三節段連續體機械臂設計與影像伺服控制

Abstract

本研究設計出一組三節段連續體機械臂。在連續體機械臂的碟片設計上，創新地採用兩種A型和B型的碟片來取代常見的萬向接頭或環氧樹酯黏合；在連續體機械臂的主幹設計上，使用矽膠包覆鎳鈦合金的方式取代常見的彈性材料或彈簧，維持彈性並降低主幹的震動，這種創新設計有助於提高連續體機械臂的精準度。同時，本研究改良了常見的片段定曲率模型，特別針對主幹固定盤面的曲段算法進行調整，以獲得更精準的運動學模型，提高了機械臂的運動精度。 由於在大部分連續體機械臂的研究中，多數只做到單節段或兩節段的連續體機械臂，對於多數干擾都可以忽略；由運動學絕對精度實驗與軌跡追蹤實驗可以發現，本研究中使用的三節段連續體機械臂明顯地受到重力影響，進而影響運動學的準確性。在運動學絕對精度實驗中，使用Motion Capture System測量10組不同的位置，算得誤差35.73毫米與標準差18.18，同時發現第一節彎曲幅度越大，誤差也跟著受影響。在軌跡追蹤實驗中，針對方形、圓形與S形使用Motion Capture System測出均方根誤差分別為18.23毫米、19.15毫米與11.37毫米，證實常見的定曲率模型假設在多節段且受重力影響的連續體機械臂上會存在誤差，進而影響多節段連續體機械臂在真實環境的應用。 因此，本研究藉由慣性測量單元(Inertial Measurement Unit；IMU)與神經網路針對連續體機械臂的末端點估測位置，本研究收集了74542筆IMU訓練資料，並使用其中的20%作為驗證集，得出2.23毫米的訓練誤差與4.09毫米的驗證誤差，在4500筆未看過的測試集中獲得4.25毫米的誤差與4.93的標準差。 在最後的實驗階段，本研究對於提出的神經網路的末端點估測方式加入立體相機、三指夾爪做夾取物品實驗。在實驗中，由於相機精度、手臂姿態與350克的夾爪，使原先利用末端無負載的手臂訓練的神經網路產生誤差，最終成功率為20%。為了驗證相機精度與手臂姿態是否對於夾取物品實驗產生影響，本研究將IMU換回Motion Capture System再次進行實驗。經過實驗顯示，使用Motion Capture System實驗的成功率為73.3%，證實相機立體精度、手臂姿態與負載對於夾取物品實驗確實產生影響。 為了進一步提升連續體機械臂以神經網路為回授的控制方式的精準度，本研究進行了結合神經網路結果與影像伺服的夾取實驗。這項實驗最終成功率為66.7%。這個成果驗證了在不依賴Motion Capture System的情況下，透過影像伺服和IMU回授的神經網路在常見環境中仍能夠有效發揮其作用。

In this study, a three-segment continuum robotic arm was designed. A novel approach was employed in the design of the arm's disks, utilizing two types of A and B disks to replace conventional universal joints or epoxy bonding. For the main body design of the continuum robotic arm, a method involving silicon encapsulation of nickel-titanium alloy was adopted, in contrast to the conventional use of elastic materials or springs. This innovative design aimed to maintain elasticity while reducing main body vibrations, ultimately enhancing the precision of the continuum robotic arm. Additionally, the commonly used constant-curvature model was refined in this study, particularly in adjusting the algorithm for segments fixed to the main body, resulting in a more accurate kinematic model and improved arm motion precision. Unlike most studies that focus on single or dual-segment continuum robotic arms, this research unveiled distinct challenges with a three-segment continuum robotic arm due to its susceptibility to gravitational effects, which can consequently impact kinematic accuracy. Absolute positional accuracy experiments, as well as trajectory tracking experiments, revealed this impact. In the absolute positional accuracy experiments using Motion Capture System to measure 10 different positions, an error of 35.73 millimeters with a standard deviation of 18.18 was recorded. Notably, it was observed that larger deflections of the first segment correlated with increased error. Similarly, in the trajectory tracking experiments involving square, circular, and S-shaped paths, Motion Capture System recorded root mean square errors of 18.23 millimeters, 19.15 millimeters, and 11.37 millimeters, respectively. This confirmed the presence of errors in the common constant-curvature model assumptions when applied to multi-segment continuum robotic arms affected by gravity, subsequently impacting their real-world applicability. Consequently, to address these challenges, this study employed inertial measurement units (IMUs) and neural networks to estimate the endpoint position of the continuum robotic arm, collecting 74,542 IMU training data points and utilizing 20% for validation. This process yielded a training error of 2.23 millimeters and a validation error of 4.09 millimeters. Moreover, in a previously unseen test set of 4,500 data points, an error of 4.25 millimeters and a standard deviation of 4.93 were achieved. In the final experimental phase, the proposed neural network-based endpoint estimation approach was combined with stereoscopic cameras and a three-finger gripper for object grasping experiments. However, due to factors such as camera accuracy, arm posture, and the gripper's 350-gram weight, the neural network trained with an unloaded arm exhibited errors, resulting in 20% success rate. To verify the influence of camera accuracy and arm posture on object grasping, the study repeated the experiment using Motion Capture System, achieving a success rate of 73.3%. This confirmed that camera accuracy, arm posture, and load significantly affected the object grasping experiment. To further enhance the precision of the continuum robotic arm controlled through neural network feedback, a fusion of neural network results and image servo was utilized for object grasping experiments. This approach ultimately achieved a success rate of 66.7%. This outcome demonstrates that, even without relying on Motion Capture System, the neural network integrated with image servoing and IMU feedback can effectively operate in common environments.