在複雜的室內環境中進行高效的物件搜尋，由於感測器範圍受限、窮舉探索成本高昂以及缺乏語義脈絡，對於移動機器人而言依然是一項重大挑戰。為解決上述問題，本研究提出了一種新穎的主動語義搜尋框架結合大型語言模型之粒子濾波與假設修剪方法（Particle lter and Hypothesis pruning with Large Language Models，簡稱 PH-LLM）。在估測目標位置方面，PH-LLM 利用源自大型語言模型（LLM）對於語義相關區域的常識先驗，並結合語義粒子濾波器（SPF）來估計目標物的機率分佈。在決策規劃方面，則透過自適應貪婪策略，以最大化預期成本效益來選擇最佳子目標點。此外，本研究亦推導並證明了在嚴格移動預算限制下的理論性能保證。實驗結果顯示，PH-LLM 在成功率（SR）與預期探測時間（ETTD）表現上，皆顯著優於現有最先進的基準演算法（如 L3MVN 與 TS-FSS）。;Ecient object searching in complex indoor environments remains a critical challenge for mobile robots due to limited sensor ranges， and the high cost of exhaustive exploration， and lack of semantic context. To address these issues， this research presents Particle lter and Hypothesis pruning with Large Language Models (PH-LLM)， a novel active semantic search framework. For estimating the target location， the PH-LLM utilizes a LLM-derived common-sense priors for semantically relevant regions and a Semantic Particle Filter (SPF) to estimate a target distribution. For decision-making， an adaptive greedy policy selects subgoals by maximizing the expected benet-cost. Furthermore， the theoretical guarantees under strict travel budget constraints are proved. The results show that PH-LLM signicantly outperforms state-of-the-art baselines (e.g.， L3MVN and TS-FSS) in both Success Rate (SR) and Expected Time to Detection (ETTD).
