機器人3D環境感知與深度學習介紹

Introduction to Robotic 3D Environment Perception and Deep Learning

施志軒

《機械工業雜誌》 473期 (2022/08) Pp. 21-30

摘要

隨著近年來機器人研究的發展，機器人逐漸走出實驗室，並廣泛應用在堆疊搬運、生產製造、產線組裝等。然而，機器人面臨的問題更加複雜，且必須與環境進行互動，因此必須具備更高階的感知功能。深度學習處理三維點雲可分為三個大類別來探討，其包括三維形狀分類、三維目標偵測與追蹤、三維點雲分割三大任務。3D感測器所收集的點雲資料屬於三維資料可以表示空間中大量的幾何、形狀與尺度資訊，然而因為資料量大、資料無序性與不會隨著旋轉改變結果。因此PointNet作為第一個能直接從點雲中學習有用資訊，達到End-to-end學習，可用同一個網路架構完成點雲分類與分割。工研院機械所團隊推出研磨拋光機器人品牌RobotSmith，提供整體軟硬解決方案，並在機器人3D環境感知處理上展示兩個成功案例：研磨周邊系統定位與焊道研磨疊代補償。期許在這波3D深度學習浪潮中，以精實的技術能力，攜手台灣產業共創新頁。

關鍵字

機器人 ； 點雲 ； 深度學習統

並列摘要

With the development of robot research in recent years, robots have gradually left the laboratory and are widely used in palletization, manufacturing, production line assembly, etc. However, the problems faced by robots are more complex and must interact with the environment and therefore must have high-level perception capability to meet real-world challenges. Deep learning for 3D point clouds covers three major tasks, including 3D shape classification, 3D object detection and tracking, and 3D point cloud segmentation. The point cloud data acquired by 3D vision sensors can provide large amount of geometric, shape, and scale information. However, point clouds are disorder, unstructured, and rotation-invariant. To overcome these problems, the pioneering work PointNet is proposed to learn per-point features using shared Multilayer Perceptrons (MLPs) and global features using symmetrical pooling functions. PointNet network can deal with the point cloud classification and segmentation for end-to-end training. ITRI MMSL robotic group roll out a new brand of grinding and polishing robot called RobotSmith. RobotSmith provides software and hardware total solution and demonstrates two successful cases in robot 3D environment perception: grinding peripheral system localization and iterative grinding path compensation. We expect that RobotSmith can create a new era of Taiwan's industries with cutting-edge technologies.

並列關鍵字

Robot ； Point cloud ； Deep learning

參考文獻

邱琬雯、莊瀅芯、黃仲宏、葉立綸、葉錦清、熊治民，“2017 機械產業年鑑 ˮ，工研院產經中心 (IEK)，2017。

R. Siegwart and I.R. Nourbakhsh, “Introduction to autonomous mobile robots,ˮ The MIT press, 2004.

“A tutorial on 3D deep learning,ˮ CVPR2017, [Online].Available：http://3ddl.stanford.edu/CVPR17_Tutorial_MVCNN_3DCNN_v3.pdf

邱威堯、謝伯璜、張津魁、呂尚杰、張俊隆，“ 堆疊物件取放技術介紹 ”，機械工業雜誌，362 期，20-28 頁，2013。

Z.Q. Zhao, P. Zheng, S.T. Xu, and X. Wu, “Object detection with deep learning: A review,ˮ IEEE transactions on neural networks and learning systems 30(11), 3212-3232, 2019.