增強式自動編碼網路應用於隨機堆疊物件之分類夾取

Robotic Random Bin Picking and Classification System Using Augmented Autoencoder

陳健倫(Jian-Lun Chen)

指導教授 ： 李志中

國立臺灣大學/工學院/機械工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202100129

摘要

本文提出一套改善的隨機堆疊物件分類之夾取流程，藉由物件切割、姿態估測以及夾取點判定，完成在隨機堆疊場景當中夾取與擺放物件的任務。整個流程首先使用深度感知相機獲得隨機堆疊場景，接著將平面影像搭配物件切割系統把場景中的目標物件逐一取出，將物件從堆疊場景取出後便放入姿態估測模型做姿態的預測，同時找尋該姿態對應的夾取資訊，最後搭配深度影像經過夾取點的判定完成本文的夾取流程。
本文提出並改善的地方分別為針對目標物件以虛擬相機自動獲得物件各個視角的照片成為姿態資料集，節省對現實資料處理的時間成本；以領域隨機化(Domain Randomization)的方式訓練自動編碼網路(Autoencoder)成為增強式自動編碼網路(Augmented Autoencoder)，避免虛擬與現實環境產生的領域差異(Domain Gap)並作為姿態估測系統；將對應的姿態經過修正的夾取經驗轉移預測夾取資訊，再經過干涉篩選與穩定度的排序，獲得最終的夾取點預測。
為了驗證本夾取流程的效果，本文架設實驗環境與機械手臂系統實際運行夾取流程，統計不同流程的夾取成功率與速度並探討本流程的特色與優點。最終本文的夾取流程針對兩樣金屬物件在隨機堆疊場景當中的夾取成功率為89.285%，整體運算的時間為1.128秒。

關鍵字

堆疊夾取 ； 實例切割 ； 姿態估測 ； 自動編碼網路 ； 領域隨機化

並列摘要

A process is proposed to improve the robotic grasping and classification system in this thesis, in which the system first uses instance segmentation technique to segment the image of the object in clutter, then proceeded by pose estimation and finally applies collision detection process to output the optimal grasping position for the robot.
This thesis focuses on the establishment of the pose estimation process by using the augmented autoencoder which uses a virtual camera to automatically crop the poses of the target object for dataset and contains domain randomization to avoid domain gap between real and synthetic data.
In order to verify the effectiveness of the process, a robotic system is set up to perform the random bin picking. It is shown that the success rate of grasping two metal objects in clutter can be up to 89.285% and the computation time is 1.128 seconds.

並列關鍵字

Clutter Grasping ； Instance Segmentation ； Pose Estimation ； Autoencoder ； Domain Randomization

參考文獻

[1] N. Correll, K. E. Bekris, D. Berenson, O. Brock, A. Causo, K. Hauser, K. Okada, A. Rodriguez, J. M. Romano, and P. R. Wurman, "Analysis and observations from the first amazon picking challenge," IEEE Transactions on Automation Science and Engineering, vol. 15, no. 1, pp. 172-188, 2016.

[2] G. Du, K. Wang, and S. Lian, "Vision-based robotic grasping from object localization, pose estimation, grasp detection to motion planning: A review," arXiv preprint arXiv:1905.06658, 2019.

[3] 蔡承翰、洪國峰, "模擬器自動圖像生成及自動標註：以深度學習結合機器人隨機堆疊取料為例," in 機械工業雜誌第449期, July 2020, pp. 28-35.

[4] N. Liu, J. Han, and M.-H. Yang, "Picanet: Learning pixel-wise contextual attention for saliency detection," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 3089-3098.

[5] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, "You only look once: Unified, real-time object detection," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779-788.