A Control Strategy for Dual-arm Object Manipulation Based on Fused Force/Position Errors and Learning Algorithms

• Main Authors: Bo-Hsun Chen, 陳柏勳
• Other Authors: Pei-Chun Lin
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/996q4n

碩士 === 國立臺灣大學 === 機械工程學研究所 === 106 === In Industry 4.0, the articulated robot arm plays an important role in intellectual manufacturing, and among robot arms, the dual arm system has many advantages, such as more degree of freedom and more firm grasping of large pieces. To operate dual arm robots, it needs multiple feedback especially force. However, there were few research about combining position and force feedback, and few using two independent position-controlled manipulators to compose the system. So in this thesis, Kalman filter is proposed to fuse position and force error measurement to estimate force feedback, and two independent position-controlled manipulators compose the dual arm robot to conduct experiments, which is more realistic in real manufacturing factories. In the part of the dual arm system coordinately grasping-and-moving objects, the controlled plant is assumed as a spring-inerter-damper model and the system identification method was proposed. Based on this model, the Kalman filter fusing measured force and position error to estimate the fusion force error was developed, and different grasping profiles are considered as different model parameters but the same model type. Then, model parameter identification, robustness of Kalman filter and comparison of the advantages of Kalman filter were carried out through experiments, and the dual arm system grasped and moved different objects along the spatial 8-figured trajectory to verify the controller structure. And in the part of the application of learning algorithm, after completely explaining the basic knowledge of machine learning, reinforcement learning and iterative learning control (ILC), ILC was applied on the task of dual arm system coordinately grasping-and-moving objects, and the improvement is proven by the theoretical and experimental method. Then, Proximal Policy Optimization algorithm and four simulation cases were illustrated, in which the usually usage of reinforcement learning to be a real-time controller is transformed to be the aided trajectory planner. The single arm pushing a spring to do force control experiments and the single arm 1D and 2D force control pseudo-grinding experiment were conducted to verify the proposed structure.