PREPRINT: Sen Gupta,G., Messom, C. H., Demidenko, S., "Real-Time Identification and Predictive Control of Fast Mobile Robots using Global Vision Sensing", IEEE Transactions on Instrumentation and Measurement, vol 54, No 1, pp 200-214 2005. Real-Time Identification and Predictive Control of Fast Mobile Robots using Global Vision Sensing Gourab Sen Gupta 1,4 , C. H. Messom 2 , S. Demidenko 3 1 IIS&T, Massey University, Palmerston North, New Zealand 2 II&MS, Massey University, Albany, New Zealand 3 School of Engineering & Science, Monash University, Kuala Lumpur, Malaysia 4 School of EEE, Singapore Polytechnic, Singapore Email: G.SenGupta@massey.ac.nz, C.H.Messom@massey.ac.nz, Serge.Demidenko@engsci.monash.edu.my Abstract – This paper presents a predictive controller for intercepting mobile targets. A global vision system is used to identify fast moving objects and uses a colour threshold technique to calculate their position and orientation. The inherent systemic noise in the raw sensor data as well as vision quantization noise is smoothed using Kalman filtering before being fed to the controller, and it is shown that this leads to superior accuracy of the controller. The predictive controller is based on State Transition Based Control (STBC) technique. As a case study, STBC has been applied to a goalkeeper’s behaviour in robot soccer which includes interception and clearance of ball. Further evaluation of the controller has been done for shooting the ball towards a target position. The system is examined for both stationary and moving objects. It is shown that predictive filtering of rough sensor data is essential to increase the reliability and accuracy of detection, and thus interception, of fast moving objects. Keywords – Vision Sensing, Real-Time Image Processing, Kalman Filtering, State Transition Based Control, Prediction, Interception, Mobile Robots I. INTRODUCTION This paper deals with a robotic system comprising three major components: the global-vision based image analysis sub- system acting as a prime source of raw data, the Kalman filter which is applied to the data obtained from the vision system and the predictive controller which uses the filtered data. The role of each component and their impact on the overall system functionality and performance are discussed. The major focus and significance of the presented work is that it shows the importance of filtering raw sensor data to improve the accuracy of a vision-based intelligent controller for fast mobile robotic applications. When developing an intelligent control system, the filter design cannot be left as an afterthought but must be integrated into the system. Many automated learning based intelligent controllers employ neural networks, genetic algorithms and genetic programming techniques on raw sensor data in the expectation that the control algorithm will learn the noise characteristics of the data [1]. This study shows that explicitly developing a filter and evolving it along with the controller is preferable, rather than relying purely on the intelligent control algorithm.