基於幾何之物體姿態估測與抓取偵測應用於備有三維視覺感測器之工業機器人隨機取物系統

Geometry-Based Object Pose Estimation and Grasp Detection for Industrial Robotic Random Picking Systems Equipped with a 3D Vision Sensor

王喻民(Yu-Min Wang)

指導教授 ： 連豊力

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202100608

摘要

本篇論文提出了一個工業機器人隨機取物系統，可用於拾起任意擺放的物體，物體姿態估測和抓取偵測的方法皆基於幾何，此研究主要的貢獻在於改善了物體姿態估測方法的可靠性，抓取偵測方法所需的處理時間可以減少和收斂，以及呈現了一個穩健的機器人隨機取物系統。
在藉由一個三維視覺感測器偵測是否有任何物體在工作空間中之後，採用了一個基於點對特徵的物體姿態估測方法，藉由一種有效率的投票機制可以獲得一些姿態，然後，在本篇論文中提出的驗證函式可以用來驗證估測姿態的正確性。
接著藉由在本篇論文中提出的排名函式判定整體最佳的抓取姿態可以偵測最適當的抓取姿態，為了使任意形狀的物體可以被成功拾起，目標物體和夾具的幾何性質需要被分析，並提出了一個基於目標物體的局部表面幾何之穩定性量測，除此之外，必須確保夾具與任何物體之間不會有碰撞，採用的碰撞偵測方法是基於表面法向量，然後，此系統可以規劃如何使機械手臂能夠到達抓取姿態並抓取目標物體，在機械手臂移動的期間可以最佳化目標物體的姿態以減少量測誤差的負面影響，並且如果有必要的話，可以連帶調整抓取姿態。
最後，本篇論文呈現一些實驗結果以證實提出方法的可行性和分析提出系統的性能，提出的系統可以藉由一個備有一台深度相機的工業機械手臂實作出來。

關鍵字

隨機取物 ； 立體視覺 ； 幾何 ； 表面法向量 ； 物體姿態估測 ； 點對特徵 ； 投票 ； 抓取偵測 ； 局部表面幾何

並列摘要

In this thesis, an industrial robotic random picking system for picking objects placed randomly is proposed. The methods for object pose estimation and grasp detection are both based on geometry. The main contributions of this research are that the reliability of the method for object pose estimation is improved, that the processing time that the proposed grasp detection method requires can decrease and converge, and that a robust robotic random picking system is presented.
After detecting whether there is any object in the workspace with a 3D vision sensor, a method based on point pair features is adopted for object pose estimation. By using an efficient voting scheme, a number of poses can be obtained. Then, a validation function proposed in this thesis can be used to validate the correctness of the estimated poses.
Next, the most appropriate grasp pose can be detected by determining the overall optimal grasp pose with the ranking function proposed in this thesis. In order that objects of arbitrary shape can be picked successfully, the geometric properties of the target object and the gripper need to be analyzed, and a stability measurement based on the local surface geometry of the target object is proposed. In addition, it must be ensured that there is no collision between the gripper and any objects. The adopted method for collision detection is based on surface normals. Then, the system can plan how to enable the robotic arm to reach the grasp pose and grasp the target object. During the motion of the robotic arm, the pose of the target object can be optimized to reduce the negative effects of measurement errors, and the grasp pose can be adjusted jointly if necessary.
Finally, some experimental results are presented in this thesis to verify the feasibility of the proposed methods and analyze the performance of the proposed system. The proposed system can be implemented with an industrial robotic arm equipped with a depth camera.

並列關鍵字

Random Picking ； Stereo Vision ； Geometry ； Surface Normal ； Object Pose Estimation ； Point Pair Feature ； Voting ； Grasp Detection ； Local Surface Geometry

參考文獻

[1: Pérez et al. 2016] Luis Pérez, Íñigo Rodríguez, Nuria Rodríguez, Rubén Usamentiaga, and Daniel F. García, “Robot Guidance Using Machine Vision Techniques in Industrial Environments: A Comparative Review,” Sensors, Vol. 16, No. 3, March 2016.

[2: Buchholz 2015] Dirk Buchholz, “Bin-Picking: New Approaches for a Classical Problem,” 1st ed., vol. 44, Switzerland: Springer, 2015.

[3: Du et al. 2019] Guoguang Du, Kai Wang, and Shiguo Lian, “Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning: A Review,” in arXiv: 1905.06658v1, May. 16, 2019.

[4: Fujita et al. 2019] M. Fujita, Y. Domae, A. Noda, G. A. Garcia Ricardez, T. Nagatani, A. Zeng, S. Song, A. Rodriguez, A. Causo, I. M. Chen, and T. Ogasawara, “What are the important technologies for bin picking? Technology analysis of robots in competitions based on a set of performance metrics,” Journal of Advanced Robotics, pp. 1-15, December 2019.

[5: Levine et al. 2017] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen, “Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection,” The International Journal of Robotics Research, Vol. 37, No. 4-5, pp. 421-436, June 2017.