大型語言模型（Large Language Models, LLMs）在自然語言理解與推理上展現強大潛力，成為智慧智能體（agent）研究的重要方向。然而，單靠LLM於具身智能體（Embodied Intelligence）中執行多步驟與目標導向任務，易受環境不確定性及空間複雜性影響，導致決策不穩。本研究提出結合LLM與符號規劃語言PDDL（Planning Domain Definition Language）的整合架構，並利用環境圖輔助決策與設計重新規劃機制。架構中，LLM負責語義理解與任務拆解，將指令轉為PDDL格式，經外部規劃器生成行動序列。環境圖提供物件位置與場景拓撲資訊，支援行動路徑與子任務規劃。當遇到執行異常（如物件缺失、目標無法達成），系統可即時重新生成或調整計劃，提升靈活性與錯誤復原能力。實驗於ALFWorld等模擬場景驗證本方法相較傳統方法在任務成功率、決策準確性與環境適應性上具有優勢。研究貢獻在於整合PDDL規劃、場景感知與語言提示，提升LLM於多層次智能體決策之可解釋性與決策穩定性。

Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and reasoning, making them a key focus for intelligent agent research. However, relying solely on LLMs for multi-step and goal-directed tasks in embodied intelligence is prone to instability due to environmental uncertainty and spatial complexity. This study proposes an integrated framework that combines LLMs with the Planning Domain Definition Language (PDDL) and incorporates scene graph to support decision-making and introduces a replanning mechanism. The LLM handles semantic understanding and task decomposition, translating instructions into PDDL-formatted planning problems, which are then processed by an external planner. The scene graph provides spatial information, such as object positions and scene topology, to assist with action path and subtask planning. When execution anomalies occur (e.g., missing objects, unreachable goals), the system can dynamically regenerate or adjust plans, improving flexibility and fault recovery. Experiments in simulated environments like ALFWorld show superior performance over traditional methods in task success rate, decision accuracy, and adaptability. This research contributes a systematic integration of PDDL planning, scene perception, and language prompting to enhance the interpretability and adaptability of LLM-driven multi-layer agent decision-making.

論文審定書 i
摘要 ii
Abstract iii
目錄 iv
圖次 vi
表次 vii
第一章：緒論 1
1.1 研究背景 1
1.2 研究動機 2
1.3 研究目的 2
第二章：相關研究 5
2.1 機器人代理與任務規劃 5
2.1.1 機器人規劃問題相關研究 5
2.1.2 ALFWorld機器人指示任務 6
2.1.3 ALFRED的基準模型 7
2.1.4 ALFWorld的基線模型 7
2.2 使用大型語言模型進行推理 8
2.2.1 LLM的問題設定和提示工程 8
2.2.2 AutoPlan 10
2.2.3 LLM-Planner 11
2.2.4 Auto-GPT 11
2.3 機器人環境圖 13
2.3.1 動態記憶與環境追蹤 13
2.3.2 路徑壓縮技術 13
2.4 規劃領域定義語言（Planning Domain Definition Language, PDDL） 14
2.4.1 從自然語言指令到結構化PDDL與自動規劃流程 16
第三章 研究方法 17
3.1 系統整體概述 18
3.2 各層次間的協作 19
3.2.1 資訊流動與層次分工 20
3.2.2 回饋機制與動態調整 21
3.3 高層次決策 21
3.3.1 高層次決策的任務解析 22
3.3.2 In-Context Learning（ICL）方式 24
3.3.3 PDDL方法 26
3.3.4 ICL 與 PDDL 方法比較 28
3.4 中層次規劃 30
3.4.1 中層次行動決策流程 31
3.4.2 場景圖理解 32
3.4.3 導航與互動行動之分離處理 33
3.5 低層次行動 34
3.5.1 低層次回饋與感知資料收集 35
3.5.2 場景圖 36
3.5.3 不同層級的場景圖使用策略 36
3.5.4 記憶模組的效益 37
第四章 研究結果 38
4.1 實驗設計與環境設定 38
4.1.1 ALFWorld資料集 39
4.1.2系統配置與開發環境 40
4.1.3 實驗參數設置 41
4.1.4 評估指標設定 42
4.2 實驗結果分析 44
4.2.1 整體任務成功率比較 44
4.2.2 子目標完成率分析 46
4.2.3 平均行動步驟數 48
4.3 實驗結果探討 49
4.3.1 整體成功率提升的因素分析 49
4.3.2 輔助場景圖與Replan機制 50
4.3.3 失敗任務分析 51
4.3.4 複合任務案例分析：環境圖於複雜跨任務推理中的潛在價值 54
第五章 結論 60
5.1結論 60
5.2未來展望 61
參考文獻 62
附錄 64
6.1使用環境圖任務執行紀錄 64
6.2未使用環境圖任務執行紀錄 66
6.3環境圖於textworld資料集任務執行紀錄 68

[1] J. Achiam et al., "Gpt-4 technical report," arXiv preprint arXiv:2303.08774, 2023.
[2] J. Wei et al., "Chain-of-thought prompting elicits reasoning in large language models," Advances in neural information processing systems, vol. 35, pp. 24824-24837, 2022.
[3] S. Yao et al., "React: Synergizing reasoning and acting in language models," in International Conference on Learning Representations (ICLR), 2023.
[4] M. Shridhar, X. Yuan, M.-A. Côté, Y. Bisk, A. Trischler, and M. Hausknecht, "Alfworld: Aligning text and embodied environments for interactive learning," arXiv preprint arXiv:2010.03768, 2020.
[5] E. Kolve et al., "Ai2-thor: An interactive 3d environment for visual ai," arXiv preprint arXiv:1712.05474, 2017.
[6] C. Aeronautiques et al., "Pddl| the planning domain definition language," Technical Report, Tech. Rep., 1998.
[7] M.-A. Côté et al., "Textworld: A learning environment for text-based games," in Computer Games: 7th Workshop, CGW 2018, Held in Conjunction with the 27th International Conference on Artificial Intelligence, IJCAI 2018, Stockholm, Sweden, July 13, 2018, Revised Selected Papers 7, 2019: Springer, pp. 41-75.
[8] M. Shridhar et al., "Alfred: A benchmark for interpreting grounded instructions for everyday tasks," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10740-10749.
[9] T. Brown et al., "Language models are few-shot learners," Advances in neural information processing systems, vol. 33, pp. 1877-1901, 2020.
[10] S. Ouyang and L. Li, "Autoplan: Automatic planning of interactive decision-making tasks with large language models," arXiv preprint arXiv:2305.15064, 2023.
[11] C. H. Song, J. Wu, C. Washington, B. M. Sadler, W.-L. Chao, and Y. Su, "Llm-planner: Few-shot grounded planning for embodied agents with large language models," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2998-3009.
[12] H. Yang, S. Yue, and Y. He, "Auto-gpt for online decision making: Benchmarks and additional opinions," arXiv preprint arXiv:2306.02224, 2023.
[13] M. Ahn et al., "Do as i can, not as i say: Grounding language in robotic affordances," arXiv preprint arXiv:2204.01691, 2022.
[14] M. Helmert, "The fast downward planning system," Journal of Artificial Intelligence Research, vol. 26, pp. 191-246, 2006.
[15] L. Wang et al., "Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models," arXiv preprint arXiv:2305.04091, 2023.
[16] Q. Wu et al., "Autogen: Enabling next-gen llm applications via multi-agent conversation," arXiv preprint arXiv:2308.08155, 2023.
[17] N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, "Reflexion: Language agents with verbal reinforcement learning," Advances in Neural Information Processing Systems, vol. 36, pp. 8634-8652, 2023.