基於深度神經網路之同心管機器人運動控制與視覺追蹤應用

Abstract

同心管機器人（Concentric Tube Robot；CTR）是一種軟性連續型機器人，由具直徑差異之彈性預彎曲同心圓管相互插入組成，並透過旋轉與平移行為進行運動控制。因其具有小體積、高靈活性等特點，常被應用於醫療手術與工業探勘中。本研究設計了一組三支鎳鈦合金圓管具6自由度之同心管機器人，首先分析基於Cosserat Rod理論的同心管機器人動力學模型，發現端點均方根誤差（Root Mean Square Error；RMSE）約為機器人總長6.3%（21mm），其原因在於模型採用了過多簡化假設，例如材料均勻和忽略摩擦力等，導致在實際應用中的誤差過大，從而限制了該方法在實際應用中的可行性和效果。 為了解決基於Cosserat Rod理論於同心管機器人動力學模型中存在的誤差問題，本研究採用資料驅動的深度神經網路（Deep Neural Network；DNN）運動學方法來改進模型的精確性和可行性，透過預蒐集資料集進行訓練，建立一個包含15層的全連接隱藏層，每層包含128個神經元的深度神經網路正向運動學模型，並於測試資料集中得到端點均方根誤差為1.03mm之大幅改善。之後，本文也提出了一種結合深度神經網路運動學和Jacobian的軌跡追蹤，實驗結果顯示，於2×2公分正方形軌跡中該方法的端點均方根誤差為1.064mm，相較具有穩定性與準確性。透過ZED 2i立體相機搭配自設計微型夾爪進行視覺追蹤定位目標物抓取，實驗之改善證明了使用結合深度神經網路運動學的Jacobian控制方法在同心管機器人具有實際應用價值。在實驗中，無航點規劃和有航點規劃操作成功率達到72%和88%。這些結果表明，本研究所提出的基於深度神經網路的Jacobian控制方法在實際應用中具有可行性和很高的精度，能夠有效提高同心管機器人在夾取任務中的控制性能。 最後，實驗比較兩種運動學方法和OptiTrack動作捕捉系統回授控制方法，在夾取任務中的總移動距離、執行時間和最後端點誤差的性能表現。運動學方法在沒有即時夾爪夾取點位置回授的情況，移動的總距離相對是比較短的，執行時間也是相對較少，但是最後端點誤差就相對較大。其中，Cosserat Rod 動力學方法中移動距離雖然是最短的，但平均最後端點誤差高達11.03mm，這將會造成夾取失敗。而DNN運動學模型因為有比較精準的運動學，所以在夾爪夾取點位置的估測上面相對比較準確，因此最後端點平均誤差在2.38mm，是可以進行有效地夾取，但這樣的誤差值有可能導致部分的夾取任務失敗。而無準確端點運動學模型狀態下，使用OptiTrack動作捕捉系統即時回授夾爪夾取點位置，在三種實驗的條件均能達到相當好的最後端點誤差，其誤差數值落在1.49-1.92mm之間，三種實驗的夾取成功率都可以確實達成。但是可以發現在沒有運動學輔助之下，夾取任務的執行時間和總移動距離是最長的，這並不符合實際機器手臂的應用。因此與運動學方法輔助使用，除了有效減少夾取的誤差，總移動距離和執行時間也大幅縮短，並提升整體夾取表現。

The concentric tube robot(CTR) is a type of soft continuum robot, constructed by inserting pre-curved different diameters elastic tubes into each other, with each tube is controlled by rotation and translation. Due to its small size and high flexibility, it is frequently used in minimally invasive surgeries and industrial exploration. This study developed a set of concentric tube robots with three Nitinol tubes with 6 degrees of freedom. First of all, the kinetics model of concentric tube robots based on Cosserat Rod theory was analyzed, and the root mean square error (RMSE) of tip position was 21mm, indicating that the Cosserat Rod kinetics model adopted simplification of assumptions, such as homogeneous material and ignoring friction, etc., leading to excessive errors in practical applications. In order to improve the kinematics model accuracy of the concentric tube robot, this study adopts the deep neural network (DNN) data-driven method to improve the accuracy of the kinematics model. The deep neural network includes 15 fully connected hidden layers. Each layer consists of 128 neurons, and tip position RMSE is 1.03mm in the test dataset. Then a method based on deep neural network kinematics Jacobian is proposed for waypoint planning. According to the experimental results, the RMSE of this method is 1.064mm in a 2×2 cm square trajectory, which is stable and accurate. Through the ZED 2i stereo camera and the self-designed micro gripper in the grasping task, the success rate for the operation without waypoint planning and with waypoint planning is 72% and 88%. These results indicated that the DNN-based kinematics Jacobian method proposed in this study is feasible and highly accurate in practical applications, and it can effectively improve the control performance of concentric tube robots in grasping tasks. Finally, the experiment compared the total movement distance, execution time, and tip position error in a grasping task using two different kinematic methods and the OptiTrack motion capture system feedback control method. Without real-time feedback on the position of the gripper grasping point, the kinematic methods had relatively shorter total movement distances and less execution time, but a higher tip position error. Among them, while the Cosserat Rod kinetics method had the shortest movement distance, its average tip position error was 11.03mm, which led to grasping failure. The DNN kinematics model had more precise position and thus provided a more accurate estimation of the gripper grasping point position. Consequently, its average tip position error was 2.38mm, which allowed for effective grasping, but this level of error might still result in some grasping task failures. In the absence of an accurate tip position kinematics model, using the OptiTrack motion capture system to provide real-time feedback on the position of the gripper grasping point resulted in great tip position error performance under three experimental conditions, with error values ranging from 1.49mm to 1.92mm. At this time, the success rate of the grasping tasks in these three experiments can be reliably achieved. However, it was noted that without the assistance of kinematics, the execution time and total movement distance for the grasping task were the longest, which is not in line with the practical application. Therefore, integrating the kinematics method not only effectively reduces grasping errors, but also greatly shortens the total movement distance and execution time, and improves overall grasping performance.