基於神經網路的軟性機器手掌動態建模與抓握控制

Abstract

由於其天然的彈性，軟性材料能使機器手指自然而然地順應外力改變彎曲角度，而無需複雜的控制策略，因此在多指機器手領域中越來越受到重視。這種天生的被動順應性確保了機器人與人類以及周圍環境之間的互動安全性。然而，目前的研究仍缺乏將軟手指動力學模型、控制架構、氣動驅動與通訊系統整合於一體的完整方案。因此，本論文旨在建立一個完整的軟性機械手系統。本研究採用了比例-積分-微分控制器，並搭配優化方法，以建立穩定可靠的軟掌控制架構，優化方法為使用粒子群最佳化，藉此尋求最佳比例-積分-微分控制器的參數。此外，本論文介紹了一種利用深度學習技術訓練的軟體機器手臂動力學模型，藉以模擬致動器行為。根據所建立的手指動力模型，進一步提出兩種模型為基礎的控制策略：分別為具感測器回饋以及無感測器回饋的控制方式。為了讓整個手掌系統能實際運作，氣動氣路系統與通訊系統也被統合設計並納入本研究中。最後，透過模擬實驗展示了本系統的效能，說明了不同控制器架構的表現差異、動力模型與控制架構的整合效果，以及在不同抓握方式下手指彎曲角度的變化情形。結果顯示，若將深度學習動力模型與控制器架構進行整合，則可使回饋角度能夠準確追蹤理想角度，誤差極小且震動不明顯，證明本系統具備良好的控制穩定性與實用性。

Soft materials are increasingly valued in the field of multi-fingered robotic hands due to their natural elasticity, which allows soft fingers to passively conform to external forces and adjust their bending angles without relying on complex control strategies. This inherent compliance ensures safer interactions between the robot, humans, and its surrounding environment. However, current research lacks an integrated approach that combines soft finger dynamics, control architectures, pneumatic actuation, and communication systems into a unified framework. Therefore, this thesis focuses on building a complete soft robotic hand system. A PID (Proportional-Integral-Derivative) controller is implemented, along with optimization methods, to establish a robust control structure for the soft palm. The optimization method used is Particle Swarm Optimization (PSO), which is employed to find the optimal parameters for the PID controller. In addition, the thesis presents a dynamic model of the soft robotic hand developed using deep learning method to simulate the actuator behavior. Based on the proposed finger dynamics model, two model-based control strategies are introduced: one with sensor feedback and the other without. To fully realize the hand system, the pneumatic components and communication organization are also designed and integrated. Finally, simulation experiments demonstrate the system’s performance, showing the effectiveness of different control architectures, the integration of the dynamic model, and the variations in finger bending angles under different grasp types. The results indicate that combining the deep learning-based dynamic model with the proposed control architecture enables the feedback angles to closely follow the desired trajectories, maintaining minimal error and reducing oscillation.