Adaptive Neural Control in Mobile Robotics: Experimentation for a Wheeled Cart

This paper presents experimental results of an original approach to the Neural Network learning architecture for the control and the adaptive control of mobile robots. The basic idea is to use non-recurrent multi-layer-network and the backpropagation algorithm without desired outputs, but with a quadratic criterion which spezify the control objective. To illustrate this method, we consider an experimental problem that is to control cartesian position and orientation of an non-holonomic wheeled cart. The results establish that the neural net learns on-line the kinematic constraints of the robot. After several on-line learning lessons the net is able to control the robot at any conngurations in a limited cartesian space. It is interesting to use multilayered networks in control of robotic systems because of their following properties: their strong generalization capabilities of multilay-ered networks (a neurocontroller can be designed even no explicit form of a control law is known), their reduced computation time, their inherent robustness to real noisy data. A standard neural control scheme is build on the academic feedback control approach. The neural net replace the classical corrector (see Fig.1). The net input is a combination between the current state of the system and a desired state. The net outputs are the control parameters. Multilayered networks are usually trained with the well known backpropagation learning algorithm (direct backpropagation). The goal of backpropagation is outputs neural network feedback robot desired state Fig. 1. Neural feedback control to minimize the quadratic error E (addition of quadratic errors E p for every input pattern p) between the net and a desired ouput (see Fig. 2). Consequently, it requires an a priori knowledge of the desired outputs (e.g. the desired controls) corresponding to the net input. weight update by gradient descent possible inputs set of neural network error E outputs rE desired state robot model inverse outputs desired of the robot Fig. 2. Standard backpropagation (direct backpropaga-tion) Determination of desired controls is the main problem of using standard backpropagation in learning control of robotic applications. Two solutions are generally used to calculate it: the rst requires a reference model: a PID or fuzzy controller 4] or a human teacher 5]. This solution does not exploit the learning and identiication capabilities of neural nets (except for the human teacher).