在未知環境中自主導航是機器人技術面臨的一個基本挑戰，特別是在 PointGoal 任務中，機器人必須在沒有預先存在的地圖的情況下到達目標位置。本論文介紹了意識前沿分析模型 (CFAM)，這是一個基於學習的新穎框架，它利用改進的部分可觀察馬可夫決策過程 (POMDP) 和強化學習 (RL) 來實現這種環境中的有效路徑規劃。所提出的模型透過引入信念狀態感知前沿評估策略，對傳統的基於前沿的方法進行了創新，該策略以修改後的貝爾曼方程式為基礎，結合抽象來降低大規模部分可觀測領域的計算複雜度。

為了解決多樣化訓練資料的稀缺性並提高普遍性，使用域隨機化技術程式化地產生合成地圖，從而實現對修改後的 ResNet18 架構的穩健訓練，以估計邊界距離成本。這些合成環境引入了房間幾何形狀、感測器雜訊和佈局複雜性的變化，緊密模擬了真實世界的條件。採用分層規劃框架，將邊界選擇的高階決策與使用來自各種感測器（包括 2D/3D LIDAR 和 RGB-D 輸入）的佔用網格圖 (OGM) 的低階控制相結合。

所提出的方法在模擬和現實環境中都得到了廣泛的評估。使用 Habitat 模擬器對 Matterport3D 資料集進行的模擬結果表明，與 LAM、ANS 和 FBM 等最先進的基線相比，成功率和路徑效率有顯著提高。此外，真實世界的實驗證實了該模型在各種室內場景（包括開放空間、雜亂區域和長走廊）中的穩健性和適應性。 CFAM 不斷展示其推廣至前所未見環境的能力，凸顯了其作為現實世界自主導航任務的可擴展且有效的解決方案的潛力。

透過學習、規劃和分層控制的整合，本論文推動了機器人自主領域的發展，使代理人能夠在沒有先前地圖的情況下在複雜、部分可觀察的環境中做出明智的導航決策。這項發現為進一步探索信念感知決策以及使用合成數據增強機器人系統中的模擬到現實的轉移開闢了道路。

Autonomous navigation in unknown environments presents a fundamental challenge in robotics, particularly in PointGoal tasks where a robot must reach a target location without access to a pre-existing map. This thesis introduces the Conscious Frontier Analysis Model (CFAM), a novel learning-based framework that leverages a modified Partially Observable Markov Decision Process (POMDP) and reinforcement learning (RL) to enable efficient path planning under uncertainty. CFAM advances traditional frontier-based methods by integrating belief-state-aware frontier evaluation, formulated through a modified Bellman equation that incorporates abstraction to reduce computational complexity in large-scale, partially observable environments.

To support the learning process within this framework, the model requires large and diverse training data that realistically reflect a wide range of navigational scenarios. To address this, synthetic maps are procedurally generated using domain randomization techniques, enabling the model to generalize effectively across varied spatial configurations and sensor conditions. A modified ResNet18 architecture is trained on these synthetic maps to estimate the cost of reaching the PointGoal via candidate frontiers. These costs are integrated within a hierarchical planning framework, which combines high-level frontier selection with low-level motion planning using occupancy grid maps (OGMs) derived from multi-modal sensor inputs such as 2D/3D LIDAR and RGB-D data.

The effectiveness of CFAM is demonstrated through extensive evaluations in both simulated and real-world environments. Experiments on the Matterport3D dataset using the Habitat simulator show significant improvements in success rate and path efficiency compared to state-of-the-art baselines including LAM, ANS, and FBM. Real-world trials further validate the robustness and adaptability of the approach across diverse indoor settings, such as cluttered rooms, long corridors, and open spaces. CFAM consistently demonstrates strong generalization capabilities to previously unseen environments, underscoring its potential for scalable deployment in practical autonomous systems.

Through the integration of belief-aware planning, deep learning, and hierarchical control, this thesis contributes a unified approach for navigating complex, partially observable environments without prior maps. The findings also highlight the value of synthetic data in bridging the gap between simulation and reality, offering promising directions for future research in learning-based navigation.