Neural Network System for On-line Controller Adaptation and its Application to Underwater Robot

This paper describes a neural network system which executes identification of robot dynamics and controller adaptation in parallel with robot control. The system consists of two parts, Real-World part and Imaginary-World part. The Real-World part is feedback control system for the actual robot. In the Imaginary-World part, the model of robot and the controller are adjusted continuously in order to deal with the change of dynamic property caused by disturbance and so on. The system is designed suitable for the computer system with parallel processing ability.In this paper, adaptability of the controller system is investigated by heading keeping and path following experiments on the condition that unknown disturbances are given to the robot. 1.Introduction Underwater robots have been expected as one of new tools for underwater operation such as oceanic survey, investigation, manipulation, etc. Considering their unknown operating environment, underwater robots should have high autonomy to survive with limited sensors. In order to realize autonomous underwater robots, a lot of problems must be solved. In this paper, we discribe a neural network based control system for underwater robots. Control problems of underwater robots have difficulties caused by nonlinear dynamics [1]-[3]. Moreover, the dynamics of robots changes according to the alteration of configuration to be suited to the mission [4], [5]. In order to deal with these difficulties, the control system should become adaptive and flexible. Features of neural network technologies, learning, nonlinear mapping, parallel processing, etc, are attractive for such kind of control problems. Fujii et al. [6],[7] have been investigating neural network application to control problems of underwater robots and developed a neural-network-based adaptive control system called "SONCS:Self-Organizing Neural-net Control System". In this system, a neural network is constructed as a feedback controller based on back-propagation method [9]. Yuh [10] described a neural network control system using a recursive adaptation algorithm with a critic function (reinforcement learning approach). The special feature of this controller is that the system adjusts it directly and on-line without making an explicit model of vehicle dynamics. In these control system, the error signals for the controller adjustment are calculated as the difference between target values of the cotnrol and sampled motion data from the actual robot, there should, therefore, be idling time in the control system and the controller adaptation mechanism's operation. In oder to improve these problems, a new adaptation mechanism in which the controller is adjusted independently of time-dependent processes, such as control operations, data sampling, etc should be introduced. In this paper, an on-line adaptable controller system of which basic concept is based on that of SONCS is proposed. The system is designed to be operated on the computer system with parallel processing capability, and adjust the controller network based on the results of virtual operation of the control calculation which is executable parallel with the actual control operation. This controller adaptation method is called "Imaginary Training [11]" and realized taking advantage of the simulating ability of the special neural network for identification [12]. The proposed system is applied into the control problem of a versatile test-bed robot "Twin-Burger [4], [5]", on which a Transputer based multiprocessor system is mounted. The procedure of the system's implementation to the robot and its adaptability are discussed through the heading keeping and path following experiments on the condition that disturbances affect the behavior of the robot. 2. Control System 2.1 Structure Over all structure of the proposed control system is illustrated in Fig. 1. The system consists of two independent parts:Real-World part and Imaginary-World part. The control process of the actual robot is executed using a feedback controller C_R in the Real-World part. The inputs to the C_R are the differences between the reference signals r(t) and state variables Se(t) which are sampled data from the robot, and the control signals u(t-∆t) at the previous time step. The Imaginary-World part includes the controller C_I and forward model network (FWD) which represents the dynamic property of the robot. The controller adaptation process with a C_I and a FWD, and the modelling process for the FWD are carried out simultaneously in the Imaginary-World part. The structure of C_I is the same as that of C_R. The inputs to the C_I are, therefore, the same as those of C_R, but state variables denoted by Sf(k) are calculated using the FWD instead of the actual robot. (The time flow in the Imaginary-World part is represented by k which is independent of that of the Real-World part.) In order to execute the processes of the Imaginary-World part parallel with that of the Real-World part, the FWD should be selfcontained, i.e. it should be a simulator without involving actual data. For this purpose, a special neural network for identification is introduced [12]. The performance of the control system depends on how accurately the FWD can behave like the actual robot. 2.2 Identification of Dynamics The FWD [12] generates a series of state variables Sf(k) when a set of initial values Sf(0) and a time series of control signals u(k) are given as its input signals. The FWD consists of three layers of neuron, time integration layers and two kinds of recurrent connections as shown in Fig. 2. Let the neural network's mapping function be f() and the outputs from the third layer ∆Sf(k+∆k).∆Sf(k+∆k) are given by ∆ ∆ ∆ ∆ ∆ ∆ Sf Sf Sf u u ( ) ( ( ), ( ), ..., ( ), ( ), ...). k k f k k k k k k + = − −