基於類別級6D姿態估測之機器人夾取

Abstract

6D物件姿態估測在機器手臂夾取任務中扮演關鍵角色。然而，過去基於深度學習的物件姿態估測方法大多是實例級 (Instance-Level)，限制了其在真實場景中機器手臂夾取的普遍應用能力。在本篇論文，我們採用類別級 (Category-Level) 物件姿態估測，此種方法不僅能夠估測物件的姿態還能夠估測物件的大小。另外，該方法在訓練過程中不需要事先準備精確且完整的物件三維模型，並具備對未見過 (Unseen) 的物件進行姿態估測的能力。因此我們提出類別級SegFormer用於6D物件姿態估測 (Category-Level SegFormer for 6D Object Pose Estimation, also known as CLSF-6DPE) 之方法，本方法結合YOLOv8物件偵測模型和嵌入一個共同的分支 (Shared Head) 的SegFormer來預測正規化的物件座標空間圖 (Normalized Object Coordinate Space Map, also known as NOCS map)，接著透過相似變換 (Similarity Transformation) 演算法來估測物件的6D姿態和大小。我們還整合CLSF-6DPE至機器人作業系統 (Robot Operating System)，並且設計一個直觀的圖形化操作介面，讓使用者可以輕鬆地執行機器手臂夾取任務。實驗結果顯示，所提出的方法在姿態估測任務中優於傳統基於CNN的模型，展現出更高的準確性，證明其在實際應用中的可行性。6D object pose estimation is essential for robot grasping. However, past deep learning-based methods were mostly instance-level, limiting their applicability in real-world scenarios. In this thesis, we apply category-level 6D pose estimation, which estimates not only the 6D pose of objects but also their sizes. In addition, this method does not require exact 3D object models during training and can estimate unseen object poses. Therefore, we propose a Category-Level SegFormer for 6D Object Pose Estimation (CLSF-6DPE). This method integrates the YOLOv8 object detection model with SegFormer, including a shared head to predict the Normalized Object Coordinate Space (NOCS) map, followed by a similarity transformation algorithm to estimate the 6D pose and size of the object. We have also integrated CLSF-6DPE into the Robot Operating System (ROS) and designed an intuitive graphical user interface, enabling users to easily perform robot grasping tasks. Experimental results show that the proposed method outperforms traditional CNN-based models in pose estimation, achieving higher accuracy and demonstrating its feasibility in real-world applications.