具挫屈關節之無指軟性夾爪及其在協作型機械手臂系統的應用──以動態接球取放為例

Abstract

軟性機器人利用軟材料的特性與結構設計，使其在與外界互動時，能憑藉材料的柔軟性和易變形特質，在無需複雜控制系統的情況下有效地抓握物品。軟性夾爪透過形態上的分類可分為無指夾爪、雙指夾爪、多指夾爪以及多關節夾爪。其中，無指夾爪透過特殊的幾何形狀或材料柔性來實現抓取，無需傳統的「手指」設計。無指夾爪的幾何特性使其在簡化控制下仍能達到穩定的抓握效果，但由於依賴材料變形來完成抓取，反應速度通常較慢。 本研究結合了實驗室先前開發的負壓挫屈關節與近期無指夾爪的設計概念，利用負壓挫屈關節的快速響應特性來提升無指夾爪的反應速度，同時保持穩定的抓握性能。雖然軟性夾爪可利用材料的自適應性來抓取不規則或脆弱的物品，但其易變形的特性通常會降低抓取精度，使得無法將物品精確定位至工具的中心點。為了解決此問題，本研究受到摺紙結構的啟發，將主動面板置於夾爪內，使夾爪在收合時，即便物體偏離預設的工具中心點，仍能透過主動面板將物體推回中心點，以達到精確的取放效果。 本研究會透過兩項應用將夾爪置於達明協作形機械手臂應用去反映上述所提到快速反應與精確取放功能，分別為透過接球實驗與視覺取放任務。在接球實驗中，研究結合了YOLO深度學習模型與支援向量回歸（SVR）技術進行飛行球體的偵測和軌跡預測。為了應對不同環境下的光線變化，研究運用了資料增強技術提升YOLO模型的泛化能力，確保模型在多種光源條件下均能穩定檢測飛行中的球體。研究還進一步比較了RealSense D435i與ZED深度攝影機在高速物體偵測中的性能差異，結果顯示RealSense D435i在準確度和穩定性方面優於ZED深度攝影機，使得本研究得以更精確地追蹤球體的運動。實驗結果表明，夾爪在0.64秒的反應時間下達到了55%的接球成功率。 在視覺取放過程中，夾爪展現了軟性材料的自適應性優勢，不僅能穩定抓握形狀不規則的物體，還能在一定的負重測試下維持結構穩定。為了進一步驗證夾爪的抓取精確度，研究測試了夾爪搭配達明機械手臂視覺取放系統的能力，成功將直徑4公分的高爾夫球放置在直徑3公分的瓶蓋上，顯示出夾爪在精確操作中的潛力。為了證明本研究設計的夾爪在具備快速反應能力的同時，亦能實現精確的取放操作，應在完成接球實驗後，進一步執行視覺取放任務。然而，由於協作型機器人內建的扭矩安全限制，導致無法在接球實驗後順利進行後續的視覺取放任務。未來將逐步調整達明協作型機械手臂的扭矩容許值，以驗證本研究開發的夾爪不僅具備快速反應能力，還能克服軟性機器人在精確取放方面的普遍挑戰。

Soft robots leverage the properties of soft materials and structural design to interact effectively with external objects, utilizing their flexibility and deformability to grasp items without the need for complex control systems. Soft grippers can be categorized by form into fingerless, two-finger, multi-finger, and multi-joint grippers. Among them, the fingerless gripper achieves grasping through special geometric shapes or material flexibility, bypassing the need for a traditional “finger” design. The geometry of fingerless grippers allows stable grasping with simplified control, though their reliance on material deformation typically results in slower response times. This study combines a previously developed vacuum-driven buckling joint from our lab with a recent fingerless gripper design, enhancing response speed by leveraging the fast-acting characteristics of the vacuum-driven buckling joint while maintaining stable grasping performance. While soft grippers adapt to grasp irregular or delicate objects through material compliance, their deformability can often reduce grasping precision, making it challenging to align objects precisely to the tool’s center point. To address this, we incorporated an active panel inspired by origami structures within the gripper, allowing it to push objects toward the tool center point during closure, ensuring precise placement even if the object is initially offset. This study implements the gripper on a TM collaborative robot arm to demonstrate its rapid response and precise pick-and-place capabilities through two applications: a catching experiment and a visual pick-and-place task. In the catching experiment, the study integrates the YOLO deep learning model and Support Vector Regression (SVR) to detect and predict the trajectory of a flying ball. To accommodate varying lighting conditions, data augmentation techniques enhance the YOLO model's generalization capacity, ensuring consistent detection of the ball across different lighting environments. Furthermore, the study compares the RealSense D435i and ZED depth cameras for high-speed object detection, showing that the RealSense D435i outperforms ZED in accuracy and stability, allowing for more precise ball tracking. Experimental results indicate that the gripper achieves a 55% catching success rate with a response time of 0.64 seconds. In the pick-and-place process, the gripper showcases the adaptive advantages of soft materials, maintaining stable grasping of irregularly shaped objects and structural stability under certain load tests. To further validate the gripper’s precision, the study tests the gripper’s performance in conjunction with the TM robot’s vision-based pick-and-place system, successfully placing a 4 cm golf ball onto a 3 cm cap, highlighting its potential in precise operations. To demonstrate that the gripper designed in this study can achieve precise pick-and-place operations while maintaining rapid response capability, it is necessary to perform visual pick-and-place tasks following the completion of the ball-catching experiment. However, due to the built-in torque safety limits of the collaborative robot, it is unable to seamlessly transition to the subsequent visual pick-and-place tasks after the ball-catching experiment. In the future, the torque limits of the TM collaborative robotic arm will be gradually adjusted to verify that the gripper developed in this study not only possesses rapid response capabilities but also overcomes common challenges faced by soft robots in achieving precise pick-and-place operations.