Impedance Control with On-Line Neural Network Compensator for Dual-Arm Robots

VTEX(za-al.kv) PIPS No.: 118796 MATHKAP JINT1359.tex; 6/12/1996; 12:46; v.5; p.1
88
SHIH-TIN LIN AND HUI-CHEN TSAI
the manipulated object and its environment. Jean and Fu [6] also proposed an adaptive scheme for coordinated multimanipulator systems. In this method, the measurement of contact force is not required. Backes [7] proposed a unified control method which decomposes forces sensed at the wrists of the two manipulators into forces in the move subspaces, which cause system motion, and forces in the squeeze subspaces, which cause internal forces. The controller proposed in this paper uses a central controller similar to the one proposed by Schneider to create the position and force command for each arm. An impedance controller is then applied to each arm to track the desired commands. A successful impedance controller requires precise characteristics of the environment. As a consequence, the performance of the controller may become unsatisfactory if the environmental characteristics, such as stiffness and geometry, are not known accurately. One way to overcome this difficulty is to make an on-line estimation of unknown parameters of the environment and adjust theቤተ መጻሕፍቲ ባይዱtarget impedance [8]. The disadvantage of this approach is that the identification equation is highly nonlinear in terms of the unknown parameters which make the identification difficult. Meanwhile, to derive the identification equation, we had to make assumptions about the dynamics and geometry of the environment which are not always true. In this paper, we proposed a feed-forward back propagation neural network compensator to overcome above-mentioned problems for dual-arm robots. The neural network was used to compensate the effects of unknown parameters of the manipulator and the environment. In order to have a better performance, the neural network is improved by adding a second momentum term in the generalized delta rule and by fuzzifying the inputs of the neural network. In the past, applications of neural networks to control problems such as position control and trajectory control of robotic manipulators have been reported [9]–[14], but few applications of neural networks to force control of robotic manipulators, [15]– [17]. One basic feature of neural networks is that they estimate functions from sample data without mathematical models. In other words, they are model-free estimators. The research described in this paper was motivated by this characteristic of neural networks. The organization of the paper is as follows. In Section 2, the impedance control law is discussed. In Section 3, a neural network compensator is developed to compensate the uncertainties of the parameters. Section 4 describes a central controller for dual-arm robots. Simulation examples and results are discussed in Section 5. The paper is concluded with comments and discussions in Section 6. 2. Impedance Control Law The control law development begins with the dynamic model of an n-link manipulator. The equations of robot contact motion are [18]
Journal of Intelligent and Robotic Systems 18: 87–104, 1997. c 1997 Kluwer Academic Publishers. Printed in the Netherlands.
87
Impedance Control with On-Line Neural Network Compensator for Dual-Arm Robots
1. Introduction Although control of robot manipulators has been studied extensively in recent years, this study has focused primarily on single-arm robots. As a result, most of the robots used in industry are only handicapped operators due to their singlearm structure. As the demand increases for robots to perform more dexterous manipulations, intensive research on dual-arm robot control has been undertaken. For dual-arm robots, undesired forces will be transmitted to the load by the arms if position control is used for both arms. Luh and Zheng [1] offered a control policy using the concept of master and slave. The trajectory planning is done only for the first arm as a leader, the second arm follows the motion of the first arm. The disadvantage of this method is that load balancing is hard to achieved. Dauchez [2] and Yoshikawa [3] applied the hybrid position/force control to dualarm and multi-arm robots. Seraji [4] adopted hybrid position/force control law to control each arm. An adaptive control law was used to adjust the control parameters in the presence of unknown object parameters. This paper, however, did not mention how to create position and force command for each arm. Schneider and Cannon [5] solved this problem by defining an object impedance between