraints and environmental feedback during execution. In this work, we propose a **Reflective** and **Re-execution** (**$Re^2$**) agent that integrates prior knowledge-driven static prompting with task rule extraction for iterative refinement. Starting from an initial knowledge-grounded prompt, **$Re^2$** agent executes an action sequence, reflects on environmental feedback, abstracts failure patterns into rules, and re-prompts itself with refined constraints. This closed-loop reasoning process enhances robustness and leads to progressive improvement in task performance. We evaluate our method on the Embodied Agent Interface (EAI) benchmark, achieving an average score of **81.36** (**86.35** on BEHAVIOR and **76.36** on VirtualHome), demonstrating the effectiveness of our approach in bridging the gap between language reasoning and grounded action. The code is available at *https://github.com/chenyang126/Re-2-Agent*.

Submission Number: 7