近年來，隨著機器學習的快速發展，端到端學習模型變得非常受歡迎。在機器人學中，越來越多研究讓機器人以端到端模型學習複雜的任務。儘管使用端到端學習模型帶來很多優點，但其是否真的優於以已知知識制定規則的先前方法仍然是一個無法決定的問題。在此論文中，我們將重點放在機械手臂的應用和端到端學習模型與基於運動學的方法之比較上，以倒水任務為例。基於運動學的方法利用深度相機與深度神經網路做目標偵測和獲取其座標，並利用逆運動學計算軌跡規劃。而端到端學習方法則是以單個神經網路將深度圖像當成輸入並輸出機械手臂的關節角度來控制機械手臂完成倒水任務。我們將兩種方法放在靜態場景和動態場景進行比較，我們的實驗結果顯示此兩種方法都能達到相似的性能與需要相似的儲存空間，但端到端學習方法更適合處理複雜和動態的場景。

With the fast development of machine learning in recent years, learning-based end-to-end neural models become very popular. In robotics, more and more works use end-to-end models to let robots learn complex tasks. Despite the many advantages of end-to-end models, it is still nondecisive whether they can outperform prior approaches that leverage known knowledge as rules. In this thesis, we focus on robotic arm tasking and compare the learning-based end-to-end approach with a kinematics-based approach, using pouring as an example. In kinematics-based approach, object detection and their coordinates are obtained with deep neural networks and depth camera respectively, and arm trajectory planning is calculated with traditional inverse kinematics (IK). In learning-based end-to-end approach, a single neural network is developed that takes RGBD images as input and outputs joint parameters of the robot arm to control the pouring task. We compare these two approaches with two scenarios, static and dynamic. Our experiments show that both approaches can achieve similar performance and require similar storage space. However, the learning-based end-to-end approach can handle complicated and dynamic scenarios better.

中文摘要 I
Abstract II
致謝 III
Table of Contents IV
List of Figures V
List of Tables VI
Chapter 1 Introduction 1
Chapter 2 Related Works 5
2.1 Inverse Kinematics 5
2.2 You Only Look Once 5
2.3 Deep Residual Network 6
2.4 Long Short-Term Memory Network 8
Chapter 3 Approaches to Be Compared 9
3.1 Kinematics-Based Approach 9
3.2 Learning-Based End-to-End Approach 11
Chapter 4 Experiments 13
4.1 Environment 13
4.2 Data Collection 14
4.3 Evaluation 15
4.3.1 Static Scenario 15
4.3.2 Dynamic Scenario 20
Chapter 5 Conclusion and Future Works 26
References 27

[1] “End-to-end learning, the (almost) every purpose ML method” [Online]. Available: https://towardsdatascience.com/e2e-the-every-purpose-ml-method-5d4f20dafee4
[2] Staffan Ekvall and Danica Kragic, “Learning and Evaluation of the Approach Vector for Automatic Grasp Generation and Planning,” in Proceedings IEEE International Conference on Robotics and Automation.
[3] S. Ekvall and D. Kragic, “Receptive field cooccurrence histograms for object detection,” in IEEE/RSJ IROS, 2005.
[4] S. Ekvall, D.Kragic, and F. Hoffmann, “Object recognition and pose estimation using color cooccurrence histograms and geometric modeling,” Image and Vision Computing, vol. 23, pp. 943–955, 2005.
[5] Sulabh Kumra and Christopher Kanan, “Robotic Grasp Detection using Deep Convolutional Neural Networks,” in IEEE/RSJ IROS, 2017.
[6] Xinchen Yan, Jasmine Hsu, Mohammad Khansari, Yunfei Bai, Arkanath Pathak, Abhinav Gupta, James Davidson, and Honglak Lee, “Learning 6-DOF Grasping Interaction via Deep Geometry-aware 3D Representations,” in ICRA, 2018
[7] “Inverse kinematics - Wikipedia” [Online]. Available: https://en.wikipedia.org/wiki/Inverse_kinematics
[8] “Program inverse kinematics algorithms with MATLAB” [Online]. Available: https://ww2.mathworks.cn/discovery/inverse-kinematics.html
[9] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” in IEEE CVPR, 2015.
[10] “An Overview of ResNet and its Variants” [Online]. Available: https://towardsdatascience.com/an-overview-of-resnet-and-its-variants-5281e2f56035
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep Residual Learning for Image Recognition,” in IEEE CVPR, 2016.
[12] Sepp Hochreiter and Jürgen Schmidhuber, “Long Short-Term Memory,” Neural Computation, 1997.
[13] “Long Short-Term Memory - Wikipedia” [Online]. Available: https://en.wikipedia.org/wiki/Long_short-term_memory