| 摘要: | 隨著網路購物快速發展與社會消費型態改變，越來越多空間被用來作為臨時存放貨物的小型倉儲。然而，碎片化的空間如果使用人力進行管理，將會大幅增加成本；若改採完全自助的方式，又可能面臨安全性與信任不足的問題。為了解決這些挑戰，本論文設計了一套低成本的無人搬運車(Autonomous Mobile Robot, AMR)系統，讓機器人能夠自動管理倉儲狀態，並協助顧客取貨，達到無人化的操作流程。
本系統的設計目標在於降低成本與計算資源需求，因此在硬體方面，捨棄了高價的光學雷達與深度相機，改以成本較低的里程計(Odometry)、慣性測量單元(Inertial Measurement Unit , IMU)以及網路攝影機(Webcam)來進行定位。而在系統架構上，搬運車不需具備完整的本地智慧運算功能，而是將任務決策與路徑規劃交由雲端的大型語言模型(Large Language Model, LLM)處理。透過這樣的設計，搬運車只需回傳自身位置、障礙物資訊與目標位置，就能由 LLM 推論出最佳的任務順序與中繼導航點，進而產生極佳的運動軌跡。
由於減少感測器數量，本系統大量依賴如 AprilTag 等視覺標籤來協助定位與進行貨架對位，使搬運車能準確掌握自身在室內空間中的位置。導航過程中，透過 LLM 所選擇的任務順序與目標位置，結合LLM-A* 與 DWB 演算法來完成地圖中與搬運車周邊的路徑規劃。實際執行搬運任務的過程中，本系統設計了具備升降與夾取功能的夾爪用以執行取件動作，並建立重置機制，避免在連續操作時因結構累積誤差而造成偏移。同時，為修正里程計在長時間使用下的誤差累積問題，本系統也導入擴展卡爾曼濾波器(Extended Kalman Filter, EKF)，結合 IMU 與里程計資料進行定位融合，提升整體移動精度。
為了方便使用者操作與管理搬運車，本系統也設計了一套網頁式的操作介面，提供貨架、貨品、障礙物與任務等功能的管理。該介面具備跨平台支援，並與機器人作業系統(Robot Operating System, ROS)整合，可在電腦或手機等不同平台上操作此系統。實驗在模擬場域中進行驗證，結果顯示本系統能順利完成搬運任務，並具備穩定性與實用性
;With the rapid growth of e-commerce and the evolving demands of last-mile logistics, an increasing number of small indoor spaces are being utilized as temporary storage facilities. Manual management of such fragmented environments incurs high labor costs, while fully self-service models often suffer from limitations in safety and reliability. To address these challenges, this paper presents a cost-effective Autonomous Mobile Robot (AMR) system designed for automated object retrieval and storage management in small-scale warehouse scenarios.
The proposed system emphasizes low hardware and computational overhead. Instead of relying on high-cost sensors such as LiDAR or depth cameras, it utilizes wheel odometry, an Inertial Measurement Unit (IMU), and a monocular webcam for localization. Task scheduling and global path planning are offloaded to a cloud-based Large Language Model (LLM), which computes optimal task sequences and navigation targets based on the robot’s current state and environmental information.
For precise indoor localization and alignment, the system employs AprilTag visual fiducials, enabling accurate positioning relative to storage racks. Navigation is achieved using a hybrid approach that combines LLM-A* for global planning and the Dynamic Window Approach (DWB) for local obstacle avoidance. A lift-enabled gripper with an automatic reset mechanism is used for object handling, while an Extended Kalman Filter (EKF) fuses odometry and IMU data to reduce long-term drift and improve localization accuracy.
A platform-independent web-based interface is developed for task assignment, inventory tracking, and real-time robot monitoring. Experimental results in a simulated warehouse field demonstrate the system’s ability to perform pick-and-place tasks reliably with high stability and low operational cost, making it suitable for flexible deployment in autonomous storage applications. |
