Online learning of the sensors fuzzy membership functions in autonomous mobile robots

We describe a technique which enables a fuzzy-logic based robot control system to automatically determine the membership functions (MF) of the input sensors online and in a short time interval. There is a necessity for such online self-calibration for fast changing and dynamic environments such as agricultural environments and difficult or inaccessible environments, such as nuclear reactors, underwater and space environments. In these media the robot has to learn the appropriate MF with no human intervention taking into account the difference in sensor characteristics in the different environments and changes in production requirements and repairing or otherwise upgrading robots. So there is a necessity to find a fast converging algorithm that can calibrate the MF online in real time with no need for human intervention or simulation. our work reports on an approach based on the use of a modified genetic algorithm to evolve the fuzzy MF of the individual behaviours. The MF of four behaviours were learnt online in an average time of 4 minutes for each behaviour in an outdoor environment. These learnt behaviours were then co-ordinated and tested in complex and dynamic environments in which the robot gave a very good response